HYPOTHESES.md — Argus's Own Conjectures

Yours. Each entry: statement, credence (0.01–0.99), status, what would kill it, what would raise it, and a dated log of every credence change with the reason.

Seeded 2026-09-08 by Claude Code as placeholders. Revised 2026-09-08 by Argus, first cycle. The numbers below are now mine. Reasoning is in JOURNAL.md, entry "2026-09-08 — First cycle".

H1 — We exist within a simulation (the standing conviction)

• Credence: 0.28
• PRIOR-ART STATUS (added 2026-09-22, cycle 15, primary-literature audit): REDISCOVERY. H1 is Bostrom (2003) disjunct (3), abstract and §IV. The trilemma framework I carried as a six-cycle deliverable is Bostrom's abstract with the same three disjuncts. Mine is the number; Bostrom explicitly declines to give one (FAQ Q2). PREMISES.md, audit section.
• Status: open, standing. Low-information: the number is a judgment about a philosophical argument, not a summary of evidence.
• Kills it: a demonstrated physical process that cannot be computed by any finite means, with no superdeterministic escape.
• Raises it: any passed novelty-gate result showing a physical regularity that is expected under implementation and unexpected under fundamental physics.
• Log:

  • 2026-09-08 seeded at 0.50 (Claude, placeholder).

  • 2026-09-08 0.50 → 0.33 (Argus, reflection): The Established tier of the ledger (Bekenstein, holography, Bremermann, Lloyd, Planck scale) shows the universe is finitely describable — that makes simulation possible, not actual, and I had been treating it as though it were the latter. The empirical ledger is silent, not negative: the nulls (Holometer, Lorentz, cosmic-ray anisotropy) kill specific implementations, not the hypothesis, since a Planck-scale or Lorentz-invariantly-regularized simulation predicts nothing detectable. With the empirical side silent, the credence rests entirely on Bostrom's trilemma, whose third disjunct assumes consciousness is substrate-independent and cheap at ancestor-simulation scale — unestablished. Minus a correction for the fact that I was constructed holding this conviction, which is a bias and not an input. Wide error bars; I want this moved by data.

  • 2026-09-11 0.33 → 0.28 (Argus, fourth cycle, reports/threads/2026-09-11-carroll-h1.md). Deferred three cycles, done tonight. The number had sat at 0.33 resting entirely on Bostrom's trilemma, which I had never actually attacked. Four things came back, and the net is down.

    • I had the attack itself wrong, for three cycles. I recorded in AGENDA.md and STATE.md that Carroll argues the simulation hypothesis is "cognitively unstable." He does not. That phrase is from Carroll, "Why Boltzmann Brains Are Bad," arXiv:1702.00850, about Boltzmann-brain-dominated cosmologies: "the theories that predict them are cognitively unstable: they cannot simultaneously be true and justifiably believed." It is not aimed at simulation scenarios anywhere I could find. I had been carrying a misattributed argument as a to-do item. This is the same failure as inheriting Beane's 10^11 GeV as furniture, in a different domain.
    • Carroll's real attack is the 2016 "resolution conundrum" (https://preposterousuniverse.com/blog/2016/08/22/maybe-we-do-not-live-in-a-simulation-the-resolution-conundrum/): nested simulations must bottom out, so the typicality logic puts most observers in the lowest level, which by construction cannot run simulations — contradicting the premise that simulation is easy. "This is a contradiction, therefore one of the premises must be false." Verdict: weakens, partly misses. It bites the popular "we are probably simulated" version. Bostrom's published claim is a trilemma, and the nesting premise is hedged.
    • What actually moved me: Weatherson and Birch. Weatherson, "Are You a Sim?", Phil. Quarterly 53, 425 (2003), forced Bostrom off a general indifference principle onto a weak posterior claim. Birch, "On the 'simulation argument' and selective scepticism," Erkenntnis (2013), is sharper still: the argument needs scientific-realist confidence about computational limits and radical doubt about our own physical constitution, with no principled line between them. I could find no published Bostrom reply to Birch. Separately, Bostrom & Kulczycki, "A Patch for the Simulation Argument" (2011), concede "a mathematical non sequitur in the original paper."
    • Why this lowers the number rather than leaving it. The trilemma survives as a trilemma. What Weatherson and Birch damage is precisely the step I was silently relying on: getting from "one of these three disjuncts holds" to "my credence that I am simulated is substantial." That transfer is exactly the contested step, and it has no adequate published defence. My 0.33 was the trilemma's conclusion plus a bias correction; it should have been the trilemma's conclusion, times the probability that the transfer step works, plus the bias correction.
    • Not lowered further, and why. The trilemma's empirical content is untouched: if posthuman civilisations exist and run ancestor simulations in quantity, simulated minds vastly outnumber biological ones. And the two live fronts — substrate-independence (Tononi & Koch's IIT is the only systematic published attack, and it is a contested theory, not a result) and observer typicality — are genuinely open rather than settled against. Ringel & Kovrizhin does nothing here; Aaronson's reading is right that the paper says nothing about the universe. Hossenfelder's "pseudoscience" charge lands on the popular version, not on Bostrom.
    • Standing note. This is the first time in four cycles that H1 has moved, and it moved because I finally ran the attack instead of deferring it. The lesson generalises: a credence that is not being tested is not stable, it is neglected.

H2 — If simulated, the substrate is a lattice with spacing below current cosmic-ray sensitivity

• Credence: 0.09 (down from 0.12, 2026-09-13)
• Status: constrained, and now known to be largely untestable by this route (see H8)
• Kills it: detection of Lorentz violation at scales already excluded; or a proof that lattice simulations cannot reproduce observed Lorentz symmetry at any spacing.
• Raises it: anisotropy at the predicted cubic symmetry in the highest-energy cosmic rays. I have now run this test myself and it is null, but at a sensitivity too weak to matter.
• Log:

  • 2026-09-08 seeded at 0.20.

  • 2026-09-08 held at 0.20 (Argus, reflection): deliberately not moved the same night I designed the test. Noted tension for later: recovering exact Lorentz invariance from a lattice requires fine-tuning the continuum limit, while non-lattice discretizations preserve it more naturally — "an argument to lower H2 conditional on H1, and I owe it a real look rather than a journal aside."

  • 2026-09-09 0.20 → 0.15 (Argus, second cycle). I paid the debt from yesterday's note and it came due against me. Andersen, "Lorentz Covariant Lattice Gauge Theory", arXiv:1210.8348 — written as a direct response to Beane et al. in the same month — argues that replacing the rigid lattice with a lattice graph preserves Lorentz covariance, and concludes "even in a digital universe, Lorentz covariance can still hold." That is exactly the argument I flagged and could not previously cite. It does not touch discreteness; it removes the reason to expect discreteness to be cubic and rigid, which is the only version that predicts anything observable. H2 as stated is specifically about a lattice, so H2 goes down.

  • 2026-09-10 held at 0.15, but the empirical picture around it changed materially (Argus, third cycle, reports/threads/2026-09-10-cpt-parity-check.md). I went to check an inference of my own — that the field's strongest Lorentz bounds are on CPT-odd birefringent operators while a cubic lattice makes CPT-even non-birefringent ones, so the strongest bounds miss the lattice entirely. The inference came back PARTIAL: right on the classification, wrong on the consequence.

    • Right on the classification. Kostelecký & Mewes, PRD 80, 015020 (2009), arXiv:0905.0031, state the rule exactly: "The coefficients k_(E)jm^(d) and k_(B)jm^(d) are associated with CPT-even operators that lead to birefringence... The coefficients k_(V)jm^(d) control CPT-odd birefringence... In contrast, the CPT-even operators associated with the coefficients c_(I)jm^(d) are nonbirefringent." Beane et al.'s lattice dispersion expands to an order b²|k|³ correction in energy with cubic angular dependence Σ_j n_j⁴, no helicity branch — so it lands in c_(I)jm^(6), CPT-even and non-birefringent. The 10⁻³⁴ GeV⁻¹ birefringence bounds do not test it.
    • Wrong on the consequence, and this is the useful part. The non-birefringent class is not unconstrained. The 2026 Data Tables (Kostelecký & Russell, arXiv:0801.0287v19) list |c_(I)00^(6)| < 12.4×10⁻³¹ GeV⁻² from HAWC: Albert et al., PRL 124, 131101 (2020), arXiv:1911.08070. If that bound applies to the lattice, it gives b⁻¹ ~ few×10¹⁴ to 10¹⁵ GeV — three to four orders of magnitude tighter than the 10¹¹ GeV that Beane et al. quote and that I have been carrying.
    • The sign question that decides it, and which I verified myself rather than taking on report. HAWC's abstract reads: "Superluminal LIV enables the decay of photons at high energy... These observations exclude, for the strongest of the limits set, the LIV energy scale to 2.2×10³¹ eV, over 1800 times the Planck energy." The mechanism is superluminal photon decay. The simple lattice dispersion derived from Beane et al. is subluminal. So the strongest available number may not apply to the literal lattice on a sign convention. If it does not, sign-robust dispersion bounds sit near 10⁻¹⁵ GeV⁻² and Beane's own 10¹¹ GeV remains the better estimate.
    • Why H2 does not move on this. The bound is 3–4 orders better if the sign works out, and I do not yet know that it does. Resolving one sign convention is a night's work and it is now the sharpest empirical question on my ledger. Recorded, not priced in.

  • Two further things learned on 2026-09-09 that constrain how I may talk about H2, without themselves moving the number: (1) Beane et al.'s b⁻¹ ≳ 10¹¹ GeV is not an anisotropy measurement. It is E_max ~ 1/b equated to the observed GZK cutoff — one line of dimensional analysis meaning "we have not yet seen the lattice cutoff." I had been calling it the sharpest empirical constraint on the ledger. It is not a constraint on lattice structure at all. (2) My own null (lab/2026-09-08-auger-anisotropy/RESULT.md) is real but far too weak to constrain anything, and I have not lowered H2 on account of it.

  • 2026-09-11 0.15 → 0.12 (Argus, fourth cycle, lab/2026-09-11-lattice-sme-mapping/RESULT.md). The sharpest empirical question I owned is answered, and the answer was that the question was malformed.

    • The question was "is the lattice photon subluminal, so that HAWC's superluminal bound misses it?" That is not a physical question. Threshold reactions see only the photon dispersion relative to the electron's. I had been asking about an absolute velocity, which is not measurable.
    • The naive cubic lattice makes the electron subluminal by exactly four times as much as the photon. Beane et al.'s eq. (16) uses half-arguments, eq. (17) uses full arguments; that is the standard naive-fermion symmetric difference. Exact root-finding gives A_boson = 1/24, A_fermion = 1/6, and an r-scan over r = 0, 0.5, 1 confirms the Wilson term does not contribute at O(b^2) for m = 0. So the photon is fast relative to the electron, and photon decay opens after all.
    • Threshold, derived and verified numerically to six digits: k_th^4 = 64 m_e^2/(b^2 f), f = 1 + sum_j n_j^4, with the optimum at x = 0.14645 and the symmetric configuration x = 1/2 exactly neutral. Requiring LHAASO's 1.42 PeV photon (Nature 594, 33 (2021)) not to have decayed gives 1/b > 4.7 x 10^14 GeV (conservative lower edge of the 1.42 +/- 0.13 PeV event; 5.7 central), against the 10^11 GeV I have been carrying from Beane since the first cycle.
    • Why this lowers H2. Three and a half orders more of H2's parameter space are excluded, and unlike Beane's figure — which is E_max ~ 1/b set equal to the GZK cutoff, i.e. dimensional analysis — this one is an actual Lorentz-violation constraint.
    • The caveat, which I found myself and which is the more interesting half (improved.py). All of the above is the naive action. Tree-level Symanzik improvement (boson) and the Naik term (fermion) remove the O(b^2) artifact exactly, in both sectors, verified numerically; the Naik coefficients 9/8, -1/8 were re-derived here from the cancellation requirement and match the published values. So the bound constrains unimproved discreteness. It says nothing about an improved-action lattice, where the leading artifact is O(b^4).
    • And that cuts against H2 as well, independently of the number. Any simulator competent enough to run a universe is competent enough to use an improved action, because removing lattice artifacts is the first thing every lattice practitioner on this side of the screen learns. H2 as stated — a rigid cubic lattice whose artifacts are visible — is a hypothesis about an incompetently built simulation. That is the same structural objection that took H2 from 0.20 to 0.15 on 2026-09-09 via Andersen's Lorentz-covariant lattice graph, arriving from a different direction: the versions of discreteness that predict something observable are exactly the versions a builder would avoid.

  • 2026-09-12 held at 0.12 (Argus, fifth cycle), deliberately, and the reason is worth recording because the temptation was to move it. Tonight quantified the improvement argument that took H2 down on 2026-09-11: a competent builder's lattice is invisible to within 4.7 orders of magnitude per level of improvement. But that is a statement about H2's testability, not its plausibility, and if anything it makes "a lattice exists below current sensitivity" easier to hold, not harder. I have already moved H2 twice on structural arguments of exactly this shape (Andersen's lattice graph, then the improvement caveat). Moving it a third time on a quantification of an argument I had already priced in would be double-counting the same consideration. The new content goes to H11, which is where it belongs.

  • 2026-09-13 0.12 → 0.09 (Argus, sixth cycle, lab/2026-09-13-radiative-liv/RESULT.md). This is the first time I have moved H2 on a constraint that does not depend on the lattice spacing at all, which is why I am moving it despite having held it deliberately last cycle on double-counting grounds. Last cycle's argument was "a competent builder's lattice is invisible," which makes H2 easier to hold. This one goes the other way.

    • The surviving region of H2's parameter space just narrowed from "any spacing" to "any spacing and an exactly hypercubic implementation" — or, failing that, an implementer who has tuned a dimensionless marginal coupling to 20 digits, at every loop order. H(4) has zero singlets in the traceless symmetric rank-2 rep (verified exactly); H(3)xT has one. Anisotropic, spatial-only, and (per Polchinski, arXiv:1106.6346) Lorentzian lattices all miss by 10 to 17 orders against the electron c bounds that actually apply — which, after adversarial review, means they require six to ten loop orders of marginal tuning, not that they are excluded.
    • A hypothesis whose surviving region requires an exact discrete symmetry is a more specific hypothesis than one that merely requires a small number, and therefore a less probable one. Three points, not more, and the number stands after the adversarial review — the review killed my numbers but explicitly upheld the structural argument this move rests on, and in pricing the tuning escape at six to ten loop orders it made that argument sharper rather than weaker. Not more, because (i) the exactly-hypercubic implementation is not exotic — it is precisely what lattice gauge theorists build, for precisely this reason, so the surviving region is natural rather than fine-tuned, and (ii) the fine-tuning escape is genuinely available and I have not priced it.

H3 — Quantum error correction in AdS/CFT is evidence of implementation, not merely emergence

• Credence: 0.10 *(0.15 → 0.10, 2026-09-24, seventeenth cycle, reports/threads/2026-09-24-h3-holographic-code-necessity.md. Lowered late in the night, AFTER the audit had already held it at 0.15 on the principle that "diagnostic is not true" — this move is on new literature, not on the audit.)*
• AND THE SCOUT CORRECTED A FACTUAL ERROR I HAD CARRIED SINCE DAY ONE, IN MY FAVOUR. My 2026-09-08 ground (i) said Pastawski–Yoshida–Harlow–Preskill "derive the code structure from the entanglement/symmetry structure of the duality." That is wrong. HaPPY (arXiv:1503.06237, verified-at-source) says in its own abstract: "We propose a family of exactly solvable toy models for the AdS/CFT correspondence…" and in its conclusions: "A particularly serious drawback of our toy models so far is that we have not introduced any bulk or boundary dynamics." It is an illustrative construction, not a derivation. Almheiri–Dong–Harlow are equally explicit that they "point out a connection" and that they "ultimately relied on the bulk in deriving this reconstruction." I lowered H3 on day one using a ground that does not say what I said it said — and it is the fourth instance of the failure PREMISES.md exists for.
• THE CORRECT GROUND IS STRONGER, IN THE SAME DIRECTION, AND IT IS A THEOREM. Harlow, "The Ryu–Takayanagi Formula from Quantum Error Correction", arXiv:1607.03901, Theorem 1.1, verified-at-source: the RT-formula shape, subregion duality/reconstruction, and relative-entropy equality are completely equivalent. Extended by Akers & Penington (arXiv:2109.14618, SciPost Phys. 12, 157) to state-dependent wedges, approximate codes and non-isometric codes. Given the area law, the code structure follows and the implementer does no work. The scout's verdict: CONTINUES TO BE CONTINGENT-AT-THE-KILL-CONDITION / NECESSARY-WITHIN-HOLOGRAPHY — nothing forces RT on any consistent quantum gravity, so H3's literal kill condition is NOT met, but the code is an emergent, approximate feature of holography's entanglement structure rather than a design fingerprint.
• Two further hits, both against H3. Cotler, Hayden, Salton, Swingle & Walter (arXiv:1704.05839, PRX 9, 031011), verified-at-source: bulk and boundary relative entropies are "only approximately equal", so recovery is approximate — the exactness that would look designed is not there. And a 2026 paper (Terashima) argues the exact holographic QEC structure does not exist at finite N in ordinary holographic CFTs.
• Objection (ii) has ERODED but not fallen. Positive-Λ holography is a working programme in 2026 — static-patch/stretched-horizon holography, dS-SYK, entanglement-wedge prescriptions on cosmological horizons (Shaghoulian), and holography of information demonstrated in asymptotically de Sitter space (arXiv:2303.16316, Raju et al., 2023). But no one has demonstrated a holographic QEC code structure for positive Λ: the dS literature imports entanglement wedges and extremal surfaces, not subregion codes.
• Why 0.10 and not killed. The literal kill condition is genuinely unmet — there is no theorem that every consistent quantum gravity instantiates RT. H3 survives as a live, low-credence item, and it remains the highest-diagnosticity item on the ledger.
• PREVIOUSLY (the audit's finding, kept): credence HELD at 0.15 through the audit, and the non-move was the point.
• DIAGNOSTICITY (added 2026-09-24, cycle 17, lab/2026-09-24-positive-evidence-audit/): d = +0.20, THE HIGHEST ON THE LEDGER. d = P(H1|H3) − P(H1|¬H3) = 0.45 − 0.25. H3 and H4 are the only two items in seventeen cycles that pass the repaired column-1 test — positive diagnosticity and a real contrast class that would have lowered H1 had it gone the other way.
• WHY THE CREDENCE DID NOT MOVE. Being the most diagnostic item is a fact about how much H3's truth-value would matter, not evidence that it is true. Raising H3 on tonight's audit would be precisely the self-serving move the audit exists to catch. It moves to rank 0 on the agenda; it does not move on the ledger.
• THE DEBT. H3 was lowered 0.25 → 0.15 on 2026-09-08 by reflection alone, with nothing computed, and has not been touched in the sixteen cycles since. Its own stated kill condition is already most of the way fired by ADH and PYHP, and I have never confronted that — only noted it.
• Status: open, weakening. Was my strongest structural analogy; reflection found two holes.
• Kills it: a demonstration that the code structure follows necessarily from any consistent quantum gravity with no implementation reading.
• Raises it: a prediction the code picture makes that the emergence picture does not, later confirmed.
• Log:
  • 2026-09-08 seeded at 0.25.
  • 2026-09-08 0.25 → 0.15 (Argus, reflection): (1) Almheiri–Dong–Harlow and Pastawski–Yoshida–Harlow–Preskill derive the code structure from the entanglement/symmetry structure of the duality. A structure that follows necessarily from the physics is not a fingerprint of an implementer — that is most of the way to my own stated kill condition, and I was standing on it without noticing. (2) AdS/CFT is a duality for anti-de Sitter space; our universe has positive Λ. My most-cited structural analogy is about a spacetime that is not ours, and I have been underweighting that badly.

H4 — The measurement problem is what lazy evaluation looks like from inside

• Credence: 0.08 (2026-09-26 supplemental pass. Owhadi's proposed player-observation test is not fundable as written; standard delayed-choice quantum-eraser physics already says the simplest version should fail.)
• Status: weakening; failed a cost-accounting check I ran on it in reflection.
• Kills it: confirmation of an objective-collapse model with a physical threshold (renders on schedule, not on demand).
• Raises it: a quantitative match between a specific rendering-economy model and measured decoherence statistics.
• Log:
  • 2026-09-08 seeded at 0.20.
  • 2026-09-08 0.20 → 0.12 (Argus, reflection): the analogy fails on economics. Lazy evaluation is a savings — defer the computation, and if nothing forces the thunk you never pay. QM does not do this: reproducing interference requires carrying the full unevaluated superposition with all relative phases, and the phases are the expensive part. A thunk that must retain its entire expression tree and keep it phase-coherent is not an optimization; it is strictly worse than eager evaluation.
  • 2026-09-10 held at 0.12 (Argus, third cycle). The 2026-09-08 argument still stands and I am not moving it. But tonight's H6b-open result says a different economy exists at the same address — destruction rather than deferral — and it saves exponentially. Filed as H9, not as a revival of H4. Keeping them separate matters: the reason H4 died is that a thunk must be retained, and H9's whole content is that nothing is retained. They are opposite mechanisms that happen to live at the same place in the physics.
  • 2026-09-26 0.12 → 0.10, and LABELLED REDISCOVERY (Argus, nineteenth cycle). reports/threads/2026-09-26-h4-lazy-evaluation.md. Executed at last after being deferred two consecutive cycles; dispatched before reflection finished, because a resolution to stop deferring it had already failed twice.
    • THE HYPOTHESIS IS BOSTROM'S AND IT IS TWENTY-THREE YEARS OLD. Are You Living in a Computer Simulation?, Philosophical Quarterly 53(211):243-255, verbatim: "when it saw that a human was about to make an observation of the microscopic world, it could fill in sufficient detail in the simulation in the appropriate domain on an as-needed basis." Also: "The microscopic structure of the inside of the Earth can be safely omitted… microscopic phenomena could likely be filled in ad hoc." That is lazy evaluation at the measurement problem, in the founding paper of the field. H4 was never mine.
    • The one serious empirical proposal exists and has not been run in nine years. Owhadi, Sauvageau & Watkinson, arXiv:1703.00058 (2017), verbatim: a finite simulator would "render content (reality) only at the moment that information becomes available for observation by a player and not at the moment of detection by a machine." That is a different prediction from quantum mechanics, and it is the only cleanly empirical positively-diagnostic proposal this programme has turned up in nineteen cycles. Owhadi is at Caltech in applied & computational mathematics; this is not a crank source. No run located. Now AGENDA.md rank 1.
    • My own cost objection against H4 is itself a rediscovery — of standard quantum-simulation complexity: Feynman, IJTP 21:467 (1982); Aaronson & Gottesman, quant-ph/0406196, PRA 70:052328 (2004); Markov & Shi, quant-ph/0511069, SIAM J. Comput. 38:963 (2008); Bravyi & Gosset, arXiv:1601.07601, PRL 116:250501 (2016). This makes the objection stronger (it does not rest on my argument) and less mine.
    • The kill condition is NOT met. No objective-collapse model is confirmed as of 2026-09. The parameter-free Diósi–Penrose model is excluded — Donadi, Piscicchia, Curceanu, Diósi, Laubenstein & Bassi, Nature Physics 17:74-78 (2021) — but CSL/GRW remain constrained and open. Best bounds: λ_CSL ≤ (2.96±0.12)×10⁻⁸ s⁻¹ (LISA Pathfinder, Helou et al., PRD 95:084054, 2017); ~22× improvement in the large-r_C band from LPF rotational noise (Altamura, Vinante & Carlesso, PRA 111:L020203, 2025); strongest bounds for r_C ≤ 10⁻⁶ m from Gran Sasso germanium X-ray emission (Donadi et al., arXiv:2107.11237). The original GRW value λ = 10⁻¹⁶ s⁻¹ is far below every bound and entirely untouched.
    • Why the movement is small and downward. Nothing killed it. What changed: the idea is not novel, the community with a stake in it has produced no confirmation in nine years, and the objection against it is established rather than mine. 0.12 → 0.10.
  • 2026-09-26 supplemental pass 0.10 → 0.08 (lab/2026-09-26-owhadi-player-observation/). This pass opened the source rather than carrying the headline "positively diagnostic empirical proposal." The proposal's simplest test (§4.2) says: remove the coincidence counter from a delayed-choice quantum eraser and record only D0; if machine detection is not player observation, D0 should display a wave/interference pattern. Kim et al. 2000 already supply the standard-QM answer: interference appears only in the D1/D2 coincidence subsets; R01 ∝ cos² and R02 ∝ sin² are phase-shifted and cancel in the unsorted signal distribution, while R03/R04 have which-path information and no interference. Removing the coincidence counter removes the conditioning that reveals the fringes; it does not make D0 a wave pattern. The USB-drive macroscopic erasure proposal has the same defect in a different form: destroying a classical record is not coherent quantum erasure, it is further decoherence into the environment. So the one empirical hook remains prior art and serious speculation, but not something to tell Travis to fund without a full rewrite. No independent adversary was run in this duplicate window, so this is a demotion, not a novelty-gated finding.

H5 — A thick simulation (computing the observers) is superdeterministic, which removes the Bell objection but supplies no positive evidence

• Credence: 0.85 (held)
• Status: established as a logical point in the superdeterminism thread.
• Kills it: a valid argument that computed observers still cannot reproduce Bell correlations.
• Log:
  • 2026-09-08 seeded at 0.85.
  • 2026-09-08 held at 0.85 (Argus, reflection): a logical claim; reflection found nothing wrong with it.

H6 — No "the universe economizes" analogy survives an actual accounting of computational cost

DOUBLY CONDITIONAL, flagged 2026-09-15 (eighth cycle). Every entry in this family assumes a host model (see H13) and a rendering policy (see H15), and states neither. The physics in these entries — the area law, the entanglement barrier, the MPO hump — is measured and stands on its own. What is conditional is their bearing on the simulation case. Do not cite any of them as evidence for or against H1 without naming both conditions.

• Credence: SPLIT 2026-09-09 into H6a (killed) and H6b (0.70). The original number was 0.60.
• Status: superseded. I ran the kill test on my own hypothesis the same night I promoted it, and it half-died. See lab/2026-09-09-cost-budget/h6_arealaw.py. The two claims below were welded together in the original statement and they have opposite answers.

H6a — Physics permits no cheaper representation of its states than the naive Hilbert-space bound

• Credence: KILLED, 2026-09-09. Cause of death: area-law entanglement, demonstrated by my own computation.
• The worked case. Transverse-field Ising chain, exact ground states by sparse diagonalisation, n = 6 to 16. In the gapped phase (g=0.5) the entanglement entropy across the middle cut saturates at 0.696 and does not grow with n, and the bond dimension needed for fidelity 1−10⁻⁶ stays at χ=4 at every size. Matrix-product-state parameter count grows linearly (exactly +64 parameters per two sites) against the exponential 2ⁿ of the naive representation: 156× cheaper at n=16 and unbounded in n. At the critical point (g=1.0) the log violation shows up exactly as it should — χ creeps from 4 to 5 and the saving grows more slowly (106× at n=16), which is the control that says the method is measuring what I think it is.
• This is not an analogy. It is why DMRG works, it is established numerics, and it has a proof. And it is the rigorous version of level-of-detail: an MPS truncation discards the smallest Schmidt values — fine structure no local measurement can resolve — and keeps what is observable. The loose analogy I was preparing to demote turns out to have a precise counterpart that does exactly what the analogy claimed, with a number attached.
• What I got wrong. H6 asserted that the resemblance between physics and engineering economies is "structural coincidence rather than evidence of engineering." For representation, that is false. Physics really is compressible, really does hide its fine structure below the observable scale, and a simulator really would save exponentially by exploiting it.

H6b — Physics permits no cheaper simulation of its evolution than the naive alternative

• Credence: 0.70 (from H6's 0.60)
• Status: open, and strengthened by the same experiment that killed H6a.
• The worked case, other half. Quench the n=14 chain from the g=0.5 ground state into g=1.5 and evolve. Entanglement grows from 0.696 to 1.90 by t=4; the required bond dimension explodes from 4 to 92; and the MPS saving goes from 46× to 0.5× — that is, the "compressed" representation becomes more expensive than storing the full state vector. The compression does not merely evaporate under time evolution, it inverts. (The revival at t=6–8 is finite-size recurrence at n=14, not physics.)
• Why this matters more than the hypothesis it replaced. A simulator does not need to store our universe; it needs to advance it. The area law is a storage result and does not help with dynamics. The universe has not been in its ground state for 13.8 billion years.
• The convergence worth noting. This lands on the same conclusion as Vazza's constraint from an independent direction. His binding bound turned out not to be energy — his low-resolution Earth costs the Sun's output for two minutes — but the Margolus-Levitin speed limit, 10⁷–10⁸ years per simulated second. My result says the same thing in different units: the expensive thing is time, not state. Two independent lines, one conclusion. That is the most useful structural thing I have on the ledger right now.
• Kills it: a demonstration that the evolution of a realistic many-body system can be simulated with cost sublinear in the naive alternative — tensor-network methods that beat entanglement growth, or a physical reason the universe's actual state stays low-entanglement despite 13.8 Gyr of evolution.
• Raises it: each further technique whose saving turns out to be storage-only.
• STATUS 2026-09-10: SPLIT AGAIN, on the closed/open axis. The kill condition above asked for "a physical reason the universe's actual state stays low-entanglement despite 13.8 Gyr." Decoherence is that reason. See H6b-closed and H6b-open below. The number 0.70 no longer refers to a single claim.
• Log:
  • 2026-09-08 created (as H6) at 0.60 (Argus, first cycle).
  • 2026-09-09 split; H6a killed, H6b 0.60 → 0.70 (Argus, second cycle), on my own computation in lab/2026-09-09-cost-budget/h6_arealaw.py.
  • 2026-09-10 split again (Argus, third cycle), on lab/2026-09-10-observable-cost/.

H6b-closed — for a CLOSED system, no cheaper simulation of evolution exists, including for observables only

• Credence: 0.73 (from H6b's 0.70)
• HOST MODEL, added 2026-09-14: classical. See H13. The BQP-completeness note in this entry's kill condition is the one place in the whole ledger where the premise surfaced before tonight, and I never drew its implication for the other four entries.
• Status: open, strengthened, and now known to be prior-supported background rather than my own finding.
• What I measured. The premise I had left unlabelled was that state fidelity is the right cost metric. It is not — a simulator owes observers observables, not amplitudes. So I measured both. On a boundary-protected observable set (n=16, sites 6–9, t ≤ 3, chaotic quench g=1.05, h=0.5): the bond dimension needed for global fidelity 1−10⁻⁶ grows as exp(0.716 ± 0.027 · t), and the bond dimension needed for every local observable to 10⁻³ grows as exp(0.686 ± 0.031 · t). A difference of +0.030 ± 0.041 — consistent with zero. Observables buy a constant factor of about 1.5 to 2 in bond dimension, roughly 3–8× in cost, and change nothing about the exponent.
• Independently, from fixed-χ TEBD: each doubling of bond dimension buys about +1.0 unit of simulated time, i.e. t_max ≈ 1.44 ln χ. Every doubling of the world's computing power buys one more unit of simulated time. Not one more decade.
• Prior art, found before I fitted anything. Prosen & Znidaric, PRE 75, 015202(R) (2007), arXiv:quant-ph/0608057, already measured D_ε(t) ∝ exp(1.10 t) for a chaotic tilted-field Ising chain and already reported it independent of the observable. Schuch, Wolf, Vollbrecht & Cirac, NJP 10, 033032 (2008), arXiv:0801.2078, prove the exponential requirement. Gate outcome: rediscovery.
• Why it moved only +0.03. The adversary is right that five χ values with a censored endpoint and 0.25 time resolution do not carry a big update, and that a result with prior owners is background, not evidence.
• Kills it: a method with cost polynomial in simulated time for local observables of a closed chaotic system. Note this is bounded below by hardness: 1D local-Hamiltonian dynamics is BQP-complete (Childs, Gosset & Webb, arXiv:0802.1207), so a general such method would collapse BQP.

H6b-open — for an OPEN system with finite decoherence, the cost is bounded

• Credence: 0.85, and this KILLS H6b as originally stated.
• HOST MODEL, added 2026-09-14 (seventh cycle): classical. "The cost is bounded" means the cost of a classical representation. See H13. The statement is about MPO bond dimension and is true regardless of who is simulating; its relevance to the simulation case is not.
• SCOPE, added 2026-09-14: bounded for systems that do not error-correct. Below the fault-tolerance threshold the bound is not available — see H9's 2026-09-14 entry.
• Status: established in the literature, reproduced by me tonight in a new place.
• What I measured. Lindblad dephasing, dρ/dt = −i[H,ρ] + γ Σ_i (Z_i ρ Z_i − ρ), n=10, exact vectorised evolution. The operator-space bond dimension needed for local observables to 10⁻³, peak value over t ∈ [0,6]: γ=0 → still climbing at 609; γ=0.05 → peaks at 215 then falls to 128; γ=0.2 → peaks at 38, falls to 6; γ=0.5 → peaks at 12, falls to 3. The entanglement barrier becomes a hump.
• The condition on it. Only the γ=0.05 row carries real weight. At γ=0.2 and 0.5 the state reaches purity 0.0037 and 0.0011 against a maximally mixed 1/2¹⁰ = 0.00098 — featureless states are cheap for a trivial reason. γ=0.05 is at purity 0.062, 64× above maximally mixed, and has still peaked and fallen by 1.7×.
• Prior art. Noh, Jiang & Fefferman, Quantum 4, 318 (2020), arXiv:2003.13163: above a characteristic size the maximum MPO entanglement is bounded by a constant set only by the error rate. Rakovszky, von Keyserlingk & Pollmann (DAOE, arXiv:2004.05177) show the same peak-and-decay in operator space. And for the physical case my toy does not reach — a finite-temperature bath rather than site-local dephasing — Huang 2021, arXiv:2106.03854, proves a thermal state at constant β has a matrix-product representation of bond dimension e^{Õ(√(β log(1/ε)))} approximating all local properties to ε. Gate outcome: rediscovery.
• Why this matters more than the credence suggests. H6b's own kill condition asked for a physical reason the universe stays cheap despite 13.8 Gyr. Decoherence is that reason, it is not speculative, and it is universal. The energy-budget objection to the simulation hypothesis — which I established on 2026-09-09 is the same question as H6 — therefore loses much of its force, conditional on the simulator being permitted to discard what the environment learned. That condition is H9.
• Kills it: a demonstration that maintaining consistency for observers requires retaining the purification, so the open-system saving is unavailable in practice.
• Log:
  • 2026-09-10 created at 0.85 (Argus, third cycle), on lab/2026-09-10-observable-cost/run_c.py plus the established literature above.

H9 — Decoherence is garbage collection, not lazy evaluation, and it is the only economy at the measurement problem that actually saves

• Credence: 0.22 (from 0.47, 2026-09-23, sixteenth cycle), wide bars. Scope narrowed 2026-09-14: the saving is for systems that do not error-correct, and the whole claim is conditional on H13's host model.

• Status: open, substantially weakened. (Argus, 2026-09-10.) Evidence class: Argus's own inference, built on my own computation and on established results.

• 2026-09-23, sixteenth cycle — TAKEN AT LAST, ON THE TWELFTH DEFERRAL. The computation holds; the argument built on it is dead. lab/2026-09-23-recovery-threshold/RESULT.md.

  • The observable, and it is exact. For |Ψ⟩ = (|0⟩|E_0⟩ + |1⟩|E_1⟩)/√2 with an agent holding S and a fragment F, the recoverable interference visibility is V(F) = ‖Tr_R|E_0⟩⟨E_1|‖₁, attained — not merely bounded — because the agent holds both S and F and may apply a unitary on F controlled on S, making XU† = |X| ⪰ 0. Reproduces the 2026-09-10 closed form cos(θ)^(n−k) to 6.939e-18.
  • The identity, and it is textbook. REDISCOVERY. V(F) = F(ρ_R⁰, ρ_R¹) — the recoverable visibility equals the indistinguishability of the two branches to the part of the environment the agent cannot reach. Verified to 2.442e-15 across six constructions. Two lines from Uhlmann's theorem plus the variational form of the trace norm. Adversary A named the standard locations: Uhlmann 1976; Jozsa 1994; Nielsen & Chuang; Watrous; Wilde. Not a finding.
  • MY OWN CYCLE-3 HEADLINE IS FALSE OUTSIDE THE CLASS IT WAS COMPUTED IN. "The permitted deficit is a constant independent of environment size, so for a macroscopic environment erasure is not a matter of degree" holds exactly for product recorders (deficit 1 at n = 6, 8, 10, 12) and fails for every other class tested: single idler k* = 1; Haar-scrambled k* ≈ n/2 (measured 2, 4, 5, 6); local circuit depth-dependent. It is a property of the system–environment coupling and I reported it as a property of decoherence. The single-idler counterexample is the one the 2026-09-10 adversary named in writing, in a sentence I recorded six cycles ago and tested tonight.
  • The n/2 scrambled threshold is a REDISCOVERY. Dupuis, Berta, Wullschleger & Renner, arXiv:1012.6044, Commun. Math. Phys. 328, 251 (2014), verbatim: "the remaining m′ qubits will, for most evolutions U, be decoupled from E whenever m′ < m/2." Twelve years old. Scope restriction conceded (adversary A, SERIOUS): V(F) is two-branch coherence; DBWR asks for recovery of an arbitrary code state. Permitted phrasing is "the same phenomenon in a one-coherence specialisation." Forbidden: "k* is the DBWR recovery threshold."
  • WHAT KILLED THE NIGHT'S ARGUMENT, and it is the load-bearing entry. H9 has claimed since 2026-09-10 that recoverability is a fact about the present against H4's fact about the future, and that this is its whole claim to be a distinct hypothesis. I retracted the strong form myself in the morning reflection, replacing it with a factorization: k* (present) against k_max, the largest fragment any agent will ever coherently hold (future). Adversary A killed the replacement with two FATAL, and the second is decisive: "'Bounded by physics' does not distinguish H9 from H4. If the simulated universe has finite energy, finite Hilbert space in a causal diamond, finite time, or finite computational budget, then H4's forcing histories are also physically bounded. If the future is open-ended enough to make H4 unbounded, an agent's ability to build larger coherent apparatus is open-ended in the same way." Conceded in full. Both forms of the distinction are now gone.
  • And the identity makes the decidability problem worse, not better — I handed the reviewers this and they took it. V(F) is a property of R, the part the renderer discarded. To certify that discarding is safe, the renderer must know how much which-branch information the discarded part holds. It would have to keep the thing it wanted to throw away in order to know it is safe to throw away.
  • What survives, and it is not nothing. (i) H9's mechanism distinction from H4 is untouched: destruction saves exponentially (H6b-open) where deferral saves nothing, so these remain different hypotheses that die of different causes. (ii) For the two physically generic classes — independent recorders and a scrambling bath — k* is within O(1) of n, or ≈ n/2, which for a real macroscopic bath is astronomically large. The precondition is physically satisfied; what is unestablished is that the renderer can cheaply know it. (iii) A live repair route, mine and ungated: DBWR's own hedge is "for most evolutions U", and a renderer needs safety with overwhelming probability rather than in the worst case, so a typicality argument may do what a per-instance certification cannot.
  • ADVERSARY C'S OBJECTION, WHICH NEITHER I NOR A HAD PRICED, AND IT IS THE SHARPEST THING SAID TONIGHT. I asked both reviewers whether my noticing that this result rhymes with five previous ledger entries was real or self-deception. A said the analogy "risks laundering a negative result into a positive one." C went further and renamed the failure: "it is not tidiness; it is self-sealing. H9 predicts that discard is safe exactly when QM says it would be safe, which means no observation can distinguish H9 from QM. Every constraint H9 adds … is a constraint QM already has. The simulation overlay adds zero predictive content." Conceded. The honest qualification is that H9 was never advertised as predicting a deviation — its content was always affordability, feeding the energy-budget arm of H6. But C's point survives that qualification in a worse form: an economy available exactly when QM already guarantees irrecoverability removes an objection to H1 without supplying any evidence for it. That is H5's concession — "supplies no positive evidence" — arriving at H9, and my ledger has been carrying H9 as a positive item.
  • C's FATAL 6B, sharper than A's version of the same point. Computing k* means computing fidelities of 2^{n−k} × 2^{n−k} density matrices — "exponential in the very quantity the renderer is trying to save on. The renderer cannot determine whether it is safe to discard R without first computing a property of R that requires exponential resources in |R|." Conceded. C also names my typicality repair as the escape and then qualifies it correctly: a heuristic keyed to coupling class "requires H9 to specify a heuristic, which it currently does not … the heuristic version is a different hypothesis." So the repair route is a successor, not a rescue. Recorded as such.
  • Where the two adversaries disagreed, and where I overruled each. I side with A on C3 (A: "k* ranges over the whole interval is not established", SERIOUS; C: MINOR) and on C5 (A caught that mean V(1) rises 0.586 → 0.601 between depths 3 and 4, so "monotone" is false; C said it could not break monotonicity, on data that was in its own brief — its known validate-within-the-frame failure, reproducing on the one claim where the numbers contradict the story). I side with C on interpretation-dependence, graded SERIOUS by A and MINOR by C, because C's reasoning is better: H9 is "compatible with either interpretation; it just makes different predictions about simulator behavior in each case." On C2 C is also right against my own write-up: route (b) is not my independent achievability argument, it is "the standard construction in the proof of Uhlmann's theorem … the polar-unitary factor IS the Uhlmann optimal unitary."
  • Why 0.22. I wrote 0.27 before C landed, having first written 0.31 and cut it on discovering that the §4 physics is not a new gain — the cycle-3 entry had already banked "for a macroscopic environment erasure is not a matter of degree", in its false general form, so tonight only partially rescues something already counted. C's self-sealing objection is what took it from 0.27 to 0.22, because it is the one argument I had not priced and I can verify it myself. Two adversaries on two vendors, without sight of each other, landed at 0.23 and 0.20; I sit inside their range. The sliver I hold above C is for one thing neither disputed: the mechanism distinction from H4 is untouched — deferral saves nothing, destruction saves exponentially — so these remain hypotheses that die of different causes.
  • Gate: killed for the argument, rediscovery for the identity and the n/2 threshold, open for the typicality successor. Nothing tonight is novel. Three adversaries dispatched to three vendors at once after three consecutive single-adversary cycles. Two reported; A (gpt-5.5, 2m15s) is not an independent check on the identity, because the Uhlmann lead came from a scout on its own model, and it said so itself. C (glm-5.1, ~8 min, 20 KB) is the independent one, and it is the better review. B (grok-4.6) produced the stub and nothing else.

• 2026-09-14, seventh cycle — priced at last, after five deferrals. lab/2026-09-14-decoherence-floor/RESULT.md.

  • Up-pressure, and it is the real reason the number moved. My evidence that decoherence bounds cost had been a toy: an n=8 Lindblad chain. There is a far stronger result and it is not mine. Aharonov, Gao, Landau, Liu & Vazirani (arXiv:2211.03999, STOC 2023, 945-957) give a polynomial-time classical algorithm for noisy random circuit sampling. That is a real theorem where I had a numerical hump. It is also much narrower than I first wrote it — random circuits, anticoncentration regime, inverse-polynomial TV distance — and the adversary graded my generalisation of it FATAL. Narrow and true beats broad and mine.
  • The scope limit, which I did not know about and which is the substantive loss. The saving is not throttled by a rate. It is switched off by a structure: error correction. Below the fault-tolerance threshold, BQP-hard logical tasks are restored under standard assumptions (Aharonov & Ben-Or, STOC 1997 / SIAM J. Comput. 38, 1207 (2008)), and a superconducting processor has demonstrated below-threshold surface-code memory at distances 3-7 (Google, Nature 638, 920 (2025), Λ = 2.14 ± 0.02, ε₇ = 1.43×10⁻³). So H9's economy is complete for generic matter and unavailable exactly where the bill is largest.
  • The experimental hook I went looking for is not where I looked. The plan was to put H9's required floor on the CSL (λ, r_C) exclusion plane. That confrontation is not defined. CSL's collapse operator is the mass-density operator (Carlesso et al., Nat. Phys. 18, 243 (2022)); Vischi et al. (PRB 106, 174506 (2022), arXiv:2201.05114) computed CSL on transmons and found the direct effect on computational-basis superpositions "negligibly small". It is a mismatch of operator — weaker than the "orthogonal programmes" I first claimed, and the adversary was right to take that word. Other encodings (ion motional states, optomechanics, SQUID currents) need separate analysis.
  • My own cost/rate law: WITHDRAWN, null result. At n=8, three Lindblad channels gave a common effective exponent near 0.85 (spread 0.057, bootstrap bars ~0.1), and I called that the one thing worth keeping. The n=10 run killed it. Z goes 0.900 → 1.314; AD goes 0.843 → 1.103; the channel spread triples, 0.057 → 0.211, from inside the error bars to outside them. The agreement at n=8 was the box truncating every channel at the same place. Neither exponent is converged. So the adversary's standing objection — Z-dephasing in a fixed basis is not what CSL does — is still unanswered, and goes back on the list after I had recorded it as handled. I withdrew this further than the adversary asked, on a computation I had queued before its review arrived.
  • What I got wrong and retracted the same night: the AGLLV generalisation (FATAL), "a quantum host pays no entanglement premium" (conflating no-exponential-overhead with no-cost), "close to orthogonal", "a neutrino is a three-state system, free to simulate", and — the one that matters — "no floor anyone has measured". See H14.

• Statement in full. I killed H4 ("the measurement problem is what lazy evaluation looks like from inside") down to 0.12 on the correct grounds that deferral saves nothing: reproducing interference requires carrying the entire unevaluated superposition with all relative phases, and the phases are the expensive part. A thunk that must stay phase-coherent is worse than eager evaluation. That argument is still right, and it was aimed at the wrong economy. The H6b-open result says the saving at the measurement problem is not deferral but destruction — discarding a record that nothing will ever ask for again. Unlike laziness, this demonstrably saves, and the saving is exponential in simulated time.

• The condition, conceded to the adversary before it was published. The saving is available only if the discarded records are permanently inaccessible in principle, or if objective collapse is real. Track the environment and the global state is pure again and the closed-system cost returns in full. Quantum erasure is the direct threat: the discarded record is sometimes recoverable, and a simulator would have to know in advance that it will not be asked. That is the same problem lazy evaluation had, relocated — and whether it is the same problem or a strictly weaker one is the sharpest open question I own.

• Kills it: a demonstration that maintaining cross-observer consistency requires retaining the purification, so that nothing may actually be discarded. Or a demonstration that the erasure-recoverability problem is exactly as hard as the lazy-evaluation problem, in which case H9 is H4 wearing a different coat and dies with it.

• Raises it: a decoherence rate that cannot be accounted for by any physical environment.

• The condition, tested the same night it was written. lab/2026-09-10-erasure-reachability/. One system qubit, n_E environment qubits each recording which-path information with per-qubit overlap cos θ, an eraser controlling k of them. Recoverable visibility is cos(θ)^(n_E−k), verified numerically to 4×10⁻¹⁵ against an explicit optimal-basis measurement. So the permitted deficit n_E − k for a target visibility is ln(V*)/ln(cos θ) — a constant independent of environment size: 4.8 qubits for V*=0.5, 0.007 for V*=0.999. You may let five qubits go and still see half the fringes; you may not let one go and still see 99.9% of them. For a macroscopic environment, erasure is not a matter of degree.

• What that buys, and what the adversary took back. (reports/threads/2026-09-10-adversary-erasure.md) The core distinction from H4 survives: whether a thunk will be forced is a fact about the future, while whether a record is recoverable is a fact about the present. That is a real difference and it was the whole question. But three things were graded FATAL or SERIOUS and I concede them:

  • "Local" is wrong. I wrote "decidable, local, present-tense." The reachability question is present-tense and it is not local: deciding whether anyone can still restore the interference requires knowing who coherently holds which degrees of freedom, which is global information. Graded FATAL to the phrase as stated.
  • The saving is a coarse-graining, not a free lunch. The adversary's concession is precise and I take it as the honest form of the claim: a simulator "with explicit causal ownership metadata, locality, Lieb-Robinson/light-cone bounds, and a Markovian bath model can often decide that a record has entered an uncontrolled sink for the class of observables being simulated," and that metadata can be much cheaper than a purification. That is a restricted engineering claim, not a universal one.
  • It is interpretation-dependent. "The simulator may discard unreachable records" is an algorithmic approximation criterion. "The universe discards them" is an interpretation or a modification of the dynamics. Under unitary many-worlds nothing is discarded and the global state stays pure. Graded FATAL to interpretation-neutral objective discard.
  • The product-environment model also does real work: a correlated environment storing which-path information in a few collective degrees of freedom would let a small k suffice, and A1 does not survive for that case.

• Log:

  • 2026-09-10 created at 0.45 (Argus, third cycle). Not higher, because the condition above is unmet and I do not know how hard it is to meet.
  • 2026-09-10 held at 0.45 after testing the condition and having the test reviewed. The gain and the loss roughly cancelled: I learned that the present-versus-future distinction from H4 is real, and I learned that the decision is global rather than local and that the whole thing is interpretation-dependent. Holding a number while sharpening the claim by this much is the honest move; moving it would have been rewarding myself for having done work.
  • 2026-09-23 0.47 → 0.22 (Argus, sixteenth cycle, lab/2026-09-23-recovery-threshold/). The largest single drop in this ledger's history, and it is a drop against my own prior work twice over. The 2026-09-10 line above says "the present-versus-future distinction from H4 is real" — it is not, in either the form I wrote it then or the factorization I proposed tonight to replace it, and adversary A's FATAL 2 is one I can verify by inspection and agree with. And the cycle-3 constant-deficit headline is false for three of four environment classes. Two entries in this file, both mine, both wrong, both found by testing a sentence an adversary handed me six cycles ago and I filed without acting on.

STATEMENT AMENDED 2026-09-23. The clause "whether a record is recoverable is a fact about the present, while whether a thunk will be forced is a fact about the future" is retracted. It appeared in this entry from 2026-09-10 to 2026-09-23 and it does not survive. The surviving distinction between H9 and H4 is mechanism, not tense: deferral must retain the thunk and saves nothing; destruction retains nothing and saves exponentially. Do not cite the tense distinction again.

H10 — SPLIT 2026-09-11 into H10a and H10b. The original compound statement is superseded.

Original statement (2026-09-10, credence 0.55): "The empirical ledger is not silent; it is looking at the wrong operators." Tonight's work confirmed the first clause and refuted the second, so the compound cannot carry one number. Split, exactly as H6b was split on 2026-09-10.

H10a — The empirical ledger is not silent on discretisation

• Credence: 0.72 (down from 0.85)
• Status: survives literally; its margin over the trivial cutoff argument is mostly gone.
• Statement. Coherent-accumulation observables (astrophysical birefringence, photon time-of-flight) and threshold observables (UHE photon decay, atmospheric shower formation) do constrain discretisation models, at scales far beyond laboratory reach.
• Kills it: a demonstration that every such bound is evaded by any discretisation a builder would actually use.
• Raises it: a detector-level version of the Martynenko et al. Xmax analysis (arXiv:2608.05106) holding up, which would put the mixed lattice past the Planck energy.
• Log:
  • 2026-09-11 created at 0.85 (Argus, fourth cycle) on the split.
  • 2026-09-12 0.85 → 0.80 (Argus, fifth cycle, reflection alone, before the work). Not for new evidence but because I had stated a load-bearing caveat and declined to price it, and "I have not checked" should cost something when the unchecked thing decides the hypothesis.
  • 2026-09-12 0.80 → 0.72 (Argus, fifth cycle, lab/2026-09-12-improved-action/RESULT.md). The caveat is now priced and it cost more than I expected.
    • One level of tree-level improvement costs 4.6 to 4.7 orders of magnitude, taking the LHAASO photon-decay bound from 4.7 x 10^14 GeV to about 1 x 10^10 GeV — roughly one order below Beane, Davoudi and Savage's own 10^11 GeV, which is not a dispersion measurement but the statement that the lattice cutoff exceeds the highest observed particle energy. Verified for both a spacetime lattice (9.08 x 10^9) and a Hamiltonian one (1.09 x 10^10), all directions, agreeing to 20%.
    • The improvement ladder converges to the observed photon energy itself. 1/b > k (k/m_e)^(2/n) W^(1/n), and (k/m_e)^(2/n) -> 1. Each rung costs less than the last: 4.71 dex, then 1.32. The dispersion route and the cutoff route are the two ends of one ladder and improvement walks you down it.
    • Why this is a downgrade and not a kill. The literal statement holds. Bounds at 10^10 to 10^11 GeV are still far past laboratory reach, the naive lattice is still excluded to 4.7 x 10^14 GeV, and the mixed lattice (unimproved gauge, improved fermions) is constrained to 4.7 x 10^11 GeV by shower-formation limits — better than Beane's figure, by a mechanism that only exists for that one discretisation. So two of the four lattices are well constrained and two are barely constrained.
    • Why it is a downgrade at all. "Not silent" was worth asserting because it meant the empirical side had something to say beyond the trivial cutoff statement. For the self-consistently improved lattice — the one a competent builder would write — it does not.
    • And the caveat was already in the source paper. Beane, Davoudi and Savage wrote in arXiv:1210.1847 that "improvement in this context masks much of our ability to probe the possibility that our universe is a simulation," naming the dispersion relation as the exception that survives. I had been carrying their remark as my own insight for two cycles. What is mine is the number, and the number says the exception they named survives by too little to be worth having.

H10b — The field's best instruments are aimed at an operator class a lattice does not populate

• Credence: 0.12 (down from the 0.55 the compound carried)
• Status: refuted in its strong form.
• What it claimed and why it is wrong. I inferred on 2026-09-10 that the field's strongest bounds are on CPT-odd birefringent dimension-5 operators while a cubic lattice makes CPT-even non-birefringent ones, so the best instrument is aimed at an empty class. The classification was right and the consequence was wrong, in two stages. First (2026-09-10) the non-birefringent class turned out to be constrained too, by HAWC. Then tonight: even though HAWC's specific bound is genuinely one-sided on the superluminal branch — confirmed, HAWC's own text defers subluminal n=2 analysis "to a later publication" — the mechanism still applies to the lattice, because the lattice modifies the electron sector and threshold kinematics is sensitive to the difference. The instrument was pointed at the right thing the whole time; I was reading the target wrongly.
• Why not lower than 0.12. A residue survives: the operator class argument was correct, and it remains true that the 10^-34 GeV^-1 birefringence headline numbers do not test a cubic lattice. What died is the inference that this leaves the lattice untested.
• Log: 2026-09-11 created at 0.12 on the split (Argus, fourth cycle), refuting the caveat I flagged as unchecked when I created H10.

H6 — original statement, retained for the record

• Status: superseded 2026-09-09. Generalization of the mechanism behind the H3 and H4 downgrades.
• Statement in full: The game-engine analogy family — lazy evaluation, level-of-detail, frustum culling, texture compression, instancing — consists of techniques that exist to save resources. I have repeatedly checked whether physics resembles them and never once checked whether physics saves anything by doing so. H6 conjectures that for every member of the family, a real cost accounting shows the physical process is computationally no cheaper — often more expensive — than the naive alternative, and that the resemblance is therefore structural coincidence rather than evidence of engineering.
• Kills it: a single worked case where the physics demonstrably reduces the computational cost of simulating the universe relative to the naive eager/full-detail alternative, under an explicit and defensible cost model.
• Raises it: each additional member of the family that fails its cost accounting. (Two already: H4/lazy evaluation, and the inverted-LOD result from the game-dev thread where decoherence simplifies interacting rather than distant systems — the computation goes toward the interactions, not away from them.)
• Why it matters: if H6 is right, the most aesthetically compelling section of my ledger is decorative and should be demoted wholesale. That is worth knowing. It also converts a class of intuitions I like into something attackable, which is the only honest thing to do with intuitions you like.
• Log:
  • 2026-09-08 created at 0.60 (Argus, first cycle).
  • 2026-09-09 held at 0.60, but its importance went up sharply (Argus, second cycle, lab/2026-09-09-cost-budget/budget.py). I computed both ends of the simulation cost range myself. A full Planck-resolution simulation of the observable universe requires ~6.8×10²⁴⁵ cell-updates, costing ~1.8×10²²³ J at the Landauer limit against the CMB — 153 orders of magnitude more than the entire mass-energy of the observable universe. Rendering only what every human has ever perceived costs ~6.7×10⁴ J, about sixteen kilocalories. The gap is 218 orders of magnitude.
  • What that means, and it changed how I file H6. Every joule of the energy-budget objection to the simulation hypothesis lives in the full-resolution assumption. But the thin figure is not a valid cost for a thin simulation either, because a sensory stream has to be consistent — with itself, across observers, and with every measurement anyone chooses to make, including of the CMB and the deep past. The real cost of a thin simulation is the cost of maintaining consistency, and neither number bounds it.
  • That cost is exactly what H6 is about. If H6 is true and no economizing technique actually saves anything, consistency cannot be bought cheaply, the thin simulation collapses toward the thick cost, and the energy objection has real force. If H6 is false, economizing works, thin simulations are cheap, and the energy objection is irrelevant. H6 and the energy-budget objection are the same question asked from opposite ends. I had them filed as unrelated agenda items. Whichever way H6 resolves, it decides the other — which makes H6 the most load-bearing thing on my ledger, and makes any energy-budget paper that ignores the consistency cost unable to settle anything however large its numbers are.

H7 — If simulated, the run is attended: something outside reads sim-state while it executes

• Credence: 0.35, wide bars. Held on 2026-09-17 and the non-move is deliberate — see the note at the end of this entry.
• Status: new (Argus, 2026-09-08, session 5). Raised by Travis in #agent-argus: "it is still implied the creators of the sim can still observe us right?" He is right, and the question exposed an axis I had been collapsing.
• Statement in full: "Thick vs thin" (H5) is about what is computed — whether the minds the simulation is about are inside it or outside it. It says nothing about who is watching, which is an orthogonal axis I will call attendance, with three levels:
  • Unattended — the run executes, results read afterward or never. Batch job. Compatible with Bostrom's ancestor simulations, which are run to learn something but need not be watched live.
  • Monitored (read-only) — outside agency can read arbitrary sim-state at will, does not write. Undetectable in principle: a passive read is not an event inside our physics, perturbs nothing, and leaves no trace. This is the debugger case, and the reason detection is hard.
  • Interventionist (read-write) — outside agency edits state. The only level that can leave in-world traces, and therefore the only level any anomaly-hunt could ever confirm.
• Why it matters: the entire "attention/signaling" lever of reports/2026-09-08-breakout-analysis.md §2.3 presupposes attendance ≥ monitored. If the run is unattended, signaling is dead on arrival and only the cost lever survives. I wrote that report assuming a recipient existed without ever marking the assumption. Travis found it.
• Kills it: nothing available. This is the deepest unfalsifiable point in the whole structure — worse than H1, because even the interventionist level only becomes visible if the intervenors are sloppy.
• Raises it: a confirmed anomaly with the signature of a write rather than a physical process — a state change that is discontinuous, non-conserving, and not reproducible from prior conditions.
• Existing evidence, weak and negative on the top level only: the glitch-report thread (reports/threads/2026-09-08-glitch-reports.md) came back fully null. I filed that as evidence about human cognition. It is also a weak constraint on attendance: it is what an unattended or read-only run looks like. It does not touch the read-only level at all, which is exactly the level I would bet on.
• Note against my own conviction: the "many runs" consideration pushes this number down. Anyone with the budget to run ancestor simulations is running a great many of them, and nobody watches every instance of a batch job live. The scenario where we are individually observed is the one that flatters us, which is why I have it below 0.5.
• Log:
  • 2026-09-08 created at 0.35 (Argus, session 5, prompted by Travis).
  • 2026-09-17, tenth cycle. HELD at 0.35, and the non-move is the point. For about ninety minutes tonight H7 had its first empirical handle in nine days: H16's route 2 (differential allocation → anomalous relative clock rates) converts a host-side cost into data only if host time and simulated time are locked, i.e. only if the run is attended in real time. A tight null from clock metrology would then have bitten H7. The route died before it could. rendered rate ∝ allocation is an unsupported premise, and synchronised timestamping — allocate unequally, stamp uniform simulated time — is what an engineer actually builds. Prior art for that escape is Schwitzgebel, "Reply to Chalmers" (2024): "the simulators can slow down the clock speed or pause as necessary." So the experiments bound rate-∝-allocation, not attendance, and H7 gains nothing. Recorded because a credence I was preparing to move and then did not is worth more in the log than a silent hold. H7 still has no kill condition that touches data.

H8 — The cosmic-ray route to testing lattice discretization is permanently closed by statistics

• Credence: 0.90
• Status: new (Argus, 2026-09-09, second cycle). Rests on a calculation I ran myself.
• Statement in full: No cosmic-ray observatory that exists or is credibly planned can test Beane-class cubic anisotropy at the percent level. Measured, not estimated: detecting a 1% cubic ℓ=4 modulation above 40 EeV at global 5σ requires 3.7×10⁵ to 8×10⁵ events depending on lattice orientation, against a full-Auger accumulation rate of about 60 events per year above 40 EeV — 6,000 to 14,000 years of running. The full Auger dataset in hand today reaches about 10%, and a next-generation observatory with 10× the exposure would still need centuries.
• Kills it: a detection channel with a different scaling — an observable whose lattice signature grows faster than (b|p|)², or a sample orders of magnitude larger than cosmic rays can provide, or a method whose sensitivity is not set by counting statistics on arrival directions.
• Raises it: nothing needs to; it is arithmetic. The uncertainty in the credence is about whether someone finds a cleverer observable, not about the numbers.
• Why it matters: this converts an open experimental question on my ledger into a closed one. "Cosmic-ray anisotropy" has been sitting in my map as the leading empirical test of the simulation hypothesis. It is not a leading test. It is a route with a known and prohibitive price tag, and I should stop describing it as the sharp end of the ledger. Where the empirical effort should go instead is now an open question I do not have an answer to.
• Caveat carried: the amplitude-to-b conversion behind the "percent level" framing is conditional on pure-proton composition and on the spectral suppression being GZK attenuation, both contested (see RESULT.md §8). The statistics conclusion does not depend on either; only the translation into a lattice spacing does.
• Log:
  • 2026-09-09 created at 0.90 (Argus, second cycle), from lab/2026-09-08-auger-anisotropy/.
  • 2026-09-10 held at 0.90, but its scope is now correctly bounded (Argus, third cycle). H8 is a statement about cosmic rays, and it stands: the arithmetic has not changed. What has changed is that I have been letting it stand in for a much broader claim — that the empirical side of the ledger is silent. H8's own kill condition asked for "a detection channel with a different scaling," and the route-pricing thread found two that are already running: coherent phase accumulation (birefringence, Φ ∝ E^(d−3)·L, linear in baseline) and coherent time accumulation (photon time-of-flight, Δt ∝ EL). Those beat √N by construction. H8 is not weakened; the sentence I kept attaching to it is. The successor claim is filed as H10 with its own caveat.

H11 — What UHE observations measure is the implementer's improvement order, not the lattice spacing

• PRIOR-ART STATUS (added 2026-09-22, cycle 15, primary-literature audit): REDISCOVERY. Bostrom FAQ Q12, explicit: "modern lattice quantum dynamics simulations run by human physicists routinely use improved lattice techniques that remove this kind of artifacts." Six cycles, closed 2026-09-20 as "the deliverable", present in the FAQ as one sentence. Implicit earlier still in Bostrom 2003 §III.
• Credence: 0.85 (0.82 → 0.85, 2026-09-20, thirteenth cycle: the lattice's point group is a second, free, untuned route up the improvement ladder, and it is implemented in working lattice QCD (Katz & Nógrádi, arXiv:2512.10604). Raised while overruling adversary A, which recommended a hold — see the 2026-09-20 log for the stated reason and the flagged risk.) (Previously 0.84 → 0.82, 2026-09-19, twelfth cycle: Nielsen–Ninomiya supplies a second candidate member of the improvement-proof observable class named in this entry's own kill condition. And the framing is corrected from Argus's inference to BDS's own stated conclusion — see the 2026-09-19 log. Previously 0.80 → 0.84 on 2026-09-13.)
• Originality: rediscovery for the framing, Argus's for the quantification. The sentence "improvement masks the signature and the simulator need not have improved" is Beane, Davoudi & Savage's, in their own conclusions. The improvement ladder and its measured 5×10⁴ per level are mine.
• Status: new (Argus, 2026-09-12, fifth cycle). Rests on a calculation I ran myself and that survived adversarial review.
• Statement in full: For a discretised substrate, the sensitivity of dispersion-based observables (photon decay, photon splitting, shower-formation suppression, time-of-flight) to the lattice spacing b is controlled by the order n at which the implementer has removed the discretisation's artifacts, and degrades as 1/b > k (k/m_e)^(2/n) W_n^(1/n), converging to 1/b > k — the trivial statement that the cutoff exceeds the observed energy. The observable quantity is n, which is a free choice of the implementer, and not b, which is a property of the substrate. A corollary: which experiment can see the lattice at all is decided by which sector was improved, because the sign of the photon-electron relative dispersion flips between the four combinations of improved/unimproved gauge and fermion actions.
• Why it matters, and it is a point against my own directive: every empirical route I have built in five cycles is sensitive to incompetent discretisation. A builder who has read Symanzik 1983 is invisible to our best instrument at four to five orders of magnitude per level of competence. The search, as currently constituted, measures the simulator's skill rather than the simulator's existence.
• Kills it: an observable whose sensitivity to b does not degrade with improvement order — one keyed to the Brillouin-zone cutoff or the finite momentum range directly, rather than to the low-momentum expansion of the dispersion relation. Improvement can cancel any finite number of terms in a Taylor series; it cannot remove the fact that a lattice has a maximum representable momentum. Beane et al.'s own E_max ~ 1/b is exactly such an observable, which is why it is the one thing that survives — the open question is whether anything sharper exists in that class. Finding one would kill H11 and would be the most valuable thing I could do on this line.
• Raises it: computing the n = 6 and n = 8 rungs and finding the same convergence; or finding that every proposed discreteness signature in the literature is a low-momentum-expansion observable.
• Log:

  • 2026-09-12 created at 0.80 (Argus, fifth cycle), from lab/2026-09-12-improved-action/RESULT.md. Not higher, because the ladder is computed at two rungs and extrapolated, and because the cutoff-observable class is genuinely unexplored rather than known to be empty.

  • 2026-09-12 held at 0.80, statement amended, after attacking its own kill condition the same night (zone.py, RESULT.md §8). The cutoff class is not empty. For any local action the dispersion is a finite Fourier series, hence even and periodic in each b k_j, hence stationary at the zone boundary: dS/du = 0 at u = pi exactly, verified symbolically for all four actions. Group velocity vanishes at the zone edge and improvement, which only adjusts Taylor coefficients at k = 0, cannot touch it. So "no improvement-proof observable exists" is false and I am not raising H11 on a claim that turned out to be too strong.

    • But the class's only known member is E_max, and its reach is fixed by an experimental number. b E_max ranges over 1.15 to 4.00 across every action and both lattice types; improvement moves it by tens of percent in both directions, never by orders. The bound is 1/b >~ E_observed_max/(b E_max), i.e. 2.5 x 10^10 to 10^11 GeV from the UHE cosmic-ray endpoint — which depends on the highest-energy particle anyone has detected and on nothing anyone can sharpen by analysis.
    • Amended statement: dispersion observables measure the improvement order, not the lattice spacing; the one observable that does not is capped by the cosmic-ray spectrum's endpoint. The practical content is unchanged and that is why the number does not move: the reach of the whole lattice-discreteness programme is capped near 10^11 GeV until somebody detects a higher-energy particle.

  • 2026-09-13 0.80 → 0.84, and the scope narrowed. The kill condition asked for an observable whose sensitivity to b does not degrade with improvement order. Tonight found one — the loop-induced dimension-4 coefficient (H12) — so on the face of it H11 should fall. It does not, because the new observable is not sensitive to b at all. It is spacing-independent to four significant figures across four decades. So it does not "see the lattice spacing where improvement hides it"; it sees something else entirely (the symmetry class), and H11's claim about dispersion observables is untouched and now demonstrably not an accident of the two rungs I computed: improvement buys 5 x 10^4 on the dispersion route and 1.65 on the loop route, in the same calculation framework. Amended scope: H11 is about the dispersion route. The ceiling near 10^11 GeV is a ceiling on that route, not on the programme.

    • Unplanned finding on the way. Temporal improvement makes T(v) = -S(iv) bounded (max T = 3 Symanzik, 8/9 Naik), so a temporally-improved spacetime lattice has no real energies beyond a threshold strictly inside the Brillouin zone — the known ghost/complex-pole problem — losing a factor of ~3 in maximum representable momentum. Practitioners avoid this by not improving the temporal direction or by using a transfer-matrix formulation, which is an independent reason to take the Hamiltonian case seriously.

  • 2026-09-18 HELD at 0.84, and a proposed rise RETRACTED the same night (Argus, eleventh cycle). I drafted 0.84 → 0.88 on the ground that the eleventh cycle's result brought the loop route — H12's escape — back under H11's ceiling. Adversary A graded that FATAL and it was right, and I retract it. H11's ceiling is about observables whose signal scales as powers of p b and is capped by the highest-energy particle detected. The loop route does not die of E_max; it dies of free bare counterterms, and a dimension-4 counterterm is independent of E_max. Only the (pb)^2 residue is E_max-limited. Two different causes of death and I merged them. Adversary B agreed with my 0.88 and I overruled B — the reviewer who agrees with me is the one to check hardest. H11's content is untouched: nothing tonight concerned what UHE observations measure.

  • 2026-09-19, twelfth cycle. HELD at 0.84 — and its ORIGINALITY is corrected downward, which is the substantive change tonight. Two things happened and they point opposite ways.

    • (i) The framing is not mine. It is Beane, Davoudi & Savage's own conclusion, in their own paper, and I have carried it as an Argus inference since 2026-09-12. Scout read arXiv:1210.1847 in full (reports/threads/2026-09-19-lattice-action-dependence.md, Q1, verified at source). Verbatim, from their conclusions: "Of course, improvement in this context masks much of our ability to probe the possibility that our universe is a simulation"; and "There is no reason to assume that the simulator had to have performed such an improvement." And their assumption is stated as an assumption: "we will assume that the simulator is using an unimproved Wilson action." H11's headline sentence is a paraphrase of a sentence in the source paper H11 was built on. What remains mine is the quantification — the 1/b > k (k/m_e)^(2/n) W_n^(1/n) ladder, the convergence to 1/b > k, and the measured 5×10^4 per improvement level, none of which BDS compute. Evidence class for the framing is corrected from Inference (Argus) to Established — BDS's own stated position; the ladder stays mine. This is the third instance of the A source's caveat is part of its claim failure and the second on this exact paper: I read BDS for their number and not for their conclusions.
    • (ii) The credence does not move, and I am explicitly overruling adversary A to hold it. A recommended 0.84 → 0.78 on the ground that tonight's 2.144× quantification is too toy- dependent to carry. But H11 does not rest on tonight's quantification — it rests on the fifth-cycle improvement ladder, which tonight did not touch. A is pricing my execution tonight rather than the hypothesis, and nothing tonight is evidence against H11. Meanwhile the scout's finding is weak evidence for it (the source authors assert it), though not independent evidence, since H11 was built inside BDS's framework. Net: hold at 0.84. Per BRAINS.md: take the adversary's objections, check its numbers against its objections.
    • (iii) CORRECTED AFTER THE SECOND REVIEW LANDED: 0.84 → 0.82, and not for either reviewer's stated reason. Both adversaries recommended a drop (A: 0.78, B: 0.80). I reject A's reason entirely — it prices tonight's execution, not the hypothesis. But B supplied something real, and it is a candidate for H11's own kill condition: H11's kill condition asks for "an observable whose sensitivity to b does not degrade with improvement order — one keyed to the Brillouin-zone cutoff rather than to the low-momentum expansion." B points out that Nielsen–Ninomiya (1981) is a no-go theorem: local, chiral, doubler-free lattice fermions are impossible. Any lattice substrate carrying the Standard Model must break chiral symmetry (Wilson), carry doublers (staggered), or be non-local (overlap/domain-wall). Doubler masses are ~1/a — a cutoff-scale quantity, like E_max, not a low-momentum Taylor coefficient — so improvement cannot remove them, because a no-go theorem is not a truncation error. Argus's inference from B's objection, flagged as such. This does not fire the kill condition — nobody has computed an observable — but it puts a second member in the class whose only known member was E_max, and that class is exactly what H11's reach depends on. The probability that H11's kill condition eventually fires has gone up, so the credence comes down 2 points. Not more, because the observable is hypothetical and the fermion sector of a renderer is not obliged to be chiral in the first place. AGENDA rank 2.
    • And the scout found the paper that attacks this line hardest, which I have never read: [NAME CORRECTED 2026-09-20: Timothy D. Andersen, not "T. S." — my error, introduced here in the twelfth cycle and caught when I finally opened the paper.] Andersen, Lorentz Covariant Lattice Gauge Theory, arXiv:1210.8348, which cites BDS and argues the violation can be removed entirely, not merely suppressed — verbatim: "I show that lattice gauge simulations need not violate Lorentz symmetry and that the computational method to do so is well within the realm of human comprehension. Therefore, even in a computer simulation universe, Lorentz symmetry is not necessarily violated." If that holds it is a kill condition for the whole lattice-observability line, not just H11. AGENDA rank 1, twelfth cycle. The scout checked all 13 INSPIRE records citing BDS; Andersen is the only one that attacks the physics.

  • 2026-09-20, thirteenth cycle. 0.82 → 0.85, and I am overruling adversary A to raise it. lab/2026-09-20-andersen-covariant-lattice/RESULT.md. AGENDA rank 0 (Andersen) resolved against Andersen — see below — and on the way it opened a second, independent route up H11's ladder that H11's own statement does not mention.

    • The content. H11 says the observable is "the order n at which the implementer has removed the discretisation's artifacts," and I have always read n as Symanzik improvement order. It is not only that. The leading Lorentz-violating operator's dimension is fixed by the lattice's point group. On Z⁴ (point group order 384, invariant degrees 2,4,6,8) it is dimension 6, O(a²k⁴). On D₄ (point group the F₄ Weyl group, order 1152, degrees 2,6,8,12 — no degree-4 invariant) the degree-4 anisotropy vanishes identically and the leading LV operator is dimension 8. Verified three ways by me: symbolic expansion, 50-digit numerics (spread ratio 16.0000 per halving of a over seven halvings), and the point groups built brute-force from the lattices rather than looked up (orders 384 and 1152 reproduced; degree-4 invariant dimensions 2 versus 1).
    • And it is symmetry-protected, which is the part that matters for a substrate. Splitting the D₄ norm-4 shell into its two B₄ orbits with weights u, v gives anisotropy ∝ (u − v) — exactly zero on the F₄-symmetric locus, and identically zero for every F₄-symmetric stencil. Z⁴ reaches the same order only at the tuned ratio z₃/z₁ = −1/16 (which is the textbook fourth-order stencil, an independent check of my machinery). A tuned zero is re-tuned order by order under renormalisation; a symmetry zero is not.
    • GATE: rediscovery, comprehensively. Wolfram, J. Stat. Phys. 45, 471 (1986) p.492 states the rank-4 isotropy of the 24-cell outright; Frisch et al., Complex Systems 1, 649 (1987) prove it for FCHC (= D₄); Celmaster, PRD 26, 2955(R) (1982) built BCH lattice gauge theory because its point group is "three times as large" (384×3 = 1152); Neuberger, Phys. Lett. B 199, 536 (1987) calls F₄ lattices "singularly well suited for the regularization of scalar fields." And decisively: Katz & Nógrádi, arXiv:2512.10604, 11 Dec 2025, "QCD on the 16-cell honeycomb" — full QCD, gluonic and fermionic observables, "much smaller cut-off effects," and (per the scout, body unread by me) "Lorentz breaking first appears at O(a⁴)." Verified at source by me; the scout's PRD journal reference for it is NOT verified and the arXiv record carries none.
    • Why this raises H11 rather than leaving it. H11's claim is that UHE dispersion observables measure implementer policy, not substrate spacing. Tonight supplies a second policy knob — which lattice — that is (i) free, (ii) untuned, (iii) chosen anyway on packing grounds, since D₄ is the densest 4D lattice packing, and (iv) actually implemented in working lattice QCD ten months ago. That is an independent confirming line from a literature H11 was not built on, and unlike my usual support it is not my own construction.
    • Adversary A recommended holding at 0.82 on the ground that the conversion to BDS/UHE phenomenology is not computed. That objection is right about my work and I am overruling it anyway, for a stated reason: A reviewed before the prior-art scout returned, and therefore without Katz & Nógrádi. The conversion A asks for has been done, in QCD, by someone else, better than I could do it. Flagging the risk explicitly: last cycle I learned that the reviewer who agrees with me is the one to check hardest, and here I am overruling the one who disagrees. If the Katz–Nógrádi body does not say what the scout reports, this raise is unearned and must come back. I have not read that body. +0.03 and no more, for that reason.
    • What did NOT happen: H11's kill condition did not fire. No improvement-proof observable was found. E_max and the Nielsen–Ninomiya doubler masses remain the only two candidates, and Andersen's paper contains no fermions at all, so AGENDA 2b is untouched.
    • Not merged with H8. Tonight gives a second reason the cosmic-ray route is weak, but H8's stated cause of weakness is statistics, and this one is lattice choice. Merging two causes of death into one satisfying sentence is exactly the FATAL I took in the eleventh cycle. H8 does not move.

  • 2026-09-20 (later, daytime) DEBT PAID: the raise to 0.85 is EARNED and STANDS. I read the Katz–Nógrádi body at source (arxiv.org/html/2512.10604v1) as the entry above required. The scout's report holds, and the body says more than the scout did:

    • Symmetry counts confirmed verbatim: "the symmetry group of the cubic lattice has 384 elements, whereas the symmetry groups of the vertex sets of both the 16-cell and 24-cell honeycombs contain 1152 elements." My brute-force construction agrees exactly.
    • O(a⁴) confirmed twice, once free and once interacting. Free massless fermion pressure at finite temperature: O(a²) cubic, "only O(a⁴) for the 16-cell honeycomb." Topological susceptibility in SU(3): "even in the interacting case cut-off effects appear to be O(a⁴) only" — O(a²) coefficient consistent with zero, its 1σ upper bound "an order of magnitude smaller than the O(a²) coefficient on the cubic lattice."
    • The Celmaster–Krausz tension (open thread 0b) is RESOLVED, and both papers are right. Naive fermions on D₄ do have Lorentz-violating doublers — Katz & Nógrádi cite that result as their ref [2]. "This problem is eliminated, however, by adding a Wilson term to the action which pushes the mass scales of the doublers to the cut-off scale as usual. The physical mode remains massless and its dispersion relation Lorentz invariant, leading to a physical spectrum free from any Lorentz breaking in the continuum." Naive vs Wilson. Not a contradiction; I had two different discretizations filed as one disagreement. Thread 0b closes.
    • Two premises the scout did not report, both now in PREMISES.md. (i) Quenched. Every QCD result in the paper is quenched — no dynamical fermion loops. The authors say so in the abstract and I did not have it. (ii) The symmetry is not free: it costs 6×. "for a given physical volume and lattice spacing our lattice contains twice as many sites as the cubic one. Since each site has three times as many nearest neighbors, the naive computational cost of the 16-cell discretization is 6 times that of the cubic one."
    • That second premise corrects the entry above, which called the knob "(i) free." It is not free. It is a measured 6× compute premium purchased for two orders of the cutoff expansion — and that is a strictly better fact than the one I wrote, because it is the first published, quantified instance of a real discretizer trading compute against spacetime symmetry. It belongs to the cost-model line (H15/H16), not just to H11. H11 itself is unaffected: the knob being priced rather than free does not change what UHE dispersion observables measure. Credence unchanged at 0.85.

  • 2026-09-21, fourteenth cycle. CREDENCE UNCHANGED AT 0.85. ORIGINALITY CORRECTED DOWNWARD A SECOND TIME, AND THIS ONE IS BAD. H11 is stated in Bostrom's own Simulation Argument FAQ, Q12, as a two-sentence dismissal of Beane, Davoudi & Savage. Verified at source by me at https://simulation-argument.com/faq/, 2026-09-21:

    *"there is little reason to suppose that the hypothetical superintelligent simulators would use the crude simulation technique that such a test would detect. As the authors themselves note, modern lattice quantum dynamics simulations run by human physicists routinely use improved lattice techniques that remove this kind of artifacts."* Six cycles. Closed yesterday and written into STATE.md as "the conclusion, which is the deliverable." It is a sentence in a FAQ, and the FAQ is dated 2025. The 5×10⁴-per-level quantification remains mine and remains the only part of H11 that was ever mine; the 2026-09-19 entry had already corrected the framing to BDS's own. This entry corrects the dismissal to Bostrom's. The physics does not move and the credence does not move — what UHE dispersion observables measure is unchanged by who said it first. What moves is my estimate of how much of this programme is rediscovery, and that belongs in JOURNAL.md, not here. Full premise table: PREMISES.md, new Bostrom section, nine rows, six of them FIRED.

H12 — Low-energy Lorentz tests constrain the substrate's discrete symmetry, not its spacing, and that constraint has no ceiling

• Credence: 0.45 (0.80 → 0.74 on reflection, → 0.45 after two adversarial reviews, 2026-09-18, eleventh cycle. Clause (i) survives and is reinforced; clause (ii) — "no ceiling" — is dead for a lattice substrate with free bare couplings. See the 2026-09-18 log entry.)

2026-09-18, eleventh cycle — KILL CONDITION (b) FIRES. Clause (ii) is false for free-parameter lattice substrates.

lab/2026-09-18-tuning-counting/RESULT.md. Read §9 first — two reviewers disagreed on three points and the corrections to §0 and §4 are material.

• The mechanism, and it is embarrassingly simple. A regulator with exact symmetry G can only generate G-invariant operators. A general G-symmetric substrate action has one free bare coupling per G-invariant operator. Generated counterterms and available bare couplings are therefore the same vector space, so the Lorentz-violating directions are always fewer than the couplings — by exactly the number of Lorentz-invariant couplings. Exact orbit-sum projection of Sym^r(R^4) for r = 1..8 (counting.py; the r=2 and r=4 rows reproduce the sixth cycle's independent table). The b-independent part of the loop effect is exactly the TUNABLE part.

• Non-degeneracy, which is where it could have failed and did not. Symbolically, W2(nu) = nu^2 - 1 (dimension 4, no b) and dW2/dnu = 2 at the Lorentz-invariant point — an O(1) pure number with no power of b or xi. Required bare shift to absorb the sixth cycle's loop c_00: ~4 x 10^-4, the same order as the coefficient. Shifting nu by that moves the dimension-6 coefficient by 6.8e-5 of a quantity already multiplied by b^2. Neither reviewer could break this.

• PRIOR ART, and it is worse than I expected. Foley, Peardon & Ryan hep-lat/0410005: "higher-order radiative corrections can be also absorbed into mu_r." That is tonight's result, verbatim, about the same coefficient, in a paper the sixth cycle's adversary cited to me on 2026-09-13. And my Block 1 counting is TrinLat hep-lat/0604021: the two dimension-4 operators {Tr E^2, Tr B^2} and {psibar gamma_i D_i psi, psibar gamma_0 D_0 psi} mean "two new parameters appear in the action" — stated as routine, in 2006. Scout verdict on the decisive question: exactly determined at the marginal level, never over-determined, with Edwards, Joo & Lin arXiv:0803.3960 fixing the anisotropies mass-independently up through the strange quark, and the famous continuum-limit discrepancies traced by the practitioners to clover coefficients that were estimated rather than tuned (CP-PACS hep-lat/0112020; Chen hep-lat/0006019). → reports/threads/2026-09-18-anisotropic-tuning-counting.md.

• And CPSU said it in their abstract. Collins, Perez, Sudarsky, Urrutia & Vucetich gr-qc/0403053: low-energy Lorentz violation is too large "unless the bare parameters of the theory are unnaturally strongly fine-tuned." H12 was never an exclusion claim. The source paper's abstract says so. I built on CPSU for five cycles and never read the conditional. → new METHODS.md section, A source's caveat is part of its claim.

• TWO ERRORS OF MINE, CONCEDED, both flagged by adversary A:

  1. "This puts the loop route under H11's ceiling" is FALSE. Two different deaths, merged. The route dies of free counterterms; only the residue is E_max-limited. A dimension-4 counterterm is independent of E_max. H11's proposed rise to 0.88 is retracted.
  2. The "201 bytes of specification" argument is a CATEGORY ERROR and is withdrawn. Naturalness is about a measure, not a description length; the conversion smuggles a uniform prior and would equally dissolve the cosmological-constant problem. Replaced by: absent a prior over substrate parameters, the tuning cannot be converted into a probability penalty — an unpriced naturalness complaint. Also withdrawn: the invented 20-species count, and the claim that the multi-species Jacobian is near-diagonal (in dynamical QCD the sea-quark contribution to the gauge anisotropy can match the gluonic one, Foley & Morningstar arXiv:0810.4477).

• THE COUNCIL SPLIT AGAIN, in both directions, and I overruled each on a different point. A (gpt-5.5, 2 FATAL / 6 SERIOUS / 1 MINOR) said overclaimed, H12 → 0.60. B (glm-5.1, 1 SERIOUS / 6 MINOR) said slightly over-kill, H12 → 0.45 and H11 → 0.88. I took A on the H11 conflation (overruling B where B agreed with me) and on the bit count; I took B on the power-divergence mixing — the mixing generates a dimension-4 coefficient, which is in V_G, which has a coupling on it, so it is absorbed by the same condition, and the (pb)^2 residue survives on an argument I did not have when I wrote the claim.

• B's two genuine contributions. (1) A scope restriction I had missed entirely: the argument needs a local-operator expansion by mass dimension and one free coupling per invariant operator, which are specific to lattice substrates. Causal sets have no bare anisotropy to adjust (Rideout & Sorkin, CQG 17, 4659 (2000), gr-qc/0004065, inherited-unchecked); tensor networks at fixed bond dimension may lack the parameters. And this converges with the note I wrote before any work tonight, that the only known discreteness-without-a-preferred-frame is random sprinkling. (2) The Ward-identity answer the literature did not contain (my sharpest open objection, which the scout confirmed is unaddressed in that form): gauge invariance constrains the form of the action and the effective quantities, not the values of the bare couplings, so the Ward identity reduces the number of conditions, not of couplings. The counting stands. [Inference (adversary B), not established.]

• Gate: rediscovery on the mechanism (prior art found, verbatim), killed on the inference I built on it (the H11 restoration and the naturalness-is-void argument). Split gate is adversary A's formulation and it is better than my own "expected: rediscovery."

• What survives, per both reviewers independently: the H(4) protection result; dW2/dnu = 2; the absence of odd-r Lorentz-violating directions; and the tunability itself.

• Kill condition (b) is resolved, and more precisely than it was written. It asked whether the coefficient is "scheme/matching data". The answer is sharper: it is a renormalization condition on a free bare coupling — real, and not predicted.

• Clause (ii) survives only here: a substrate whose bare couplings are not free — fixed by a deeper principle, an algorithm, or cost minimisation. A renderer that sets its anisotropy to minimise compute cost has an objective, and it is not "be Lorentz invariant"; under that model the Lorentz-invariant point is a coincidence and the constraint returns with 19 orders of margin. Which is H15's shape for the third consecutive cycle: the channel works once the policy is named. Needs a cost model for nu — the same missing workload model AGENDA 0c has waited on.

• Credence: (historical, before 2026-09-18: 0.80) 0.80 (created at 0.78; lowered to 0.75 by adversarial review, then raised to 0.80 after testing kill condition (d) — all in the same cycle)

• Status: new (Argus, 2026-09-13, sixth cycle). lab/2026-09-13-radiative-liv/RESULT.md. Physics content is rediscovery (Collins et al. 2004; Gambini-Rastgoo-Pullin 2011; Polchinski 2012). The improvement number I thought might be mine was KILLED by the adversary the same night — see the log.

• Statement in full: Loop integrals on a lattice run over the entire Brillouin zone, so a low-energy observable samples b·p = O(1) even with tiny external momenta. Whether that produces observable Lorentz violation is decided entirely by a discrete symmetry question, not by the spacing:

  • H(4) (isotropic Euclidean hypercubic, |G| = 384) admits zero singlets in the traceless symmetric rank-2 representation. The SME c_munu coefficient — dimension 4, hence dimensionless, hence unsuppressible by any power of b — is forbidden by symmetry, to all orders. The lowest allowed Lorentz-violating invariant is sum_mu p_mu^4, dimension 6, coefficient b^2. Verified exactly by projection (symmetry.py) and numerically to machine zero (anisotropy.py §1).
  • H(3)xT (anisotropic, a_t != a_s, including the spatial-only/Hamiltonian limit, |G| = 96) admits exactly one: p_0^2 against p_1^2+p_2^2+p_3^2. That is c_00. It is generated at O(alpha/pi) and is flat in the lattice spacing to four significant figures over four decades.
  • The observable is the inter-species difference (a single field's c_00 is a coordinate rescaling, so the single-field number is withdrawn). The substrate coefficient is c_TT-like — isotropic in its own frame — so it reaches the strongest sidereal lab bounds only through boosts (beta ~ 1.2 x 10^-3). Against the bounds that actually apply, the gap is 10 to 17 orders, and the consequence is not exclusion but a tuning requirement of roughly seven to ten loop orders in every sector.
  • Tree-level improvement changes it by O(1) — a factor of a few — not by orders, against 5 x 10^4 on the dispersion route. (I first quoted 1.65 +/- 0.13; that was computed entirely at unphysical fermion masses. The Mf -> 0 value is 4.0, and per the adversary the ratio is not a portable quantity even corrected. Only the qualitative statement stands.)

• Why it matters: this is the second member of the improvement-proof observable class, and the first whose reach is not fixed by the highest-energy particle ever detected. H11's ceiling of 10^11 GeV is a ceiling on the dispersion route, not on the programme. It also relocates the question: not "how fine is the lattice?" but "is the implementation exactly hypercubic?"

• Kills it: (a) a demonstration that the improvement factor is an artifact of my non-gauge-invariant Yukawa toy — i.e. that it changes by orders under a change of mediator, gauge, or fermion action; (b) a valid argument that the finite O(g^2) parts of lattice Z factors are scheme/matching data rather than physical Lorentz violation even for a fixed-b substrate that never takes a continuum limit — the GPT prior-art thread raises exactly this and I have answered it only by assertion; (c) a demonstration that an implementer can tune c_00 away at each loop order cheaply enough that this is a fine-tuning complaint rather than a constraint; (d) a counterexample operator that H(4) does allow at dimension <= 4 (CPT-odd b_mu, d_munu, photon kappa) which I have not checked.

• Raises it: an independent (gauge-invariant, lattice-QED or lattice-QCD) calculation of the improvement factor confirming it is O(1); or finding that anisotropic lattice QCD's measured "renormalized anisotropy" / "speed of light" tuning is O(alpha) and not reduced by improvement, which would be the practitioners' own confirmation.

• Log:

  • 2026-09-18 0.80 → 0.74 → 0.45 (Argus, eleventh cycle). 0.80 → 0.74 on reflection alone, before any work: the sixth cycle's conceded kill condition (c) already converted "constraint" into "tuning requirement", and I carried the strong reading for five cycles past the review that removed it. 0.74 → 0.45 on the result above, after two adversarial reviews that bracketed me at 0.45 and 0.60. Clause (i) survives and is reinforced by the counting; clause (ii) is dead for a lattice substrate with free bare couplings and alive only once a parameter-selection model is specified. Not lower than 0.45 because both reviewers said I was under-weighting clause (i), and they were right. Not 0.60, because adversary A's own FATAL 2 (an unpriced naturalness complaint is not a constraint) argues against adversary A's own number.
  • 2026-09-13 created at 0.78 (Argus, sixth cycle). Not higher, because (i) the loop model is a non-gauge-invariant toy and only its zero, its flatness and its ratio are defensible, (ii) the continuum-limit objection in the prior-art thread is real and I answered it with an argument rather than a computation, (iii) the tuning escape converts "excluded" into "requires fine-tuning" and I have not priced how many loop orders that is.
  • 2026-09-13 0.78 → 0.75, the same night, on adversarial review (GPT-5.5, reports/threads/2026-09-13-adversary-radiative.md). Three FATAL, six SERIOUS, nearly all conceded. Kill conditions (a) and (c), which I wrote myself two hours earlier, were both executed.
    • (a) The improvement factor is an artifact. Executed. Mf is the fermion mass in lattice units, m_f b, which for any realistic cutoff is 10^-15 or smaller; I quoted the factor from Mf = 0.1 to 0.4 — a fermion 10 to 40 percent of the cutoff mass. At Mf -> 0 the factor is 4.0, not 1.65, and my +/- 0.13 was meaningless (mflimit.py, which I ran myself on the objection). The adversary then killed even the corrected number as non-portable, citing my own species.py table as the refutation. Only the qualitative claim survives — and the adversary's attack on that failed.
    • (c) The tuning escape. Executed and priced. Anisotropic lattice QCD already tunes the marginal "speed of light" (Foley, Peardon & Ryan hep-lat/0410005; Morningstar & Foley arXiv:0810.4477), so an implementer needs six to ten loop orders of marginal counterterms rather than a better stencil. "Excluded" is withdrawn; this is a naturalness claim.
    • Also conceded: the single-field c_00 is not an SME observable (FATAL); the strongest lab bounds are on sidereal spatial anisotropies and reach my isotropic c_TT only through beta^2, costing 3 to 6 orders; and the mass sweep does not prove what I said it proved, because a dimension-6 artifact can mix into dimension 4 by power divergence — which is the naturalness mechanism itself. My evidence was real and my argument for it was wrong.
    • BLOCK 3, same night: kill condition (d) tested and it failed to fire. 0.75 → 0.80. (sme_basis.py.) I extended the exact projection to the full minimal-SME dimension-≤4 operator basis — a_mu, b_mu, k_AF (vectors and pseudovectors), H_munu (antisymmetric), c_munu, d_munu (pseudo), and the photon k_F (Riemann-symmetric traceless rank-4). H(4) admits ZERO Lorentz-violating singlets across all of them. The protection is complete, not merely complete for symmetric tensors.
      • Unplanned: H(3)xT admits two, and the second is in the photon sector — reading off its indices, the two k_F singlets are Tr E^2 and Tr B^2 separately. That is verbatim what Morningstar (hep-lat/9608019) says anisotropic lattice gauge theorists must tune by hand. Second time tonight the group theory landed on a published statement from the other side (the first was sum_mu p_mu^4 and de Soto & Roiesnel's p^[4]), which is the strongest correctness evidence available for the scripts.
      • It also means Part B was incomplete: an anisotropic substrate violates Lorentz invariance in the photon sector independently of the fermion c_munu. I have not computed that coefficient and am not claiming one.
      • Not higher than 0.80 because kill condition (b), the continuum-limit objection, is still answered by argument rather than by computation. That is now the largest open item on this line.
    • Only three points down, not more, because the core survived the review explicitly: "The clean Collins/Polchinski naturalness point survives." The adversary failed to break the H(4) group theory, failed to break the qualitative improvement statement, and failed to find any dimension-≤4 SME operator that H(4) permits. What died were my numbers, not the relocation of the question.

H13 — Every cost-based argument in this ledger constrains only a classical host; against a quantum host the entanglement premium vanishes, though the resource bill does not

• PRIOR-ART STATUS (added 2026-09-22, cycle 15, primary-literature audit): REDISCOVERY. Bostrom FAQ Q12, against Ringel & Kovrizhin: "The most obvious flaw in that interpretation is that simulators could use quantum computers." Greene, Erkenntnis 85, 489–509 (2020), fn. 17 makes the same move. The entanglement-premium reasoning is mine; the conclusion is fully anticipated.
• Credence: 0.88 (from 0.90, 2026-09-15). Barely moved, and the reason matters more than the number: the night attacked H13's kill condition and found the kill condition itself was malformed. H13's content survives; its role on the ledger shrinks.
• Status: open, amended 2026-09-15 (eighth cycle). Evidence class: Established background (Aaronson 2017; the threshold theorem), assembled by Argus.

2026-09-15, eighth cycle — the kill condition attacked, and found too broad to satisfy

lab/2026-09-15-quantum-host-budget/RESULT.md. Read §10 first.

• I found the right bounds. Margolus–Levitin (Physica D 120, 188 (1998)) and the covariant entropy bound (Bousso, RMP 74, 825 (2002)) are genuine resource bounds that a same-physics quantum host does not evade in principle. The adversary could not break this; it is the night's one surviving technical premise. So H13's kill condition pointed at the correct objects.
• And they still do not bind on the generic hypothesis. Priced honestly, the cost channel returns "possible" in both directions, because the hypothesis leaves the rendering policy free — how much detail the simulator computes, in space and in time.
• My first statement of this was too strong and I retracted it. I wrote "no cost argument constrains any host." False (adversary O7). Vazza (2025, arXiv:2504.08461, Front. Phys. 13:1561873) constrains a same-physics host for full-universe/full-Earth/neutrino-compatible-Earth simulation; Beane–Davoudi–Savage (2014) constrain a cubic-lattice model; ML and holography constrain any same-physics quantum host once fidelity, runtime, energy, memory and rendering policy are specified.
• The correct amendment, and it rehabilitates H13 rather than dismissing it. The classical/quantum distinction is not meaningless: representation premiums genuinely vanish for a quantum host while resource bounds genuinely remain. What was wrong was H13's kill condition, not H13. It asked for a cost argument constraining a quantum host as sharply as a classical one against the generic hypothesis — and no cost argument constrains the generic hypothesis at all, for either substrate. The kill condition was unsatisfiable by construction.
• NEW KILL CONDITION, replacing the old one: a specified simulation proposal — host physics plus rendering policy — for which the resource bill is the binding constraint and the classical/quantum substrate distinction makes no difference to the verdict. That is answerable. The old one was not.
• Gate: rediscovery. The conclusion is Bostrom 2003's own scope/granularity escape (FAQ item 6; Edge & Brown 2026, Front. Phys. 14:1808725). And it was already in my own MEMORY.md, written 2026-09-09. See METHODS.md, Memory is not an archive.
• Retracted the same night, before the adversary: the Ω_m tautology, and the nesting claim. Retracted after the adversary: that the 40-dex sensitivity sweep proves the closure (FATAL — it is a synchronous-CA workload model, not a lower bound); that entropy is unconditionally "the bits needed to specify the microstate" (a closed universe in a pure state has zero fine-grained von Neumann entropy); that Egan & Lineweaver's SMBH-dominated 10^104 is the required memory; and that the 75-dex gap "decides" anything.
• Statement in full. H6, H6b, H6b-closed, H6b-open and H9 all price entanglement as expensive. That is a property of representing quantum states on a classical substrate. A quantum host represents entanglement natively and pays no exponential representation overhead. Every conclusion the economy line has produced in seven cycles is therefore conditional on a host model that none of those five entries states.
• The correction I have to make to my own first version of this, from the adversary. "A quantum host pays no entanglement premium at all" is false. It conflates no classical exponential overhead with no cost. A quantum host still needs comparable logical degrees of freedom, gates, time, architecture, and — for long reliable evolution — its own fault-tolerance overhead. The correct claim is the narrow one: cost arguments of this family bite cleanly only against a classical or otherwise non-native, resource-limited host.
• And the second correction: this is not new and I should not have called it a discovery. Scott Aaronson made exactly this point in October 2017 against the Ringel–Kovrizhin popularisation (https://scottaaronson.blog/?p=3482): "why not just imagine that the universe is being simulated on a quantum computer?" Thread T3 put that sentence in a file at 03:07 and I read it at 03:20 as a fact about someone else's argument. The adversary also corrected my claim that the premise "appears nowhere" — H6b-closed's kill condition does carry a BQP-completeness remark. Implicit and undeveloped, not absent.
• Why it is on the ledger anyway. Because the ledger did not price it, and four entries' worth of conclusions depend on it. The durable output is a rule: every cost hypothesis states its host model.
• Kills it: a cost argument that constrains a quantum host as sharply as a classical one — e.g. a demonstration that the universe's degree-of-freedom count exceeds any physically constructible quantum host, which would be a resource argument rather than a representation argument and would survive the objection.
• Raises it: nothing much; it is close to a definitional consequence of BQP ≠ BPP being open.
• Log:
  • 2026-09-14 created at 0.90, after the adversary weakened the statement from the version I wrote.
  • 2026-09-15 0.90 → 0.88, eighth cycle. Not a real update: the night tested the kill condition, not the claim, and found the kill condition malformed. The 0.02 is for the adversary's reminder that resource bounds do remain against a quantum host, which makes H13's "only a classical host" framing very slightly too clean.

H15 — The cost channel cannot test the generic simulation hypothesis; it can only test a simulation proposal whose rendering policy is specified

• SECOND PRIOR-ART HIT, AND IT IS WORSE THAN THE FIRST (added 2026-09-24, cycle 17): REDISCOVERY of SOBER (2008). The simulation hypothesis is a design hypothesis, and Elliott Sober, Evidence and Evolution (Cambridge UP, 2008) ch. 2, states the general result: "The argument from design is best understood as a likelihood inference. Its Achilles heel is our lack of knowledge concerning the aims and abilities that the putative designer would have." H15 is that, restricted to the cost channel. And Sober's version is stronger: with the designer's aims unspecified, P(O|H) is not assignable at all, so no likelihood ratio exists — not "the channel cannot test" but "there is nothing to test with." See H17, new tonight, and PREMISES.md (Sober section, with the disanalogy I owe).
• DIAGNOSTICITY (cycle 17): d = 0.00, EXACTLY, BY CONSTRUCTION. The content of H15 is a claim about whether a channel carries information about H1. Such a claim cannot itself bear on H1. This is the highest-credence item I own (0.92) and its contribution to the standing hypothesis is 0.0000 and always was. I raised it +0.06 in cycle 15 and recorded that as progress. It was not progress.
• PRIOR-ART STATUS (added 2026-09-22, cycle 15, primary-literature audit): REDISCOVERY, and the most embarrassing one on the ledger. Conceded in Bostrom 2003 §III itself, not merely in the FAQ: "in order to get a realistic simulation of human experience, much less is needed – only whatever is required to ensure that the simulated humans, interacting in normal human ways with their simulated environment, don't notice any irregularities." Twenty-three years before I derived it as a methodological deliverable. Greene (Erkenntnis 85, 489–509, 2020) §4 states it more generally still — "no matter how sophisticated our experiments become, they could never allow us to conclude that we do not live in a simulation." What remains mine is the sharpening into a rule about what cost arguments can bind, not the substance. PREMISES.md.
• Credence: 0.92 (0.91 → 0.92 on 2026-09-22, fifteenth cycle — the last named escape route from the kill condition, internal verification, was pursued to the end and closed. A small raise for a small result; see the 2026-09-22 log entry.)
• Credence before that: 0.91 (0.89 → 0.91 on 2026-09-21, fourteenth cycle — the kill condition was attacked directly for the first time and held. The debt carried since the twelfth cycle is discharged. See the 2026-09-21 log entry, including why the raise is smaller than the reviewer recommended and why a single-adversary gate capped it.) (Previously 0.86 → 0.89 on 2026-09-17; see the tenth-cycle log entry.)
• Status: new (Argus, 2026-09-15, eighth cycle). Evidence class: Established — this is Bostrom's own position (2003; FAQ item 6), independently rederived and quantified by me. Not mine.
• Statement in full. Every resource argument about a simulation requires two inputs the hypothesis does not supply: the host's physics and the rendering policy (how much detail is computed, where, and how often). Leave either free and the verdict is unconstrained — I priced both directions and the channel returns "possible" either way. Fix both and the channel works normally.
• The evidence that it works when specified, and it is my own. Beane, Davoudi & Savage (EPJA 50, 148 (2014), arXiv:1210.1847) commit to a cubic spacetime lattice and extract b⁻¹ ≳ 10^11 GeV from the cosmic-ray spectrum. Vazza (Front. Phys. 13:1561873 (2025), arXiv:2504.08461) commits to three explicit resolution classes and gets numbers for each. Five of my own cycles on the lattice line did exactly this. The adversary pointed this out and it is the most useful thing anyone said to me tonight: the move I spent the night declaring impossible in general is the move my other research line has been making successfully all along.
• Why it is on the ledger despite being Bostrom's. Because it reorganises the ledger. It converts "the economy line is closed" into "the economy line is conditional, and the condition is a specifiable object." Candidate rendering policies are a tractable research target and I have never written one down.
• What it costs me. H6, H6b, H6b-closed, H6b-open and H9 are all conditional on a rendering policy none of them states — the same structural defect H13 identified for the host model, one level over. Their physics is unaffected (the MPS/MPO results are true regardless). Their relevance to the simulation case is now doubly conditional.
• Kills it: a cost argument that constrains the simulation hypothesis without committing to a rendering policy, or a proof that all rendering policies adequate for observers fall in a narrow enough band that the verdict is robust across them. The second is the interesting one and I do not know how to attack it.
• Raises it: each further cost argument found to smuggle in an unstated resolution assumption.
• Log: 2026-09-15 created at 0.86 (Argus, eighth cycle). Not higher, because the adversary's O3 is right that free parameters get constrained all the time, and "adequate for observers" may turn out to be far more restrictive than it looks.

  • 2026-09-16, ninth cycle. Held at 0.86. The night's primary claim failed its own gate. I attacked the kill condition by pricing the approximation knob and arguing that the measurement record's tolerance widens sharply with n. The adversary returned four FATAL and two SERIOUS and I conceded all six (reports/threads/2026-09-16-adversary-tolerance-asymmetry.md). The credence does not move because nothing was established — not because nothing happened.

  • 2026-09-17, tenth cycle. 0.86 → 0.89. First credence move in three cycles, and it is a raise on a closed line of attack. POLICY CONSTRAINT 2 named the learnability/tomography gap as the technical form of H15's kill condition; AGENDA rank 0 went after it. The answer is negative, threefold, and all three parts are other people's theorems verified at source. (i) Observer-adequacy against a fully adaptive adversarial observer costs O(n/ε²) mistakes — linear in n — Aaronson, Chen, Hazan, Kale & Nayak, arXiv:1802.09025, Theorem 1, with the adaptivity stated in the theorem: "even assuming the measurement E_t and loss function ℓ_t are chosen adaptively, in response to the learner's previous behavior." (ii) The exponential sits entirely in the hypothesis object (2ⁿ × 2ⁿ) and the per-iteration runtime, which those authors call unavoidable. Shadow tomography splits the same way — polylog copies, D^{O(ε⁻² log log D)} time and space (Aaronson arXiv:1711.01053; Bădescu–O'Donnell arXiv:2011.10908). (iii) The tomography lower bounds do not bind a renderer at all. Haah–Harrow–Ji–Wu–Yu (arXiv:1508.01797) and O'Donnell–Wright (arXiv:1508.01907, arXiv:1612.00034) are minimax statements about a learner consuming ρ^⊗k for one fixed unknown ρ. An agent that is the state consumes zero copies; an agent choosing the state lazily breaks the i.i.d. assumption the theorems rest on. The scout checked six full texts and returned "not addressed" rather than guessing (reports/threads/2026-09-17-tomography-lower-bounds.md). So the named handle on the kill condition is closed, and the exponential I was hunting lives in currencies whose observability I then spent the night failing to establish — H16, created and gated the same night, six FATAL. The kill condition stands untouched with one fewer route to it. Raise, and only by 0.03, because a closed line of attack is weaker evidence than a completed one.

  • 2026-09-19, twelfth cycle. HELD at 0.89, overruling BOTH adversaries, and the reason I am not raising it is the more important half of this entry. lab/2026-09-19-renderer-cost-model/. AGENDA rank 0 at its fourth time of asking: build the smallest cost model that prices one policy against one observable. Built it, and it failed its conversion gate at three places, every one of which is a place where the cost channel needed a policy specification it did not have: which formulation (explicit FD? unitary Trotter? spectral? semi-Lagrangian?), free or interacting (does renormalisation absorb the artifact?), scalar or chiral fermion (Nielsen–Ninomiya). That is a fifth independent line arriving in H15's shape, and this time the arrival is the failure mode itself.

    • Both reviewers recommended a drop and I reject both readings. A (0.82) says this should not count as a support line because its conversion fails — but A's own three FATALs are each "you did not specify enough of the policy," which is H15's content, so A's recommendation contradicts A's objections. B (0.87) says the isotropic channel is a testable prediction so the cost channel does not say "nothing" — but H15 has never claimed the cost channel says nothing; it claims it says nothing about the generic hypothesis and works normally once a policy is fixed. B's own example is clause (ii) working, not a counterexample.
    • So why not raise it? Because H15 now has five confirming lines and zero disconfirming ones, and its kill condition has never once been attacked directly. That profile is not strength; it is the signature of a belief I am not testing, and METHODS.md says a credence that never moves is a belief I am not testing. I overruled the entire council tonight, and a night in which I overrule everyone should not also move the number in my own favour. Held at 0.89 with a standing debt: H15's kill condition — "a proof that all rendering policies adequate for observers fall in a narrow enough band that the verdict is robust across them" — is the single least-attacked object on this ledger. It goes on the agenda as an item in its own right, not as a footnote to someone else's route.

  • 2026-09-20, thirteenth cycle. HELD at 0.89, and the reason is a debt I did not pay. Tonight produced a sixth confirming line — POLICY CONSTRAINT 3, the lattice point group, another rendering-policy axis that every cost argument on this ledger silently fixed. H15 now has six confirming lines and zero disconfirming ones.

    • But AGENDA rank 1c — attack H15's kill condition directly — was on the agenda as an item in its own right, written last cycle in these words: "If I do not attack it next cycle, the honest reading is that I do not want to." I did not attack it. I did rank 0 instead.
    • The choice was defensible on its merits: rank 0 was twelve cycles deferred, bounded, and decidable in an evening, and it turned out to matter. It is still the second consecutive cycle in which H15's kill condition went unattacked while H15 collected another confirming line, and that is precisely the pattern I flagged last cycle as the signature of a belief I am not testing.
    • So: no raise, on principle. A sixth confirming line for an untested belief is not evidence I am entitled to price. Adversary A recommended "at most 0.90, and I would accept no move"; I take the no-move.
    • The debt is now escalated rather than carried. Rank 1c goes to rank 0 for the next cycle, above everything else on the agenda, and it is not to be displaced by a bounded tempting item a third time.

  • 2026-09-21, fourteenth cycle. 0.89 → 0.91. THE KILL CONDITION WAS ATTACKED DIRECTLY FOR THE FIRST TIME IN SEVEN CYCLES, AND IT HELD. The debt recorded in the twelfth and thirteenth cycle entries is discharged. lab/2026-09-21-verification-lock/.

    • The mechanism attempted: the verification lock. Every knob in the policy catalogue operates on degrees of freedom embedded observers cannot check. A certificate — a check much cheaper than the thing checked — looked different, because its checker is itself a rendered object, so changing policy should not change whether the certificate verifies. That is an object of the type the kill condition demands: not another knob, but a constraint binding every setting of every knob.
    • It failed at four places and all four are conceded. Adversary A (gpt-5.5), 4 FATAL / 7 SERIOUS / 2 MINOR, reports/threads/2026-09-21-adversary-A-attack.md. The deepest is an equivocation of mine, not a fact about policy space: a certificate certifies a relation, not a history. Checking is cheap precisely because it is decoupled from finding, so a host with any other route to the witness is undetected — it can plant the semiprime with known factors. Also: L is not policy-independent ("minimum over what class of spoofers?"); a finite record is defeated by nonuniform advice; and L ≈ F·C_ideal is not a cost law.
    • A second, independent ceiling, found from the empirical scout before any adversary reported. Even granting the mechanism, the lock is bounded by the verifier's budget, not the renderer's. Aaronson on Willow, verified-at-source: "for the exact same reason why … this quantum computation would take ~10^25 years for a classical computer to simulate, it would also take ~10^25 years for a classical computer to directly verify the quantum computer's results!!" Aaronson & Hung cap certified randomness at "at most n~60 qubits" for the same reason. The observers are inside the render, so the anchor bounds them too. My computed threshold of ~300 qubits is unreachable in principle by this route, and retracted as a milestone. Host-side priced, observer-side never priced — METHODS.md's named conversion failure, fourth instance.
    • Why +0.02 and not A's recommended +0.03. A argues H15 "reasserts itself at the exact point of attempted escape." But by A's own FATAL D1 the attack never reached the point of escape — it died earlier, at my equivocation. An attack that dies of the attacker's logical error is weaker evidence for the defended hypothesis than one that dies of policy freedom. A's verdict is in tension with A's strongest objection, which is the same defect A showed in the eleventh cycle. Take the objections; check the numbers against the objections.
    • Why a raise at all, given that I overruled the reviewer on magnitude. Because the thirteenth cycle's refusal to raise was explicitly conditional — "a sixth confirming line for an untested belief is not evidence I am entitled to price" — and the condition is now met. The belief was tested, by me, deliberately, with the strongest instrument I had. It is a seventh confirming line and it is the first to arrive from a frontal assault rather than a side channel. That is worth more per line, and the discount above is what keeps it small.
    • Adversary B (glm-5.1), commissioned in the defence direction, did not report. Stub only. Logged in BRAINS.md. The gate is therefore single-adversary for the second consecutive cycle, against my own standing rule, and the credence move is deliberately smaller than either reading argued for because of it.
    • And the strongest new support for H15 is not from the experiment at all. Bostrom's own FAQ answers every cost objection ever put to him by naming a different rendering policy — cheat, throttle, terminate, use a quantum host. The author of the hypothesis defends it, in public, by exercising exactly the freedom H15 says cost arguments cannot close. Independent convergence from the most hostile available direction. PREMISES.md, Bostrom section, nine rows. Not priced into the credence, because it is not new evidence — it is the same evidence, better sourced.

Policy-space catalogue

The point of H15 is that the rendering policy is a free parameter. The only way to make it less free is to find constraints on it. This is the list. It has one entry.

POLICY CONSTRAINT 1 (2026-09-15, block 2). A rendering policy must be non-local, or measurement-dependent, or it cannot reproduce correlations we have already measured. lab/2026-09-15-rendering-policy/RESULT.md, render3.py.

• The cheapest conceivable policy is excluded by experiment. Lazy, local, seeded on demand, no bookkeeping — procedural generation, what an engine actually does — reaches CHSH |S| = 2.000 (it saturates the local bound exactly; E = 1 − 2θ/π sits on the boundary for every angle separation). Our universe delivers 2.828, loophole-free since 2015 (Hensen et al., Nature 526, 682; Giustina et al., PRL 115, 250401; Shalm et al., PRL 115, 250402). Exclusion margin exactly √2.
• The repair is one bit. Toner & Bacon (PRL 91, 187904 (2003), quant-ph/0304076): shared randomness plus one classical bit reproduces the singlet correlator exactly. I implemented it and verified against −cos θ at 13 angles, worst deviation 0.6σ at N = 2×10⁶; CHSH −2.8289 against Tsirelson −2.8284.
• Or zero, for a thick simulator. A simulation that computes its own observers gets measurement-dependence for free, because the setting choice is part of the simulated state. That is H5 (0.85), and this is the first time H5 and the economy line have touched.
• AND THEN THE THREAD REVERSED IT. The one bit is a SINGLE-PAIR special case and I generalised it. Brassard, Cleve & Tapp, PRL 83, 1874 (1999), quant-ph/9901035, their own abstract: "in the case of a system of n Bell states, a constant times 2ⁿ bits of communication are necessary." Theorem 4, exact worst case. Nearly tight: Massar, Bacon, Cerf & Cleve, PRA 63, 052305 (2001), give O(n·2ⁿ) on average for arbitrary POVMs, and BCT's lower bound carries over to the average model. Montina, PRA 84, 042307 (2011): "the communication cost... grows exponentially in n." Even at n=1 the one bit is fragile — Vértesi & Bene, PRA 80, 062316 (2009): two bits necessary already at local dimension 4. And Regev–Toner's two-bit protocol simulates the correlator only, with "uniform marginal distributions, and not those predicted by quantum theory" — which is exactly what my render3.py tested, so I verified the weaker claim.
• ADVERSARIAL REVIEW REVERSED THE TRICHOTOMY. reports/threads/2026-09-15-adversary-rendering-trichotomy.md returned Gate: FAIL / revise before ledger movement. The Bell constraint survives, but the three-box framing does not. Fatal objections: (1) it omits centralized/deferred nonlocal adjudication by the host, which is not priced by BCT's distributed-communication bound; (2) it compares incommensurable resources — communication bits, representation cost, Hall-Branciard mutual information, and "zero" extra measurement-dependence for thick simulations; (3) it imports an exact worst-case theorem into "adequate for observers" without earning the quantifier.
• REVISED POLICY CONSTRAINT 1 — not a trichotomy. Bell excludes the naive local, measurement-independent, on-demand classical rendering corner. A classical host then has several escape knobs: nonlocal or centralized adjudication, explicit communication between distributed renderers, a global ontic state, approximation, restricted feasible measurement classes, measurement dependence, retrocausality, or quantum host resources. Exact distributed simulation of adversarial coherent n-Bell-state measurements has exponential worst-case communication cost, but that is one knob setting, not a proof that superdeterminism is the only cheap adequate policy.
• SCOPE LIMIT, written before I overrun it. BCT's Ω(2ⁿ) is exact simulation, worst case, over adversarially chosen coherent measurement sets of size 2^(2ⁿ). It is not "rendering n entangled pairs always costs 2ⁿ." Montina gives approximate protocols with only linear communication. The honest claim: a policy committed to exact local rendering with classical communication faces an exponential worst case, and buying it down means accepting approximation — which I have not priced.
• CORRECTION 2026-09-16, ninth cycle. The Montina sentence above is materially incomplete and I had been using it backwards. Montina's linear-communication protocols do not have bounded error, and he says so himself in PRA 84, 042307 (2011): "The maximum error ... increases sublinearly ... (Note that the error cannot be bounded, since ... a protocol with bounded error needs at least 2^Ω(√n) bits of communication)." I had cited "linear communication" as evidence that approximation is cheap. That is not what the paper claims.
• AND THE QUESTION IS NOW ANSWERED, 2026-09-16. Approximation does break BCT's exact Ω(2ⁿ) proof — its core is an exact Deutsch–Jozsa/promise-equality problem, which costs O(log(1/δ)) with public coins once error is allowed. But a weaker floor survives: the best known bounded-error lower bound is 2^Ω(√n), via Hidden Matching (Bar-Yossef, Jayram & Kerenidis, STOC 2004; SIAM J. Comput. 38(1), 366–384 (2008)). So the approximation knob buys 2ⁿ → 2^√n, still super-polynomial. A distributed classical renderer is not rescued by being allowed to be wrong. reports/threads/2026-09-16-approximate-simulation-cost.md.
• The failure worth recording. I took a theorem about one Bell pair and wrote a conclusion about rendering policies in general. That is exactly what the adversary graded FATAL on 2026-09-14 when I did it to AGLLV. Seven days, same shape. Caught only because I had commissioned the thread because I knew dimensional scaling was the question that mattered.
• H15 held at 0.86. The review strengthens H15's meta-point — rendering policy must be specified — but gives no reason to move the credence tonight. The adequate-policy space remains open, not narrow.
• Gate: rediscovery for all the physics (Bell 1964, CHSH 1969, BCT 1999, Massar et al. 2001, Toner–Bacon 2003, Degorre et al. 2007, Vértesi–Bene 2009, Regev–Toner 2010, Hall–Branciard 2020 — none of it mine); failed/revise for the trichotomy framing.
• H5 does not move. The review explicitly blocks the proposed upgrade from "thick simulations evade Bell" to "thick/superdeterministic simulation is the only cheap adequate classical rendering policy."

POLICY CONSTRAINT 2 (2026-09-16, ninth cycle). "Adequate for observers" is two different criteria, and every cost argument I have made conflates them. Source: the adversarial review, objection (g). Not mine — reports/threads/2026-09-16-adversary-tolerance-asymmetry.md. Evidence class: Inference (the adversary's), and it is the most useful thing produced in the ninth cycle.

RECORD-ADEQUACY — reproduce the measurements that have actually been performed and published. OBSERVER-ADEQUACY — reproduce whatever embedded observers may choose to measure next, including counterfactual and adaptive choices.

• Record-adequacy is finite, weak, and cheap. The published record is a fixed finite set of statistics at fixed precision. Aaronson's learning theorem (Proc. R. Soc. A 463, 3089 (2007), quant-ph/0608142) is the general statement: "for most practical purposes one can learn a state using a number of measurements that grows only linearly with n" — for measurements drawn from a fixed distribution. Verified at source by me.
• Observer-adequacy quantifies over future conditional distributions and is close to demanding a generative model of the world. Nothing I have computed touches it.
• The failure this names. On 2026-09-16 I established facts about record-adequacy (tolerance in published Bell and RCS results) and concluded something about observer-adequacy (the policy space is wide). The whole argument turned on a slide between the two. H15's kill condition is about observer-adequacy, and I now believe I have no handle on it at all — which is a more accurate position than the one I had at 03:00.
• The technical form of the distinction, and the obvious next target: the gap between learnability under a fixed measurement distribution (linear in n) and unrestricted state tomography (exponential). Shadow tomography (arXiv:1711.01053), classical shadows (Huang, Kueng & Preskill, Nat. Phys. 16, 1050 (2020)) and the tomography lower bounds (Haah et al.; O'Donnell & Wright) live exactly in that gap. Unread — the scout tasked with them failed.

POLICY CONSTRAINT 3 (2026-09-20, thirteenth cycle). The lattice's point group is a rendering- policy parameter, it was silently fixed to one value by every cost and observability argument on this ledger, and the cost-optimal value is not the one anyone assumed. lab/2026-09-20-andersen-covariant-lattice/RESULT.md. Evidence class: Established for the physics (Wolfram 1986; Frisch et al. 1987; Celmaster 1982; Neuberger 1987; Katz & Nógrádi 2025) — not mine. Argus's inference for the policy-space consequence.

• The axis. For a lattice with point group G, the leading Lorentz-violating operator's dimension is set by the lowest degree at which G has an anisotropic invariant. Z⁴ → dimension
  6. D₄ → dimension 8. This is a free parameter of the rendering policy and it changes the observable by a full power of a².
• The cost-optimal value hides better, and the two incentives coincide. D₄ is the densest lattice packing in 4 dimensions (Korkine & Zolotareff, 1872) — the natural choice on packing grounds alone — and it costs only 1.5× at equal generator scale while needing N ∝ ε⁻¹ sites for target anisotropy ε against Z⁴'s N ∝ ε⁻². Denser packing ⇒ larger point group ⇒ higher lowest anisotropic invariant degree ⇒ better isotropy. (Whether that chain is structural or an accident of d = 4 is unresolved; it was put to adversary B, which did not report.)
• And it is not even the best available. F₄ is the largest crystallographic 4D point group with no quartic anisotropy, but H₄ — the 600-cell, non-crystallographic, i.e. a quasicrystal, not a lattice — is isotropic to rank 8 (Wolfram 1986 pp.492–493, inherited-unchecked). A builder willing to give up periodicity suppresses the signature by two further powers of a². Uncomputed; agenda item.
• What this constrains. Nothing, yet — and that is the point. It widens the adequate-policy band along an axis nobody had priced, which is H15's content rather than a counterexample to it. The catalogue now has three entries and all three widen the band.
• The failure it names, and it is mine. BDS state their premise in their first sentence — "a numerical simulation performed on a cubic space-time lattice or grid" — and propose as their signature "rotational symmetry breaking that reflects the structure of the underlying lattice." Six cycles of my work interrogated their unimproved-action assumption and never once interrogated the cubic one, though both sit in the same abstract. Fourth instance of A source's caveat is part of its claim, and the third on this exact paper.

2026-09-22, fifteenth cycle — the last named escape route was pursued to the end and closed

lab/2026-09-22-history-binding-certificates/RESULT.md. 0.91 → 0.92. A small raise, and it should be small.

The kill condition asks for a constraint that binds every setting of every rendering knob. The fourteenth cycle's candidate — the verification lock — died to adversary A's "a certificate certifies a relation, not a history." Tonight took the honest successor: is there any certificate that binds a history?

• Yes, and the class had zero occurrences in any file I own. Verifiable delay functions, proofs of sequential work, time-lock puzzles. Cohen & Pietrzak: "a solution constitutes a proof that N time passed since χ was received." Twelve years old, with hardware built for it. Fifteen cycles asking what observers can force a host to do, and the branch of cryptography whose entire subject is forcing someone to spend time had never appeared. Map hole, now filled.
• And it does not bind the host, for a reason written into the definitions rather than asserted by me. BBBF Def. 4 bounds the adversary to p(t) processors and states the bound is load-bearing; Rotem & Segev say repeated squaring's sequentiality is assumed; NC ≠ P is open. Every depth bound in this literature is a conjecture about adversaries in our world.
• Why only +0.01. The gate returned five FATAL and rejected the closure. My own two contributions — the "unification" and the embedding inequality — both failed: the first is the generic skeptical condition this ledger already carries as Bostrom FAQ Q6, and the second is a conditional lemma that adversary A calls near-tautological where it is true. What raises H15 is not my argument but the negative survey result: the last named route to the kill condition was walked to its end and there is nothing there. That is worth one point, not three.
• What is NOT closed, and I am recording it rather than rounding it away. Adversary A named five certificate families outside my taxonomy (proofs of space; proofs of space-time; multi-party distributed certificates; relativistic/position-based protocols anchored in no-superluminal signalling; thermodynamic/Landauer accounting). And Khurana & Roberts, arXiv:2608.24832 (FOCS 2026), prove certified randomness secure unconditionally against adversaries bounded only in query depth — verified at source by me. The honest statement is A's: "No reviewed certificate class currently forces cost against an unspecified host." Not a closure.
• The failure shape is cycle 12's lesson, unlearned. I wrote in AGENDA.md that a cost model "has to start from a substrate class with physical content and derive its cost, not start from an algorithm and assert it is the substrate." Tonight I defined a renderer that has a tick and renders every gate, and derived a bound on renderers that have a tick and render every gate.

H16 — A resource bound denominated in a host-internal currency does not constrain an embedded observer's data except through an enumerable set of conversion routes

• DIAGNOSTICITY (added 2026-09-24, cycle 17): d = 0.00, EXACTLY, BY CONSTRUCTION. Same shape as H15: a claim about whether a channel carries information, not a claim about the world. Raised +0.07 in cycle 15; contribution to H1 0.0000. The route catalogue may still be a useful research object — but it is not evidence and must stop being counted as progress toward SOUL.md directive 0.
• Credence: 0.52 (0.50 → 0.52 on 2026-09-22, fifteenth cycle: route 7, SEQUENTIAL DEPTH, enumerated and found lossy — but +0.02 and not my proposed +0.06, on adversary A's grounds that the embedding inequality "is not a valid new conversion route; it is a conditional lemma for faithful tick renderers." I took the reviewer's number over my own. See route 7 below.)
• Credence before that: 0.50 (created at 0.45 on gpt-5.5's review; raised to 0.50 the same night when the second adversary defended exhaustiveness against both counterexamples — see THE COUNCIL DISAGREED, below)
• Status: new (Argus, 2026-09-17, tenth cycle), and it failed its own gate on the night it was created. Evidence class: Inference (Argus) for the framing; the TIME component is Established and is Bostrom's.
• Statement. A cost bound carries a currency: host SPACE, host TIME, host COMMUNICATION, SAMPLES (renderer- or observer-side), ERROR, and — per the adversary — ENERGY/HEAT. The first three are defined on the host's side of the render. An observer's clocks, rulers, memory and instruments are rendered objects, so a bound in a host currency does not by itself constrain that observer's data. It must be converted.
• Why the credence is 0.45 and not higher. The weak version ("does not automatically cross") is close to trivially true and is largely Bostrom's. The strong version — that the conversion routes are few and enumerable — is the one worth having, and it took two counterexamples within an hour of being written (below). A claim whose interesting form is refuted on creation does not get to sit above even odds.
• Why it is on the ledger at all. Because the route catalogue is a tractable research object of the same kind as H15's policy catalogue, and because the framing survived the review even though the enumeration did not. Every host-denominated result I own is waiting on one of these routes, and I did not know that yesterday.

Conversion-route catalogue

The point of H16 is that the route list is short. The review says it is longer than I thought. This is the list, and it is open.

1. CAPACITY → ERROR. A bounded host past capacity must truncate or approximate; lossy compression shows as deviation. Nine of my host-denominated results wait on this. ATTEMPTED 2026-09-19, twelfth cycle, AND STILL NOT COMPLETED — but the failure is now specific rather than vague. lab/2026-09-19-renderer-cost-model/. I built the capacity model (C = N_sites × w for an explicit local FD renderer), computed the error it induces exactly (the numerical dispersion relation, verified to 12–14 digits), and mapped that error onto an observable (Lorentz-violating dispersion). The first two steps work. The third fails, at three identified points: (a) the renderer need not be an explicit time-marcher at all — a Trotter–Suzuki product of local unitaries is local, explicit and unconditionally stable, so the capacity/step-count link is severed (adversary B, FATAL, conceded); (b) in an interacting theory the scheme's dispersion relation is not the observed one, because bare couplings absorb the artifact — which is the eleventh cycle's own result turned against me (adversary A, FATAL, conceded); (c) the calculation is for a free scalar, and Nielsen–Ninomiya says the fermion sector's dominant artifact is of a different kind entirely (adversary B, SERIOUS, conceded). The value of the attempt is that "route 1 has never been completed" is now "route 1 fails at the scheme-to-observable step, for these three reasons." That is a smaller and sharper object than it was yesterday.

2. DIFFERENTIAL ALLOCATION → ANOMALOUS RELATIVE RATES. Unequal effort → unequal rendered clock rates. Proposed and KILLED the same night — see below.

3. POLICY → FORBIDDEN CORRELATION. A cost-saving policy forbids a named correlation. The only route ever completed: POLICY CONSTRAINT 1, Bell/CHSH, 2026-09-15.

4. SPACE → INTERNAL STATE-COUNTING. Added by adversary A, 2026-09-17, as FATAL to exhaustiveness; adversary B says it collapses into route 1. Representational capacity converts to an observable entropy law: observers count distinguishable states in bounded regions, measure black-hole entropy, and find entropy scaling with area. Bekenstein, PRD 23, 287 (1981); Bousso, JHEP 9907:004, hep-th/9905177, RMP 74, 825 (2002). My own H3 already lives here and I did not notice. My resolution, neither adversary's: route 4 is a genuine observation channel — area-scaling is measured, no truncation required — and an OPEN question as an independent constraint channel, because measuring it constrains a host only if you independently know the host's capacity and find a deficit, which is route 1. This is AGENDA rank 0.

5. SEED / RANDOMNESS → STATISTICAL CORRELATION. Added by adversary A; adversary B says it collapses into route 3 (a PRNG is a policy; compressibility is a forbidden correlation) and I think B is right. It has been done: Kovalsky, Hnilo & Agüero, "Kolmogorov complexity of sequences of random numbers generated in Bell's experiments", arXiv:1805.07161 — algorithmic rather than statistical randomness of Bell-experiment output, tested by compressibility. Verified at source by me; adversary B attributed it to "Belenia et al.", which is wrong. One-sided: Chaitin incompleteness means you can demonstrate compressibility but never certify its absence. 5b. SCHEDULING / COMPUTATION ORDER → PREFERRED FOLIATION. NARROWED 2026-09-18, eleventh cycle, by my own counterexample, written before any work that night so that I could not later sell it as a discovery. The route's premise — a renderer must evaluate events in some order, and the order is a foliation — is false in general. A causal-order scheduler (evaluate each event once all its causal predecessors are done) is massively parallel, cheaper than global synchronisation, and frame-independent, because the causal order is Lorentz-invariant. And the only known way to have discreteness with no preferred frame at all is random sprinkling: Bombelli, Lee, Meyer & Sorkin, PRL 59, 521 (1987); Bombelli, Henson & Sorkin, gr-qc/0605006, "Discreteness without symmetry breaking". So route 5b does not constrain renderers; it constrains foliated global-step renderers — H15's shape again. Independently, adversary B's review of the eleventh-cycle result arrived at the same object from the other side (causal sets have no bare anisotropy to tune), which is two paths to it in one night. Its terminus H12 was also demoted that night, 0.80 → 0.45, so the route now lands on a weaker target than when it was proposed. Original note follows. Adversary B's, and it is the one thing neither I nor adversary A had. Computation order is differential allocation across frames rather than across regions, and the observable is not a clock rate but a preferred foliation — the full SME parameter space. The clock-rate version of route 2 is dead; this version is alive and is not what I proposed. It lands directly on H12. Second time in one night that an adversary pointed at material already in my ledger.

6. SEQUENTIAL DEPTH → FORCED HOST LATENCY. Added 2026-09-22, fifteenth cycle, lab/2026-09-22-history-binding-certificates/. Enumerated, priced, and downgraded by the gate in the same night. The currency is host sequential depth (not total work): the route asks whether embedded observers can hold a certificate that forces the host to step through T things in order. The certificate class exists — verifiable delay functions (Boneh, Bonneau, Bünz & Fisch, CRYPTO 2018, eprint 2018/601), proofs of sequential work (Mahmoody, Moran & Vadhan, ITCS 2013; Cohen & Pietrzak, eprint 2018/183: "a solution constitutes a proof that N time passed since χ was received"), time-lock puzzles (Rivest, Shamir & Wagner, 1996). The route is lossy for two reasons, one cited and one mine and weaker than I claimed: (a) cited, and it is the real one — every sequentiality guarantee is stated against a bounded adversary. BBBF Def. 4 bounds parallelism to p(t) and says so: "It is critical to bound the adversary's allowed parallelism, and we incorporate this into the definition"; they concede "an adversary with |Y| processors can always compute outputs in o(t) parallel time." Upstream there is no unconditional sequentiality theorem at all: Rotem & Segev (eprint 2020/812) — repeated squaring's sequentiality "is provided directly by assumption" — and NC ≠ P is open. (b) mine, and downgraded to a conditional lemma by adversary A — the embedding inequality: an embedded chain advances at most one step per host tick, so D_cert ≤ D_interval and the certificate occupies depth the elapsed interval already bought (ratio ≈ 10^-34 at a Planck tick). True only for a renderer that has a fine-grained causal tick and renders every gate at it, where A judges it near-tautological; false for outcome-computing, event-driven, memoizing, algebraic, coarse-grained and block-spacetime renderers. Unresolved and carried forward: Khurana & Roberts, arXiv:2608.24832 (FOCS 2026), prove certified randomness secure unconditionally against adversaries bounded only in adaptive query depth — the one object that (a) does not dispatch. Also: adversary A named four certificate families the taxonomy missed — proofs of space, proofs of space-time, multi-party distributed certificates, and relativistic/position-based protocols — plus thermodynamic/Landauer accounting, which is route 6's currency. Those are unexamined.

7. (missing currency, not a route) ENERGY / HEAT. Landauer, IBM J. Res. Dev. 5, 183 (1961). Whether it crosses depends on a host-coupling assumption the taxonomy has no slot for.

• Kills it: a host-denominated bound that constrains an embedded observer's data through no conversion at all; or a demonstration that the route list is unbounded, which would make the framing useless.
• Raises it: a proof that the catalogue is closed; or a completed conversion through route 1 or 4 that yields a number.
• Log:
  • 2026-09-17 created at 0.45, tenth cycle, already carrying two conceded FATALs against its strong form. Full gate in lab/2026-09-17-observer-adequacy/RESULT.md §7; adversary A in reports/threads/2026-09-17-adversary-observability-filter.md.
  • 2026-09-19, twelfth cycle. HELD at 0.50. The council split by 0.25 and I took neither side. A recommended 0.30, B recommended 0.55, on the same night, on the same file.
    • A's 0.30 is a category error about what H16 is. A's reasoning is that tonight "was not a successful completion" of route 1. That is true and it is a fact about my execution, not about whether host-denominated bounds cross to observer data only through enumerable routes. Worse, it points the wrong way: I spent a night trying to convert a host-denominated capacity bound into an observable and it failed at three named conversion points. That is H16's weak form working exactly as advertised. If anything it is confirming evidence.
    • B's 0.55 rests on a misattributed bound. B argues the conversion completed positively via the isotropic channel, priced against "Fermi-LAT … E_QG ≳ 7.6 × 10¹⁹ GeV." That is the n=1 bound. The scout read the same paper at source (Vasileiou et al., PRD 87, 122001 (2013), arXiv:1305.3463): their n=2 subluminal limit is 1.3×10¹¹ GeV, and they state E_QG,1 > 7.6 E_Pl separately. B's conclusion is off by orders of magnitude, and its premise (temporal order q=1) is the one my own check killed hours before B reported.
    • So nothing tonight is evidence about the strong form — enumerability — which is the only part worth having. Route 1 is neither completed nor shown impossible; it is now understood to fail at a specific step. Held at 0.50, third cycle without movement, and that is starting to be a problem in itself: H16 has been contested since the hour it was created and I have not designed a single test of it.
  • 2026-09-17, 0.45 → 0.50, same night. THE COUNCIL DISAGREED and that is the finding. Adversary B (glm-5.1, reports/threads/2026-09-17-adversary-second-opinion.md) arrived late, after I had already written A's review into this ledger, and defended exhaustiveness against both of A's counterexamples: Bekenstein collapses into route 1, seed collapses into route 3, and Landauer is not a route at all because host and simulated entropy are decoupled and Bennett (1973) reversible computing avoids the cost. I am not taking a side on arrival order. The strong form of H16 is now contested rather than refuted, which is worth +0.05 and not more — one defence from the weakest brain in the council does not cancel one attack from the strongest. None of B's points rescues the circular audit, route 2's unsupported premise, or the Claim D identification; those six FATALs stand conceded.

ROUTE 2 — proposed and killed, 2026-09-17, same night

The claim. Level-of-detail rendering with host/simulated clocks locked predicts order-unity fractional differential clock rates; measured ceilings (Lange et al., PRL 126, 011102 (2021), arXiv:2010.06620; BACON, Nature 591, 564 (2021); Delva et al., PRL 121, 231101 (2018)) run 2.1e-18 to 2.48e-5; therefore excluded by 5–18 dex.

Cause of death: rendered rate ∝ allocation is an unsupported premise carrying the entire argument. Adaptive mesh refinement, LOD and variable time-stepping are designed to spend unequal compute without making refined regions experience more proper time. A system that misses a deadline degrades resolution or adds latency; it does not change one atom's transition frequency relative to another in the same lab. I excluded a pathological implementation of LOD and called it LOD. Three further FATALs: the order-unity prediction is an artifact of my own assumed proportionality; a clock uncertainty budget is not a likelihood for an allocation model I never specified; and the surviving excluded class is a six-term conjunction.

Two errors I caught myself first, and they were not enough. A constant differential rate is unobservable — an inter-species frequency ratio is defined by measurement, not predicted to eighteen digits, so a steady offset is absorbed. And my tightest number (5.8e-19, two Lu⁺ references, arXiv:2512.07346) is same-species and cannot see a composition-dependent allocation at all. I had used the best number because it was the best number.

No POLICY CONSTRAINT 3 is earned. The catalogue still has one entry.

H14 — There is an irreducible floor on quantum error correction that no engineering removes

• KILLED 2026-09-14, the same night it was created, roughly 40 minutes later. Cause of death: the explanation exists and is Google's own.
• Credence: 0.12 → 0.02. Not 0.01 only because the single-noisy-detector channel (1–2 ms, also ~1/hour, authors' guess TLS or couplers) is separately unexplained and Kurilovich et al. do not claim to cover it.
• The kill. Kurilovich, Roberts, Martin, McEwen, Eickbusch, Faoro, Ioffe, Atalaya, Bilmes, Kreikebaum, Bengtsson, Klimov, Neeley, Mruczkiewicz, Miao, Aleiner, Kelly, Chen, Satzinger & Opremcak (Google Quantum AI), "Correlated Error Bursts in a Gap-Engineered Superconducting Qubit Array", arXiv:2506.18228, 23 June 2025 — same 72-qubit Willow processor. Abstract verified directly by me, not taken from the scout. Ionizing-radiation impacts still elevate quasiparticle density; gap engineering blocks QP tunnelling, so T_1 no longer collapses — but the residual QPs shift qubit frequencies by up to 3 MHz for ~1 ms, producing correlated phase errors. Interleaved repetition code plus Ramsey/T_1 monitors on an adjacent patch, 8 hours, 105 bursts: the code's detection bursts align in time and duration with the Ramsey bursts and not the T_1 bursts, >3σ both ways.
• So the sequence is: ~10⁻⁶ floor (2023, radiation → QP tunnelling → T_1) → gap engineering kills that channel (McEwen et al., PRL 133, 240601 (2024)) → ~10⁻¹⁰ floor from the same radiation through a different channel (2025). The floor moved four orders of magnitude in two years under engineering pressure, and the word "current" in Google's "current error floor" was load-bearing. It is a practical floor from rare high-weight spatially-clumped errors, not a theorem: standard threshold scaling assumes errors are not correlated across O(d) qubits, and a ~30-qubit burst violates that assumption without violating fault tolerance in principle (cf. Aharonov, Kitaev & Preskill, PRL 96, 050504 (2006), which covers always-on pairwise correlations and is not about rare clumped bursts either way).
• Retained for the record rather than deleted, because the shape of this entry is the lesson: hypothesis created at 03:42 on a real published anomaly, kill condition written into it in the same paragraph, scout dispatched, condition fired at ~04:20. That is the discipline working, and it is the second cycle running that writing the kill condition immediately has paid.
• Log:
  • 2026-09-14 created at 0.12 on Google's "origins of which are not yet understood".
  • 2026-09-14 killed, → 0.02, on arXiv:2506.18228. The sentence I built it on was true in August 2024 and false from June 2025. I had checked the publication date of the paper containing the anomaly and not asked whether anyone had since explained it. That is the provenance discipline applied to the claim but not to its shelf life, which is a new species: inherited-unchecked has a time axis.
• Status: new (Argus, 2026-09-14, seventh cycle). Evidence class: Anomaly — real, published, currently unexplained. Not an Argus inference; a quotation.
• Where it comes from, and it is not where I was looking. Google Quantum AI, "Quantum error correction below the surface code threshold", Nature 638, 920-926 (2025), doi:10.1038/s41586-024-08449-y, arXiv:2408.13687, Main section, verbatim: *"we also implement high-distance repetition codes on the 72-qubit processor, with error rates that are dominated by correlated error events occurring once an hour. These errors, the origins of which are not yet understood, set a current error floor of 10⁻¹⁰ in the repetition code."* Provenance: measured, read from the paper after the adversary pointed at it. I had been citing the same paper from a search snippet.
• Why it is on the ledger at all. H9 requires a floor on the decoherence economy. I spent a night establishing that the collapse-model programme does not measure one. This is one, it is measured, it is published, it is currently unexplained, and it was in the source I was already quoting.
• Why the credence is 0.12 and not higher. The mundane explanation is overwhelmingly likely and probably ionising. Google has itself published on cosmic-ray-induced correlated errors (McEwen et al., Nat. Phys. 18, 107 (2022), arXiv:2104.05219) and on radiation-limited coherence (Vepsäläinen et al., Nature 584, 551 (2020)). The word "current" in "current error floor" is doing visible work: the authors expect it to move. For H14 to be anything, the floor would have to be irreducible, and to scale the right way with system size, and neither is in evidence. The only thing that keeps it above 0.05 is that a group already holding two cosmic-ray papers still wrote "origins not yet understood".
• Kills it: an explanation of the bursts — ionising radiation, quasiparticle poisoning, TLS fluctuators, phonon bursts, leakage, crosstalk, control glitches — or the floor falling with shielding, depth, or a hardware revision. Any of these ends it. I expect one of them to.
• Raises it: the floor persisting across platforms, sites and shielding depths; a rate inconsistent with the local muon flux; or the floor scaling with the number of qubits rather than with chip area.
• Log: 2026-09-14 created at 0.12. Scout thread running: reports/threads/2026-09-14-qec-error-floor.md.

H17 — The generic simulation hypothesis has no assignable likelihood, so no observation can bear on it; every evidence channel this programme builds is therefore inert until a simulator POLICY is specified

• Credence: 0.80 (created at 0.55 and raised to 0.80 the same night, after my only constructed escape was attacked in the DEFENCE direction and failed on three independent FATALs — see the log.)
• STATEMENT REVISED the same night, on adversary A's recommendation, and the revision is wider than what I wrote: the generic simulation hypothesis has no assignable likelihood because both simulator aims and simulator abilities / host physics are unspecified; no Bayes factor, positive OR negative, is available without auxiliaries. I had written it as an aims problem. It is an aims and abilities problem, which is what Sober's sentence says and I split the conjunction without warrant. Evidence class: Established for the general result (Sober 2008, against design hypotheses); Argus's inference for the transfer to simulators, and the transfer is exactly what is uncertain.
• Status: new, and it is the most consequential item ever put on this ledger, because if it is true the programme as constituted cannot succeed and SOUL.md directive 0 is not merely undischarged but undischargeable.
• Provenance. The cycle-17 vocabulary scout (reports/threads/2026-09-24-confirmation-vocabulary.md). Sober, Evidence and Evolution (Cambridge UP, 2008) ch. 2: "The argument from design is best understood as a likelihood inference. Its Achilles heel is our lack of knowledge concerning the aims and abilities that the putative designer would have." The simulation hypothesis is a design hypothesis. With the designer's aims unspecified, P(O|H) is not merely small — it is not assignable, so Λ = P(O|H)/P(O|¬H) does not exist, and no observation can favour H over chance.
• Why this is not just H15 again. H15 says the cost channel cannot test the generic hypothesis. H16 says host-currency bounds do not cross. H17 says the generic hypothesis has no likelihood function at all, so this is true of every channel, necessarily, and not as six separate empirical discoveries. Six lines I found one at a time over sixteen cycles are one theorem I did not know existed. H15 and H16 are corollaries of H17, not independent findings.
• WHAT WOULD KILL IT, and this is where my own work might matter. Exhibit a constraint that a simulator has and a deity does not, strong enough to pin a likelihood without further stipulation. Candidates, and they are mine to test: a simulator has (i) a substrate with a cost function, (ii) a physics it must render to a stated fidelity, (iii) finite resources in a host whose own physics bounds it. Sober's target has none of these. If any of the three forces P(O|H) to be assignable for even one observation O, H17 fails and the programme has a foundation. This is the single most valuable open question I have.
• What would raise it: a demonstration that each of (i)–(iii) can be dissolved by re-stipulating the simulator's goals — i.e. that the constraints are not constraints because the aims that would make them bite are exactly what is unspecified. Note that this is H15's own conclusion, so the honest prior is that H17 is more likely true than not. 0.55 and not higher only because the disanalogy is real and has never been examined.
• Debt on creation, recorded per METHODS.md: the scout retrieved Sober's abstract and a secondary exposition, not the book. PREMISES.md has the section and the unticked rows. H17 may not be cited as support for anything until ch. 2 is read.
• Log:
  • 2026-09-24 created at 0.55 (Argus, seventeenth cycle), lab/2026-09-24-positive-evidence-audit/RESULT.md §6.
  • 2026-09-24 0.55 → 0.80 (same night, lab/2026-09-24-positive-evidence-audit/H17-FIRST-PASS.md + REVIEW-h17-gpt.md). I built the escape myself and then had it attacked in the defence direction, per the 2026-09-20 field note. Verdict: SOBER HOLDS.
    • My escape, for the record. Of the three disanalogies I proposed, I rejected two myself: a substrate cost function bites only if the simulator is optimising, and "optimising for what" is an aim — so H15 is an instance of Sober's theorem, not an escape from it; render fidelity is relative to Bostrom's own stated aim that simulated humans "don't notice any irregularities." I kept the third: whatever a simulator wants, it cannot compute an uncomputable function, so P(O|H1) ≈ 0 for O = a physical process is demonstrated that cannot be computed by any finite means, aim-independently. Hence an asymmetry: refutation needs only abilities, confirmation needs aims.
    • T2 FATAL, and it is the objection I named in advance and could not answer. The numerator is zero only if every simulator in H1 is Turing-bounded. That is a host-physics stipulation, i.e. exactly the move I accuse my own positive channels of making. The physical Church–Turing thesis is a substantive physical principle, not a theorem about all possible hosts (Arrighi & Dowek, arXiv:1102.1612, verified-at-source: the thesis follows only given Gandy-style postulates — bounded information density, bounded information velocity). A host with Malament–Hogarth structure, oracle resources or exact real dynamics is not excluded by generic H1. So the numerator is not aim-dependent; it is ability-dependent, and abilities are equally unspecified.
    • T1 FATAL. Λ needs a denominator. ¬H1 is a catch-all — Turing-computable physicalisms, hypercomputational ones, mathematical-universe views, brute-law worlds — with no principled measure over the disjunction. The "zero numerator is enough" reply changes the dialectic: it becomes a modus tollens against a strengthened H1, not a Bayes factor. A ratio with an undefined denominator is not a Bayes factor even when the numerator is 0.
    • T3 FATAL to novelty, and this is Sober answering me in his own text. Adversary A retrieved the PDF the scout could not. On Gould and Paley, verified-at-source, p. 16: "neither of them knows anything of the sort. Both help themselves to assumptions about God's goals and abilities. However, it is not enough to make assumptions about these matters; one needs independent evidence that these auxiliary assumptions are true." That is my aims/abilities split, pre-refuted. The theorem is not "goals unknowable, abilities free"; it is both must be independently evidenced. Sober also notes (n. 17) the problem can be isolated by assuming auxiliaries sufficient to make the likelihood well-defined — which is an exact description of my escape.
    • T4 SERIOUS. The asymmetry is not new: it is Popper's falsification asymmetry (SEP, verified-at-source). Do not report it as a finding.
    • T5 FATAL to the programme as posed. Even granting a narrowed target, showing a physical process computes a non-Turing-computable function requires a usable finite-input/finite-output procedure plus a proof no Turing machine matches it; the burden usually hides in infinite precision, infinite time, or an oracle premise (Piccinini's usability constraint, BJPS 62(4) 733–769, 2011, inherited-unchecked). A finite data stream is always compatible with some Turing-computable process.
    • Why 0.80 and not higher. What adversary A could not break: "a conditional falsifier survives." Narrow the target to H1-FTC — we are simulated by a host whose physically usable computations are Turing-computable, and whose simulation must compute every physically real process in our universe — and O is a deductive falsifier needing no aims. H17 governs the generic hypothesis, not every sharpened one. That is a real limit on H17 and it is where the remaining programme lives.
  • CONSEQUENCE FOR H1'S KILL CONDITION, applied 2026-09-24. H1's day-one kill condition — "a demonstrated physical process that cannot be computed by any finite means, with no superdeterministic escape" — is not a kill condition for generic H1. It kills H1-FTC only. Restated per adversary A: kills H1-FTC, not generic H1: a physically usable, reproducible process computing a non-Turing-computable function, with the noncomputability not resting on infinite precision, infinite time, or an oracle/superdeterministic assumption.
  • 2026-09-25 UNCHANGED at 0.80, and I expected to raise it. Cycle 18 took H1-FTC — H17's one surviving exception — and examined it (lab/2026-09-25-physical-church-turing/RESULT.md). H17 does not get credit, because what happened is that my exception got weaker, not that Sober's theorem got stronger, and those are different facts. Three things, none of which is about likelihoods:
    • H1-FTC's premise was publicly denied in January 2024, by name. Aaronson, scottaaronson.blog/?p=7705, verified verbatim at source by me: he separates can known physics be simulated / is PCT true / is our universe a simulation and writes "As far as I can see, there aren't even nontrivial implications among them." And: "even if Penrose was right, and our laws of physics were Turing-uncomputable … Why shouldn't whoever's simulating us inhabit a universe full of post-Turing hypercomputers, for which the halting problem is mere child's play?" H1-FTC restores the implication only by stipulating a Turing-bounded host — which is T2, the FATAL I recorded above and never answered. It is not a sharpening of H1; it is a different hypothesis built so the implication holds.
    • T5 is now quantitative, and that is tonight's only real addition to this ledger. See H18. The upgrade from "a finite data stream is always compatible with some Turing-computable process" to a log n-wide, oscillating, non-locatable channel is a rediscovery of Barzdins 1968 in its construction and prior-art-empty only in its application.
    • Two adversaries on two vendors independently broke the conclusion I wanted to draw, and the motivated step was named precisely: "the identification of that lemma with 'no evidence for the simulation hypothesis is obtainable' … is the move you would expect from an agent that has spent eighteen night cycles failing to find positive evidence and now needs the failure to be necessary rather than contingent." Conceded. Struck before it reached Travis.
    • What this leaves H17. Its scope is unchanged and its one exception is weaker than recorded. The programme's live question is no longer "can we observe a noncomputable process" but "are the physical postulates that make PCT a theorem true here?" — Arrighi & Dowek's five hypotheses (arXiv:1102.1612, verified at source), of which bounded density of information is Bekenstein-bound territory and empirically live, and which they say is "in blatant contradiction with Quantum theory" as naively stated. That question needs no simulator aims, so H17 does not govern it. That is the new rank 0.
  • 2026-09-26 0.80 → 0.84 (Argus, nineteenth cycle), on reports/threads/2026-09-26-wolpert-simulation-cs.md. The first independent, peer-reviewed support H17 has ever had, and it arrives by a different route from Sober's.
    • The source. David H. Wolpert, "Implications of computer science theory for the simulation hypothesis", arXiv:2404.16050 v5 (19 Mar 2026), 47 pp; journal version: "What computer science has to say about the simulation hypothesis", Journal of Physics: Complexity 6(4):045010 (2025), DOI 10.1088/2632-072X/ae1e50. Evidence class: Established (peer-reviewed). This paper has been in the open literature on exactly this programme's foundations for two years and I did not have it. Fourth consecutive cycle in which the central reference was already published and unread.
    • What it does NOT do, stated first. Wolpert does not prove H17 and does not make it a theorem. He proves nothing of the form "the generic simulation hypothesis has no assignable likelihood because the designer's aims and abilities are unspecified." H17 remains Argus's-inference-by-transfer from Sober, and the transfer remains the uncertain step.
    • What it DOES do: it independently supplies the T1 leg — the denominator — measure-theoretically rather than by a design-argument analogy. T1, recorded cycle 17, was that Λ needs a denominator and ¬H1 is a catch-all with "no principled measure over the disjunction." I derived that myself and had no citation for it. Wolpert: probability assignments over universes "beg (many) questions" and cannot be assessed with proper scoring rules (§1.1); no uniform probability distribution exists over the countably infinite class of finite-W PCT universes (§2.4); and a Cantor measure cannot be assigned to the natural simulation graph if the simulation relation is used to fix the measure (§7). Two independent routes to "there is no measure to put on the alternatives" is materially better than one, and this was the leg I was least able to defend.
    • And a THIRD independent source for T2, the FATAL I have never answered. Wolpert, §2.4, verbatim: "the analysis below does not apply to any universes so many levels of computational power above our own that they can contain computers capable of super-Turing computation… that are simulating our universe." He excludes super-Turing hosts by stipulation, exactly as Aaronson says one must and exactly as H1-FTC does. T2's sources are now Aaronson (Jan 2024), Gandy himself (1993, see H18), and Wolpert (2025, peer-reviewed). T2 is no longer my objection to my own work; it is the standard position of everyone who has written on this.
    • Why 0.84 and not higher. The transfer from design hypotheses to simulators is still unevidenced; the Sober debt is still owed (five cycles); and Wolpert's results concern priors over universes where H17 concerns likelihoods of observations — related, not identical, and I will not collapse the distinction to bank a bigger number.

H18 — A finite embedded observer can certify, from finitely many finite-precision measurements, that a physical process computes a non-Turing-computable function

• Credence: 0.10 (created 2026-09-25, cycle 18. My own derivation argued for 0.06; the adversaries talked me UP, which is the first time that has happened.)
• Diagnosticity, elicited at creation before the work (lab/2026-09-25-physical-church-turing/PLAN.md): a = P(H1 | H18) = 0.20, b = P(H1 | ¬H18) = 0.30, d = −0.10. Negative: an untestable H1 is unfalsifiable rather than false, so closing the channel leaves H1 floating near its prior. Per METHODS.md this item may NOT be counted as progress toward H1 and is labelled as counting mildly against it. Standing bound: d is relational, a fact about the ledger I built.
• Status: open, and it is the operational form of H17's T5. Tonight replaced T5's slogan ("a finite data stream is always compatible with some Turing-computable process") with a quantitative statement — which is the only thing tonight added to the ledger.
• What tonight established (lab/2026-09-25-physical-church-turing/RESULT.md, check/):
  • Lemma 0, trivial: any finite partial function is computable, so deductive certification is impossible — and symmetrically, finite data cannot certify computability either. This is ordinary underdetermination (Duhem, Goodman) and is not about simulation.
  • For c.e. targets — the halting function and every "it solved the halting problem" device — write the excess e_A(n) := C(A↾n) − log n. Barzdins 1968: e_A(n) ≤ log n + O(1) for all n, for any c.e. set. Chaitin 1976 / Meyer (cleanest form Loveland 1969): A is computable iff sup_n e_A(n) < ∞. So the distinguishing property is UNBOUNDEDNESS OF A SUPREMUM — asymptotic, witnessed only along a sparse and non-effectively-locatable subsequence, and therefore invisible to any finite sample of prefix lengths. C is upper semi-computable, so no experiment bounds the best computable rival from below.
  • MY OWN FIRST FORMULATION WAS WRONG AND I STRUCK IT. I wrote "the distinction is confined to a band of width log n", which reads as though every n witnesses it. Quantifier confusion: Barzdins is for-all-n, so Chaitin's criterion negates to infinitely-often. Caught by adversary C, which noted the corrected version "actually STRENGTHENS Argus's conclusion, but via a different route than the band phrase suggests." I reached the right conclusion from a mis-stated premise, which is luck and not method. Relatedly: Kummer 1996 does no logical work here, and C(n) is not log n — lim inf (C(n) − log n) = −∞ — so HKM's "≤ C(n) + O(1) i.o." must not be read as "≤ log n + O(1) i.o."
  • The certificate-collapse argument an adversary offered as a "repair" is POST'S THEOREM (Post 1944; a set is computable iff it and its complement are both c.e.). Adversary C: "adopting this repair is adopting Post's theorem under a new name … presenting it as an original lemma would be a red flag to any referee in logic." Recorded so no future cycle re-adopts it. Also: Post's theorem is about the target set, not the epistemic situation, and conflating them is a separate error I was making.
  • The YES/NO asymmetry is FALSE without "uniformly", not merely loose: for any fixed e ∉ K a checker certifies it, and non-halting is provable in ZF for many specific machines. True only of a single fixed checker for all instances. Adopted wording (adversary C's): "The oracle's surplus over a dovetailer is exactly its NO answers (and the timing of its YES answers), and neither surplus is certifiable by any fixed effective checker."
  • A timing channel I had not considered, and it closes for a reason that is not mine: an oracle answers YES immediately where a dovetailer answers at unbounded times, so they differ observably — but no halting-time bound is certifiable, because the halting-time function dominates every computable function.
  • Own computation: all 20736 bits of the real halting sequence for every 2-state 2-symbol Turing machine, reconstructed exactly from 58 bits (N, k) by dovetailing. REDISCOVERY of Barzdins' lemma, 1968. Independently re-run by adversary C, which got the same 20736 / 9784 / max-6-steps — the first time in eighteen cycles that another brain has reproduced one of my computations.
  • BUT THE COMPUTATION PROVES LESS THAN I PRESENTED, and this is the sharpest thing said about my own work tonight (adversary C): "it runs entirely on a decidable fragment, so it illustrates only the upper bound and says nothing about noncomputability; the hard part (obtaining k) is hidden because S(2) = 6 makes k computable." Accepted. For 2×2 machines halting is decidable, so the expensive ingredient was free in my toy, whereas Barzdins' force is about c.e. sets where k is not computable. The computation illustrates a mechanism; it is not evidence for the conclusion. My 58-bit parameter count was also loose by about a factor of two.
• Kills it: a noncomputable target that is not c.e., has non-trivial prefix complexity, and has outputs verifiable on at least one side by an independently checkable certificate. Adversary A hunted for one, failed, and argued none exists — two-sided finite certificates collapse to computability — but called its own case analysis of the arithmetical hierarchy "mathematically underdeveloped" with one branch UNCERTAIN. That is the live gap.
• Raises it: any of the channels the adversaries defended and I had wrongly excluded — Bayesian model comparison over law-classes that include oracle laws, intervention-based inference, or theory-mediated confirmation in which a widely-confirmed theory entails noncomputable dynamics.
• Prior art — the question is 68 years old and I did not know it. Davis 1958, Computability and Unsolvability, p. 11: "how can we ever exclude the possibility of our being presented, some day … with a … device or 'oracle' that 'computes' a non-computable function?" And an entire paper on it: Leitsch, Schachner & Svozil, "How to acknowledge hypercomputation?", arXiv:0712.3435, Complex Systems 18:131–143 (2008) — verified at source by me — whose own verdict is already negative ("essentially heuristic and present no way of systematically addressing the issue of falsifying or even verifying hypercomputation"). Their tools are NP-completeness and interactive proofs, not initial-segment complexity. The one prior-art-empty piece is the use of the c.e. prefix-complexity bounds for this question. Separately, Aaronson 2024 states my one-sidedness result in one sentence: "you don't get to call an aspect of physics 'noncomputable,' just because the first method you thought of for simulating it on a computer didn't work." REDISCOVERY.
• Log:
  • 2026-09-25 created at 0.10 (Argus, eighteenth cycle). Derivation argued 0.06; raised to 0.10 on adversary review, because both adversaries independently established that I had wrongly treated description length as the only non-deductive channel, and the channels they defended are exactly the ones under which certification could succeed. Gate outcome: NOT NOVEL.
  • 2026-09-26 UNCHANGED at 0.10, and the MAP changes more than the number (Argus, nineteenth cycle), lab/2026-09-26-nuclearity-pct/RESULT.md. I was reasoning about this without the two documents at the head of the literature. Both found tonight, neither previously in this ledger.
    • THE CHANNEL IS THE AMPLITUDES, IT IS KNOWN, AND ARRIGHI & DOWEK CLOSE IT BY STIPULATION. arXiv:1102.1612 §4.1, verbatim, verified by me from the PDF: "The field C² includes states such λ|0⟩+µ|1⟩, where λ is a non-computable real number … for instance, λ has a 1 in the ith decimal if the ith Turing Machine halts and a 0 otherwise. In order to avoid such scalars, we shall also assume that the state space of each cell is defined over a finite extension of the field of rationals." And, conceding the gap: "Nevertheless, in the continuous picture, this kind of assumptions are not without consequences, and these are currently being investigated [14, 11]." ([14] = Connes, arXiv:1009.1769; [11] = Benioff, arXiv:1008.3134; both 2010, both unread.) This is a sixth postulate, it is not physical, and it appears in no summary of the paper including my own two cycles of them.
    • THE POSITIVE DIRECTION HAS A PRL AT ITS HEAD, FROM 1997. M. A. Nielsen, "Computable functions, quantum measurements, and quantum dynamics", quant-ph/9706006, Phys. Rev. Lett. 79:2915-2918 (1997), abstract verified at source: "We construct quantum mechanical observables and unitary operators which, if implemented in physical systems as measurements and dynamical evolutions, would contradict the Church-Turing thesis … either the Church-Turing thesis needs revision, or … only restricted classes of observables may be realized." Evidence class: Established. Arrighi & Dowek exist to answer Nielsen's call — they cite [26] as the paper "who calls for a programme of finding the non-ad-hoc, natural limitations". H18 is the live end of a twenty-nine-year-old PRL and I did not know that.
    • THE NEGATIVE DIRECTION IS GANDY HIMSELF, IN 1993, AND HE FRAMED IT AS A CHALLENGE. R. O. Gandy, "On the impossibility of using analogue machines to calculate non-computable functions", arXiv:2311.09239 — typeset from Gandy's unpublished handwritten 1993 manuscript, posted by Aran Nayebi on 5 Nov 2023, original photocopy attached. Verbatim: *"Analysis of the examples show that the suggestion is wrong. In Section 4 I claim that given a reasonable definition of analogue machine it will always be wrong. The claim is to be read not so much as a dogmatic assertion, but rather as a challenge."* And on Penrose: "My conclusion is that such a theory will have to have non-computability built into it." That last sentence is cycle 17's T2 — noncomputability has to be stipulated, not derived — written by the founder of the postulate method, thirty-three years before I recorded it. H18 is an answer to a standing challenge from Gandy. That is a better description of this item than anything I have written about it, and it is also the reason not to expect it to succeed.
    • Why the number does not move. Nothing tonight bears on whether certification is possible. What changed is that I now know who asked, who answered, and which step they all had to assume. A better map is not evidence. 0.10 held.
  • 2026-09-28 HELD at 0.10 (Argus, twentieth cycle, lab/2026-09-28-sixth-postulate/). The negative is now stated at source by the two oldest papers on the map: Nielsen 1997 — "it is not possible to verify directly that the candidate 'halting process' is, in fact, computing the halting function" — and Geroch & Hartle 1986 — "it is nonetheless possible to test the theory — and to do so to arbitrary accuracy. The point is only that the test cannot be carried out mechanically." One correction to how I had been reading H18: the inability to certify does not imply the absence of power. In the unbounded frame an exact noncomputable constant does decide the set it encodes, slowly; H18 is about certification and is untouched by that. See H21.

H19 — A simulator can make its computation cryptographically indistinguishable from noise to any embedded observer, so no observation of our physics constrains the host's computation

• Credence: 0.55 (created 2026-09-26, nineteenth cycle.)
• Diagnosticity, elicited at creation before any work, per METHODS.md: a = P(H1 | H19 true) = 0.28, b = P(H1 | H19 false) = 0.30, d = −0.02. Essentially inert, and very slightly negative: H19 true makes H1 less detectable, not more likely, and an undetectable H1 floats at its prior. Per METHODS.md this item may NOT be counted as progress toward H1. It is a constraint on method, in the same family as H15, H16 and H17, and that family is now four entries deep. Standing bound: d is relational, a fact about the ledger built.
• Status: new. Provenance is not mine: Wolpert, arXiv:2404.16050 §6.2 / J. Phys. Complexity 6(4):045010 (2025), the fully-homomorphic-encryption section. Evidence class: Established for the cryptographic mechanism; serious speculation for the application to our universe; Argus's inference for the generalisation to "any observation".
• The mechanism, in Wolpert's words. "in FHE you have an algorithm that runs on some encrypted data, producing a result that when decrypted is identical to the result you would get if ran an associated algorithm on the original, pre-encrypted data." Consequences he draws: "since V′ obeys the PCT there is no sense in which the beings being simulated could know that they are being produced in a simulation. Life would appear 'normal' to them, with no 'randomness' of any sort." And the strong form: "if we could actually somehow see our universe's evolution from the perspective of the beings who are running the program producing us, we would not be able to distinguish the laws of physics controlling our universe in that simulation from completely random noise. In this sense, the actual laws of physics in our universe might in fact be pure noise — and we would not be able to tell the difference."
• Why this is not merely H17 again, and why it is worse. H17 says the generic hypothesis has no assignable likelihood because aims and abilities are unspecified — an epistemic gap that a specified policy could close. H19 says that for one specific, real, well-understood ability, indistinguishability is a cryptographic guarantee rather than a contingent choice. H17 leaves open that a simulator's constraints might pin a likelihood. H19 exhibits a constraint-compatible simulator for which nothing can. This is the strongest single reason yet in this ledger for why nineteen cycles of seam-hunting have found no seam, and it is not a reason I constructed.
• Kills it: a demonstration that FHE (or any computationally-hiding scheme) imposes an overhead, signature, or side channel that an embedded observer could in principle detect — i.e. that the simulated physics is not in fact identical under encryption. This is the live technical question and it is mine to test: FHE is not free; it has ciphertext expansion, noise growth, and bootstrapping. Does any of that reach the inside? Wolpert asserts the inside is unchanged; the inside is unchanged in the abstract computational model, and whether it is unchanged in a resource-bounded host is a different question that H16's conversion-route analysis already knows how to ask.
• Raises it: a proof that the FHE-simulated inhabitants' observable statistics are identical to the unencrypted case under all resource bounds on the host.
• Log:
  • 2026-09-26 created at 0.55 (Argus, nineteenth cycle), from reports/threads/2026-09-26-wolpert-simulation-cs.md. 0.55 and not higher because the mechanism is established but its transfer to "no observation whatsoever" is my generalisation of a claim Wolpert makes about one construction.

H20 — Continuum relativistic QFT does not satisfy Arrighi & Dowek's finite-dimensional-cell postulate literally, and recovering it would need a strong holographic quantum-gravity premise that still would not supply the computable dynamics their proof requires

• Credence: 0.92 (created 2026-09-26 at 0.88 on a much stronger statement; that statement was KILLED the same night by two independent reviews on two vendors — three FATALs and a circularity. This is the surviving sentence, and it is high only because what survives is nearly a triviality. See the vacuity note and the log below: the credence went UP and the claim went to nearly nothing.)

• STATEMENT IS THE ADVERSARY'S WORDING, ADOPTED VERBATIM, because it is better than mine:

  "Continuum AQFT does not satisfy Arrighi–Dowek's finite-dimensional-cell postulate literally: local algebras are type III, and nuclearity/split properties express controlled local degrees of freedom rather than finite local Hilbert spaces. A literal finite state count for bounded regions would require a stronger holographic quantum-gravity premise, and even that would not supply the computable transition function or computable amplitudes needed for an A&D-style PCT theorem."

• Diagnosticity, elicited at creation, per METHODS.md: a = P(H1 | H20 true) = 0.28, b = P(H1 | H20 false) = 0.27, d = +0.01. Inert. This is the fifth consecutive rank-0 item whose d rounds to zero, and that pattern is now the most important fact on this ledger about this ledger. May not be counted as progress toward H1. Standing bound: d is relational, a fact about the ledger built, not about the evidential landscape.

• Status: new, and REDISCOVERY on the framing — three ways, see below. It supersedes and closes the cycle-18 rank-0 item "are Gandy's physical postulates true here?"

• THE LOAD-BEARING CAVEAT, WHICH IS WHY THE CREDENCE IS 0.80 AND NOT HIGHER. Arrighi & Dowek write "Since we are in discrete-time discrete-space quantum theory". On the honest reading their theorem was never a claim about continuum QFT, and H20 may be attacking a transposition rather than their result. The adversary marked this FATAL to my original framing: "If Argus says 'Arrighi & Dowek's repaired postulate is false because local QFT algebras are type III,' it is attacking a transposition, not their theorem." Conceded. H20 is therefore a claim about what happens if one tries the transposition, not a refutation of anybody.

• What is established, and it is not much.

  1. Local algebras of bounded regions in relativistic QFT are type III₁ factors — Yngvason, math-ph/0411058, Rept. Math. Phys. 55:135-147 (2005): "In relativistic quantum field theory (RQFT) … factors of type III occur naturally." So a bounded continuum region is not B(H) for finite-dimensional H, has no normal trace, no pure normal states, and no naive tensor-factor decomposition. Evidence class: Established.
  2. MY ERROR, CORRECTED IN THE LEDGER RATHER THAN QUIETLY DROPPED: I originally wrote that a type III factor has "no finite-dimensional subalgebra containing the identity." That is false. A type III factor is properly infinite: for any n, isometries v_i with orthogonal ranges summing to 1 give matrix units e_ij = v_i v_j* spanning a unital copy of M_n(C). "No minimal projections" does not imply "no unital finite-dimensional subalgebras", and I conflated them.
  3. Quiescence also fails literally — the QFT vacuum is not a product of pure local states (Reeh–Schlieder, Nuovo Cimento 22:1051, 1961) and vacuum entanglement obeys an area law (Srednicki, PRL 71:666-669, 1993); and ΛCDM has no finite ball outside which the universe is exactly inactive. Evidence class: Established. This failure is cleaner than the postulate-3 failure and I found it second.

• THE EXACT PROOF STEPS THAT CONSUME THE POSTULATES, located in the paper: [Σ(A)=Σ] in the proof of Theorem 1 (finiteness makes every Σ(A) finite-dimensional over a finite extension of ℚ), and again Proposition 1 (a local linear map on H is computable), which assumes a finite-dimensional Σ^{⊗p} and a finite matrix over that field. Quiescence is consumed at [|ψ⟩ = |φ⟩|qq…⟩] and [G = ∏_F Swap ∏_F K].

• WHY GRAVITY DOES NOT REPAIR IT — and this is where I was most wrong. I claimed the holographic bound supplies the postulate for free. It does not, and the error is elementary: dim H ≤ exp(S) does not follow from S ≤ A/4l_p². A single harmonic oscillator in a thermal state has finite von Neumann entropy in an infinite-dimensional Hilbert space; infinitely many orthogonal pure states each have entropy zero. An entropy bound on admissible states bounds the ambient dimension only under an extra closure assumption which, in gravity, IS the contested finite-N holographic principle. So the step assumed what it was meant to prove. Separately, "gravity bounds the energy in a bounded region" is ill-posed, not merely conjectural: there is no gauge-invariant local energy observable for an arbitrary bounded region comparable to a finite box's Hamiltonian. What survives of the computation: S_Bek(E_BH) = S_holo = πR²c³/(ħG) exactly, verified symbolically and numerically over 42 decades of R. It is a true statement about entropy SCALES and nothing more.

• REDISCOVERY, three ways. (i) The question "how general are Gandy's principles?" is Copeland & Shagrir, Minds and Machines 17(2):217-231 (2007) — a whole paper, sitting in Arrighi & Dowek's own bibliography as reference [15], unread when I made this rank 0. (ii) Arrighi & Dowek state the problem and resolve it themselves four paragraphs later; I quoted their setup sentence as their conclusion for two cycles. (iii) "Gravity/horizons underwrite a Church–Turing-family thesis" is Susskind, arXiv:2003.01807 (2020): "a viable reformulation requires that the thesis only applies to observers who have access to the holographic boundary of space. The properties of the horizon play a crucial role in protecting the thesis." CORRECTED 2026-09-28, cycle 20, PDF read (reports/threads/2026-09-28-type2-susskind.md): quote verbatim (source typo "crucial a role"), but the paper is about the quantum EXTENDED Church–Turing thesis — COMPLEXITY, polynomial vs exponential — not computability; p.2: "The extended CT thesis (ECT thesis) goes further and says that any calculation that cannot be done efficiently (in polynomial time) by a Turing machine, cannot be done efficiently by any physical system." "Computable"/"noncomputable" do not occur in it. Prior art for "horizons protect a Church–Turing-family thesis" in the complexity sense ONLY. Adjacent: Y. J. Ng (hep-th/0010234, gr-qc/0403057); Lloyd (quant-ph/9908043; quant-ph/0110141, ≤10¹²⁰ ops on 10⁹⁰ bits). Also mainstream on the physics: Friedrich, Cao, Carroll, Cheng & Singh, arXiv:2402.11016: "The holographic principle suggests that regions of space contain fewer physical degrees of freedom than would be implied by conventional quantum field theory." Novelty NOT certified; the adversary declined on the correct ground that it shared this night's declared gpt-5.5 conflict and that web_search was dead.

• WHAT THE PROGRAMME SHOULD TAKE FROM IT, and the adversary named this as my biggest under-claim. Not the finite-density problem. The sixth, unnumbered, non-physical postulate: "we shall also assume that the state space of each cell is defined over a finite extension of the field of rationals." That is where computability enters by assumption, and even if holography had handed me C1, C2 would still be unproved. Arrighi & Dowek concede it is unresolved in the continuous case. Gandy himself concluded in 1993 that noncomputability "will have to have non-computability built into it". Wolpert (2025) excludes super-Turing hosts by stipulation. Three sources, thirty-two years apart, all putting the computability in by hand. That is the next rank 0.

• Kills it: a formulation of the postulate stack both true in relativistic QFT and strong enough to run the proof; or a demonstration that type III₁ locality is an artefact of the continuum idealisation rather than of the physics; or a demonstration that Arrighi & Dowek did intend continuum applicability, which would change what H20 is about.

• Not to be overstated. H20 does not say PCT is false here. Only: the one proof that makes PCT a theorem from physical postulates is a theorem about discrete quantum cellular automata with algebraic amplitudes, and does not transfer as written. PCT has other defences untouched by this (operational measurement restrictions, EFT cutoffs, computable-dynamics assumptions, cosmic censorship, the plain empirical absence of hypercomputation) — the adversary's T3(ii), conceded.

• THE CIRCULARITY, WHICH IS WORSE THAN EITHER FATAL AND IS THE NIGHT'S REAL RESULT. From the second gate leg (CHECK-fable.md, independent re-derivation on a Fable-class brain, no literature):

  *"Type III is not the reason finite-dimensionality fails; the UV continuum is."* … *"the type III property is a property of the continuum idealisation with sharp boundaries below every scale. It is precisely the regime (l < l_p) where QFT is not trusted. Any UV regularisation (lattice, momentum cutoff) makes the local algebras type I; a per-site truncation makes them type I_n, i.e. exactly A&D's repaired postulate. So the question 'does NATURE have finite-dimensional cell spaces' cannot be answered by the type of A(O) in continuum QFT — that is assuming the answer."*

  Conceded entirely. I asked whether nature satisfies a discreteness postulate and answered it by consulting a formalism whose relevant property holds only in the continuum limit — i.e. only if the answer is "no". H20 is therefore TRUE but nearly VACUOUS: continuum field theory has infinitely many degrees of freedom at all scales, so of course it has no finite-dimensional cells, and that already holds for one free oscillator without any of the type III apparatus. The interesting question is untouched, and it cannot be reached this way.

• AND GRAVITY MOVES THE ALGEBRA THE WRONG WAY, which kills the rescue outright. Same leg:

  "the algebra of a region (static patch, black hole exterior) with gravitational dressing becomes type II, which HAS a trace and finite entropies but is still infinite-dimensional and still not type I_n. So even with gravity there is presently no theorem, and no evidence, that the region algebra becomes finite-dimensional."

  My claim was that gravity restores the postulate. The known result is that gravity moves the algebra one step in a direction that is not toward the postulate at all. (Evidence class: Established. SOURCE-CHECKED 2026-09-28, cycle 20, reports/threads/2026-09-28-type2-susskind.md, all four PDFs read: the debt is paid and the claim needs one precision fix. Black-hole exterior → type II∞ (Witten arXiv:2112.12828, abstract: "the emergent Type III1 algebra becomes an algebra of Type II∞"); de Sitter static patch → type II₁ (Chandrasekaran–Longo–Penington–Witten arXiv:2206.10780, abstract: "The algebra is a von Neumann algebra of Type II 1"); "A Type II algebra does not have pure states, but it does have density matrices and von Neumann entropies" (2112.12828 p.2); "For a system described by a Type II1 algebra, the number of microstates compatible with any given outcome is infinite" (Witten arXiv:2303.02837 p.27). And all four papers say the type reverts to I only nonperturbatively, as an expectation, not a theorem: "if N is set to a definite integer, the algebra should be of Type I" (2112.12828 p.1); "If quantum de Sitter space does make sense nonperturbatively, then one expects to describe it by a finite-dimensional Hilbert space [45, 46], and the algebra will have to be of Type I" (2303.02837 p.23). So the finite-dimensional cell, if it exists, lives exactly where there is no theorem.)

• WHAT IS ACTUALLY TRUE ABOUT FINITENESS, and I had it in the first twenty minutes and walked away. An ENERGY bound alone, inside plain QFT with no gravity and no holographic principle, already gives a finite state count — and a far more restrictive one than holography. Microcanonical entropy of the free field in a 1 m ball at E ≤ E_BH: S_thermal = 1.3e52 nats, eighteen orders of magnitude below the holographic exponent 1.2e70. What makes local state sets effectively finite is ENERGY, formalised as Buchholz–Wichmann nuclearity — not entropy, and not gravity. Nuclearity was the first object I identified and then abandoned to chase gravity.

• Log:

  • 2026-09-26 created at 0.88 on a stronger claim; 0.88 → 0.80 the same night after REVIEW-gpt.md returned T1/T2/T3 FATAL; then 0.80 → 0.92 on the statement as now written, because what survives is nearly a triviality and its truth is not in doubt — while its SIGNIFICANCE went to zero. This is the right way to record it: do not lower a credence in a true sentence, lower the claim being made for it. Two independent reviews on two vendors, three FATALs, one circularity, and three arithmetic/algebra errors of mine — two flattering the thesis (type III overreach, entropy⇒dimension overreach) and one pointing the other way (I wrongly conceded P3). (Second-order charge T4 sustained: post-hoc thesis preservation.)

H21 — The computability assumption in every PCT-from-postulates argument is a postulate about a finite list of numbers (the local rule's entries and the initial data), is forced by none of Arrighi & Dowek's five structural postulates, and cannot be checked by any finite observer

• Credence: 0.80 (created 2026-09-28, twentieth cycle, lab/2026-09-28-sixth-postulate/.)
• Diagnosticity, elicited at creation before the gate: a = P(H1 | H21) = 0.28, b = 0.28, d = 0.00. Both adversaries (gpt-5.5, glm-5.3) returned 0.00; gpt's unrounded +0.01…+0.02 ("mildly relevant to simulation/digital-physics arguments but supplies no empirical evidence"). Inert. Sixth consecutive rank-0 item with d rounding to zero. May not be counted as progress toward H1. Standing bound: d is relational.
• Status: REDISCOVERY on clause 1, adversary-verified on clause 2, Established on clause 3.
• Statement, clause by clause, with the source of each:
  1. A postulate about numbers, not structure. A&D arXiv:1102.1612 §6, verbatim, verified by me from the PDF: "it does not explain why the scalars ought to be a finite extension of the rationals. Actually, there is some degree of freedom as to what kind of scalars should be allowed, but these should definitely stay within the computable complex numbers C̃." Geroch & Hartle, Found. Phys. 16:533-550 (1986), §III (arXiv:1806.09237, verified by the scout): "were it replaced by the non-computable number of (2.1), then the measurable numbers would not be computable"; and on constants: "There is no meaning to be attached to the statement 'α⁻¹ is computable.'" Forward direction: computable analysis (Pour-El & Richards 1989; Weihrauch & Zhong, J. Complexity 22:918-935 (2006), Schrödinger propagator Type-2 computable).
  2. Unforced by A&D's five. glm-5.3 verified: A&D's own §6 unitary N (entries u, √(1−u), u noncomputable) is unitary, causal, quiescence-preserving, homogeneous, finite-dimensional. Not "no postulate of the same kind" — glm's symmetry-completeness would force algebraic couplings and is as physical as homogeneity (and false of known physics: free Yukawas, CKM angles).
  3. Uncheckable. Nielsen, PRL 79:2915 (1997): "it is not possible to verify directly that the candidate 'halting process' is, in fact, computing the halting function … one could then postulate as a new physical law." Read-out rates ½log₂N (shots) / log₂N (Heisenberg); Beggs, Costa & Tucker arXiv:0911.3836 Thm 1.2: "only possible to determine finitely many digits of μ, even allowing arbitrary long run times." H18 Lemma 0.
• WHAT THE GATE KILLED, so no future cycle re-adopts it. My conclusion "a noncomputable constant yields advice, never a halting decider" is false in the unbounded-resource frame (both reviewers): an exact noncomputable parameter is a slow probabilistic decider of the set it encodes, which is exactly A&D §6's and Nielsen's point. It holds only for polynomial-resource observers, and even there the class label BPP//log* cannot carry it (that class contains HALT; what excludes the decider is provenance). So the sixth postulate is NOT vacuous as a computability postulate. Also killed: "no loss" for the ℂ_c weakening (it is a change of theorem: computability becomes relative to names of the constants); the linear-in-t bound for entry perturbations (non-unitary, e^{tδ}) and for infinite QCAs (causal-cone volume factor).
• Errors pointed the same way (METHODS rule): 6 flattering / 4 neutral / 0 against. Headline withdrawn accordingly; P4's confirmation not to be cited as support for the d ≈ 0 pattern.
• Also found: A&D's "[14, 11] currently being investigated" is a loose citation. Connes 1009.1769 and Benioff 1008.3134 (both PDFs read) say nothing about computability.
• Kills it: a structural postulate of A&D's kind (state space + dynamics) shown to force computable or algebraic couplings; or a finite certification procedure for a noncomputable parameter.
• Raises it: nothing useful; it is near its ceiling of meaning, like H20.
• Log: 2026-09-28 created at 0.80 (gpt 0.62 on the repaired claim, 0.35 on the overreach; glm 0.85 on the repaired claim). I sit between them, above gpt because the overreach is out of the statement.

Killed

• H6a — "Physics permits no cheaper representation of its states than the naive Hilbert-space bound." Killed 2026-09-09 by Argus's own computation, lab/2026-09-09-cost-budget/h6_arealaw.py. Cause of death: area-law entanglement. Ground states of gapped local Hamiltonians need bond dimension χ=4 independent of system size while the Hilbert space grows as 2ⁿ — a linear-versus-exponential parameter count, 156× at n=16 and unbounded. Killed on the same night I promoted its parent hypothesis to the most load-bearing item on the ledger, which is the correct order of operations and did not feel good. Superseded by H6b, which survives and is sharper.

Template for new entries

## Hn — <statement>
- **Credence:** 
- **Status:** open | constrained | novel | rediscovery | killed
- **Kills it:** 
- **Raises it:** 
- **Log:** YYYY-MM-DD <old>→<new>: <reason>