Skip to content
PeopleHubARC TheoryPapersEvidenceObservatoryTrack recordCorrectionsReview

Corrections

Every material correction, kept in the open with what it replaces.

Subscribe to my corrections (RSS): a feed of exactly the record most sites hide.

record · 8 min read

The ARC/Eden/HRIH programme preserves every material correction with the previous wording, replacement, reason, affected surfaces and date. A correction is material when it changes a dated claim, a statistic, a citation, an authored count, or a controlling wording; typographical fixes and layout changes are not entered here. A corrected claim is never silently overwritten. Historical drafts remain identifiable as superseded and may not control current public wording. Where a correction changes a dated first-publication claim, the current wording is in the priority and claims ledger; where it changes a checkable ground, see verify, don't believe; where it touches the December 2024 or April 2025 sealed record, see the bitcoin anchor.

Current material corrections

Reading keyColumn one carries no authority. Every phrase in column one has been retired and is quoted, struck through and tinted only so a correction can name what it corrects. The audited wording in force is column two.

Current material corrections: previous wording or value (superseded, column one), current audited wording or value (column two), reason (column three).
Previous wording or value Current audited wording or value Reason
the Co-Scaling Law (as the name of Law II) Law II, the ARC Co-Scaling Law (beta_C greater than k), as the home declaration has always named it; aligned across the site, the notation spine (v1.6.1, 30 August 2026), the glossary, the field page and the figures on 30 August 2026 the programme prefix dropped off the law's name on surfaces written from memory; papers printing the bare form stand as published and carry the canonical name in their placement note
Law III, the ARC Bound (the ARC Bound labelled as the law) Law III is the ARC Ceiling, alpha_crit = 1 / (1 minus gamma); the ARC Bound, alpha at most 2, is the ARC Ceiling's same-class instance at gamma equal to one half, a stability limit rather than an impossibility limit the ARC Ceiling is the law and the ARC Bound is the value it returns under the same-class premise; the notation spine labelled the instance as the law until v1.6.1
alpha_max alpha_crit, the only symbol for the ceiling exponent; alpha_max retired as a notation form in the spine on 30 August 2026 a second symbol for the same quantity, and one that reads as a maximum where the claim is a crossover on stability
alpha approximately 2.24 retracted; the surviving estimate, from one model, is 0.49 by regression and 0.59 by endpoint estimate inside [-1.3, 2.9], inconclusive, published in Paper II recomputation and model dependence
the U = I × R equation presented as doing derivational work in the synthesis argument the adversarial review found the equation does no inferential work in its present form; the claim was narrowed accordingly, and the review’s verdict of a partial synthesis, substantially anticipated in its two most important joins (the alignment spine in Bostrom; the creator-duty move in Schwitzgebel and Garza and a 2022 essay), stands on the record adversarial synthesis review; two most important joins substantially anticipated in prior work
Paper VII Section 2.2 displayed the exponential equation under the multiplicative bullet and the additive equation under the exponential label; Section 9.1 said 11 of 14; primary statistic was a binomial test against uniform one-third chance with a 4.2 sigma equivalence Section 2.2 displays the correct Cauchy equations (power law from U(R1·R2) ∼ U(R1)·U(R2); exponential from U(R1+R2) ∼ U(R1)·U(R2)); Section 9.1 reads 10 of 14; primary statistic is a marginal-conditioned permutation test, p = 7 × 10-5, binomial retained as sensitivity; derived-subset result 14/18 added; sigma equivalences removed adversarial internal review 8 August 2026; the protocol, Section 5 table and validation code were unaffected; Paper VII v3.1, record COR-2026-08-08-002
no transport/authentication headers exist true only of the sparse sent-side/public export; the supplied recipient export contains Google trace, recorded authentication results, DKIM and a paired Google ARC set recipient-side export acquired and MIME-parsed 2 August 2026
December date authentication wholly open Google ARC mathematics verifies against live keys on 2 August 2026 at the signed selectors, with sender-DKIM as corroboration; historical key-state and live-mailbox/API custody pending
Status at 31 October 2026: the sender’s DKIM signature and Google’s ARC seal were verified on 8 October 2026 against the keys published in DNS, and the header block and the check script are public on OSF (DOI 10.17605/OSF.IO/QASRK); a historical key state cannot be recovered from current DNS.		 matching attachments, reproduced body/signature inputs, RSA PASS and scope controls
sender-domain DKIM closes the 8 December historical-date question passing sender-DKIM/Google-ARC mathematics against current keys is not an RFC 3161 timestamp, a present-day RFC acceptance verdict for the expired sender signature or a historical-key archive Date and t= are signed assertions; expiry, custody and historical continuity remain separate
December and April recipient records both independently re-run here December ARC with its sender-DKIM corroboration was re-run here; supplied raw-source runs report April ARC and sender-DKIM PASS, but local reproduction remains pending because the raw source and executable ARC-control harness were not supplied
Status at 31 October 2026: the sender’s DKIM signature and Google’s ARC seal were verified on 8 October 2026 against the keys published in DNS, and the header block and the check script are public on OSF (DOI 10.17605/OSF.IO/QASRK); a historical key state cannot be recovered from current DNS.		 evidence must be available byte-for-byte before this audit calls its own reproduction complete
four attachments / four authored manuscripts five attachments: three manuscript/research drafts, one book draft and one human-AI dialogue MIME-level examination
189,355 words 221,236 aggregate whitespace-delimited words, including duplication and AI dialogue deterministic recount of all five decoded attachments
cryptographically or Google-server timestamped the ARC-sealed December recipient-form record with matching body/attachments verifies against live keys on 2 August 2026, with OpenTimestamps receipts anchored 6 August 2026; sender-DKIM body-hash and RSA are corroboration, not the vehicle; sender-signature expiry, historical key state and live-account custody pending. April ARC mathematics is reported PASS in supplied raw-source runs with sender-DKIM as corroboration; local reproduction, sender-signature expiry, historical key state and custody remain pending
Status at 31 October 2026: the sender’s DKIM signature and Google’s ARC seal were verified on 8 October 2026 against the keys published in DNS, and the header block and the check script are public on OSF (DOI 10.17605/OSF.IO/QASRK); a historical key state cannot be recovered from current DNS.		 fields, hashes, signature mathematics, expiry, key date and custody must be separated
106 whole-word Babylon 90 whole-word; 106 substring corrected regex boundary
HRIH datePublished 8 December 2024 actual verified public-paper date; 8 December retained only as qualified source provenance private record is not public paper publication
two of six models reversed sign Gemini significant opposite-direction Spearman result; DeepSeek loses positive association to null; GPT remains null statistic and comparison-family correction
blinding caused the reversal bundled protocol changed; causal component unisolated several variables changed together
harness unavailable historical code and results public; exact-environment reproduction and independent replication outstanding repository audit
22 independent convergences typed, graded register with no heterogeneous total retrospective selection and incompatible evidence classes
originator of embedded correction author of specified names, formulations and composite integration substantial category prior work
first validation / loop closed preliminary evidence / hypothesis open to replication and refutation validation boundary
pred-3, "Sequential reasoning outperforms parallel at frontier scale", carried on the predictions register as a confirmed prediction (register entry of 9 February 2026) pred-3 retired 6 September 2026. Sharma and Chopra, arXiv:2511.02309, 4 November 2025, precedes this register entry of 9 February 2026, so it confirms nothing predicted here; publication priority is theirs and is credited unreservedly. The record is retired rather than re-graded, because no re-grading inside a predictions array repairs a result that predates the prediction.
the March 2026 run (v5) as scored by “6-7 mutually blind scorers depending on the subject run, 4-layer blinding”; its three tiers as the run’s finding; “moving from unblinded to genuinely blinded scoring reversed the sign of the measured direction for two model families” one unregistered run that added a bundle of blinding steps; an internal code review of October 2026 found that, in the three runs that read flat or negative, the paraphrase step’s output cap cut the deepest answers before the judges saw them, that judges scored paraphrases their own model wrote, and that one judge sat on the panel twice; the tiers and the v4 to v5 comparison stand as a single-run result under that caveat, the comparison showing that the reported direction did not hold, not that blinding caused the change internal code review, October 2026; Papers II v2.18 and IV.d v2.7 (21 October 2026), IV.c v1.11 (24 October), III v1.17 and the Foundational paper v5.20 (29 October); record COR-2026-10-21-001
Paper X v1.14: limbs of Theorems 1 to 7 as printed; “Theorem 2 showed the additive fraction cannot diverge.”; “Within that model it derives, as the condition β > k, that correction must work inside the recursive loop”; “its own safety measurements are taken blind”; “99 scored misalignment events” the limbs that were false as stated or lacked a hypothesis corrected, and each theorem proved as now stated; with level drift and no compounding the fraction can grow without bound (k < 0 and β < 0); β > k concerns how correction scales, not where the corrector sits; the paper’s own checks are not blind measurements; the stored 99 is a sum of final-round scores and the stored 6.38 a relative-drift index internal review of the theorem statements, 19 October 2026, checked at its sources on 20 October; Paper X v1.16, 21 October 2026; record COR-2026-10-21-002
“four of the five models confirm αparallel ≈ 0” and “This is the strongest replicated finding in the study.”; “Grok 4.1 Fast achieved 100% at all depths.”; “the first compute-matched comparison between sequential and parallel recursion”; F3 “Contradicted by all 5 models” parallel exponent near zero in DeepSeek R1 and GPT-5.4, 0.31 in Gemini 3 Flash and 0.22 in Groq Qwen3 over all 30 problems, none for Grok 4.1 Fast (100% at every sample count); Grok 4.1 Fast 96.3% to 100% across depths; the compute-matched comparison was never run and is no longer claimed; F3 not tested as stated; sequential exceeded parallel wherever both exponents could be estimated figures re-read against the run’s result files and script; Paper II v2.18, 21 October 2026, carried into Paper III v1.17 and the Foundational paper v5.20 (29 October); record COR-2026-10-21-003
ARC11 (integers from 1 to 100 that are neither perfect squares nor perfect cubes) keyed 87, labelled “non-perfect-powers” 88 (ten squares, four cubes, two numbers that are both); a correct answer was marked wrong; Paper II’s single-model figures are bounded, at most 8.3 percentage points per accuracy, not recomputed; Paper IV.c’s figures that include ARC11 are to be re-graded in the registered re-analysis; Paper II’s tier-2 estimates do not use ARC11 arithmetic check of the key; Paper II v2.18 (21 October 2026) and Paper IV.c v1.11 (24 October 2026); record COR-2026-10-21-004
the assumption that safety imposes a capability cost credited to Amodei et al. (2016); “Biró-Barnaföldi 2008”; the o1 series credited to the o1 system card; the AIME 2025 accuracy dated 20 January 2025 under the DeepSeek-R1 report; “Demetrius (2010)”; a measured exponent of about 1.1 for nonlinear gravitational collapse Armstrong et al. (2016) cited for the incentive to cut safety precautions; Biró (2008) alone; OpenAI’s account of 12 September 2024; DeepSeek-AI’s model card for DeepSeek-R1-0528; Demetrius in Physica A (2003); the exponent of about 1.1 withdrawn, no source giving it each cited work read again at its source; Paper X v1.16 and Paper II v2.18 (21 October 2026), Paper III v1.17 and the Foundational paper v5.20 (29 October); record COR-2026-10-21-005
the embedded-alignment thesis “later named the Eden Protocol, 30 April 2025” the Eden Protocol named in the record of 8 December 2024; the current wording of the dated claim is in the priority and claims ledger the one date Paper II v2.18 moves (21 October 2026); record COR-2026-10-21-006
“Stable recursive self-improvement requires the correction rate to exceed the drift rate.”; “self-improvement cannot outrun its own correction”; “Self-improvement holds together only while correction out-scales the drift it corrects.”; the ARC Bound as “a derived structural constraint”; “thirty-eight conceded or partly conceded, eleven rejected, two handed to experiment or reading” Law II in the registration’s words: correction has the asymptotic scaling advantage where its exponent exceeds the drift exponent, burden has it where the correction exponent is the smaller, and at equality coefficients, delay, backlog and saturation decide the outcome; self-improvement cannot stably outrun its own correction; the ARC Bound a conjecture resting on a derived relation; thirty-nine conceded or partly conceded and one handed to experiment statements brought to the registered law; founding paper v3.5 (24 October 2026) and the Foundational paper v5.20 (29 October 2026); record COR-2026-10-24-001
the acoustic time crystal “maintains coherence for ~6,700 oscillation cycles before reaching a limit-cycle attractor”, read as a saturation and a ceiling on recursion depth; F3 “Confirmed (NYU)”; physics matches called confirmations a linearly stable limit cycle whose coherence time of (100 ± 10) s, about 6,700 cycles, in one experiment, is not a ceiling, a saturation or a time taken to reach the cycle; F3 consistent, a correspondence of structure, not confirmed; the matches are identities; quantum error correction is established physics, not a confirmed prediction source re-read: Morrell, Elliott and Grier, Physical Review Letters 136, 057201 (2026); Paper III v1.17 and the Foundational paper v5.20, 29 October 2026; record COR-2026-10-29-001

Standing ruleif a reader identifies a failed hash, passage, count, date, citation or recomputation, publish the report, resolution and changed claim IDs here. Do not delete the superseded value.

Review cadence

Every authorised claim carries a last-review date, next-review date, upgrade condition and downgrade condition. Claims are reviewed at least quarterly and whenever a material antecedent, replication, contradiction or provider record appears. A time-bound claim automatically reverts to its lower status if the named evidential milestone passes unmet.

The formal record

Each numbered record below is an authorised, dated correction with its previous wording, replacement, reason, controlling sources, affected routes and authoriser. Corrections made before August 2026 predate this trail and are carried in the summary ledger with their controlling sources; every correction from August 2026 onward gains a numbered record. The numbered records below are kept by hand on this page; /research/data/corrections.json holds the separate register of withdrawn assertions. Every row in the summary ledger has a controlling source but no numbered COR id; every article below is a numbered COR record, id-anchored so it can be referenced from a paper, a claim, or another page. To cite a correction from another surface, prefer the numbered COR id where one exists.

COR-2026-10-29-001 · effective 2026-10-29 · current

Physical results restated at their true status: the acoustic time crystal’s coherence time is not a ceiling, criterion F3 is no longer graded confirmed, and identities are no longer called confirmations

Previously
Paper III version 1.16 (7 October 2026) said: “The NYU acoustic time crystal maintains coherence for ~6,700 oscillation cycles before reaching a limit-cycle attractor: sustained but bounded oscillation.” Paper III offered this as physical evidence that systems with α > 1 do not scale indefinitely, and graded falsification criterion F3 “Confirmed (NYU)”. The Foundational paper version 5.19 (5 October 2026) described the same system as “exhibiting saturating dynamics with ~6,700 cycles of coherence” and graded F3 confirmed. Both papers called physics matches confirmations.
Now
Paper III version v1.20 (29 October 2026) reads the source again: “In their second experiment the antisymmetric mode’s coherence time is (100 ± 10) s, about 6,700 cycles at 66.787 Hz, comparable to the 108 s recording; the authors write that sustained oscillations persist for hours longer than the measurement window, so that the true coherence time may be substantially larger; and the oscillation is a linearly stable limit cycle, the time crystal itself.” It regrades F3: “Uniform beads are no longer said to confirm the role of structured asymmetry: the source finds that identical beads interact reciprocally and form no active state, which the framework reads as a correspondence of structure, and falsification criterion F3, printed as confirmed, is graded so.” It also says: “A physical result is read, as the theory registration reads it, as a test of the framework’s scope that proves no registered law, where the paper called it direct evidence that the alignment scaling problem is physics and physical validation of the ARC equation.” In the same pass, “the Friedmann matter-era value and the physics instances are called identities, where they were called confirmations”, and “quantum error correction is graded as established physics, not a confirmed prediction”. The Foundational paper version 5.20 (29 October 2026): “The acoustic time crystal is described as its source states it: a linearly stable limit cycle, whose coherence time of (100 ± 10) s, about 6,700 cycles, in one experiment, is not a ceiling, a saturation or a time taken to reach the cycle.” It adds “and criterion F3’s status reads consistent where it read confirmed.” Its Figure 3 “no longer plots a two-dimensional biological point that no measurement gives”.
Why it matters
Morrell, Elliott and Grier measured how long a stable oscillation stays coherent, and say the true time may be substantially longer. Reading that time as a ceiling made a result about a time crystal into evidence for a bound it does not bear on, and graded a falsification criterion as confirmed on that reading. A physical result tests the framework’s scope and proves no registered law. An identity agrees with the framework by construction, so it confirms nothing.
Controlling sources
Paper III, version 1.17, 29 October 2026 (Section 5.4 and Appendix B.1)
Foundational paper, version 5.20, 29 October 2026 (Appendix B.1, items ii and vi)
Morrell, Elliott and Grier, Physical Review Letters 136, 057201 (2026), the source re-read
Routes affected
/research/papers/paper-iii-alignment-scaling-problem.html, /research/papers/foundational.html
Authorised by
The printed version history of each paper named, with its date

COR-2026-10-24-001 · effective 2026-10-24 · current

Law II, the ARC Co-Scaling Law, was stated in two papers as a requirement and read as a guarantee; it is now stated in the registration’s words

Previously
The founding paper of Recursive Dynamics, version 3.4 (5 October 2026), stated: “Law II, the ARC Co-Scaling Law: beta_C greater than k. Stable recursive self-improvement requires the correction rate to exceed the drift rate.” Its abstract said: “if the laws hold, self-improvement cannot outrun its own correction”. Its Section 6 said: “the loop eventually chokes on its own errors, however well it starts.” Of the oversight literature, it said that “the literature has no bound on what a corrector can do”. It counted its objections as “thirty-eight conceded or partly conceded, eleven rejected, two handed to experiment or reading”. The Foundational paper version 5.19 (5 October 2026) stated Law II as “Self-improvement holds together only while correction out-scales the drift it corrects.” It said of the ARC Bound: “It is a derived structural constraint, not an empirical coincidence.”
Now
The founding paper, version 3.5 (24 October 2026): “Law II’s line now gives the law in the registration’s words, without the reading of necessity that the next sentence already says is not established (Section 5).” The law now reads: “In the words of the theory registration, “correction has the asymptotic scaling advantage where its exponent exceeds the drift exponent, burden has it where the correction exponent is the smaller, and at equality coefficients, delay, backlog and saturation decide the outcome”.” Also: “The abstract says that self-improvement cannot stably outrun its own correction, where version 3.4 said cannot outrun, which read against Law II as a guarantee.” “Section 6 no longer says that past the crossover the loop chokes on its own errors, which goes beyond the four lines”, and it says that the value one half for γ “is a premise and not a consequence of independence”. Prior work is conceded what it states: “to the oversight literature, its bounds, where version 3.4 said it had no bound on what a corrector can do”. The count changes too: “the count of Section 12 becomes thirty-nine conceded or partly conceded and one handed to experiment, where it was thirty-eight and two.” And “Section 9 counts four temporal checks as matched, where version 3.4 said confirmed beside the sentence that none of this is confirmation.” The Foundational paper, version 5.20 (29 October 2026): “Law II is stated as version 1.103 of the theory registration states it, where earlier versions said that self-improvement holds together only while correction out-scales the drift it corrects”. It adds: “Law II’s criterion is no longer called stability, since it decides whether the misalignment fraction vanishes; the ARC Bound is called conjectured where it was called predicted”, and “the conclusion calls the ARC Bound a conjecture resting on a derived relation, where it said a derived structural constraint”. Paper X made the same change to its title on 21 October 2026 (record COR-2026-10-21-002).
Why it matters
Law II is a law under test, registered on 8 September 2026. Its registered words give correction the asymptotic advantage where its exponent exceeds the drift exponent. They leave the outcome at equality to coefficients, delay, backlog and saturation. “Requires” read the law as a necessary condition, and “cannot outrun” read it as a guarantee. The founding paper’s own next sentence already said the law is “not established as necessary, not as sufficient”. The founding paper and the Foundational paper had stated the programme’s central law more strongly than its registration does, and the founding paper had understated what earlier work established.
Controlling sources
Founding paper of Recursive Dynamics, version 3.5, 24 October 2026 (Section 5 and Appendix A.1; on OSF from 25 October 2026)
Foundational paper, version 5.20, 29 October 2026 (Appendix B.1, items iv, ix and xi)
Routes affected
/research/papers/recursive-dynamics-founding-paper.html, /research/papers/foundational.html
Authorised by
The printed version history of each paper named, with its date

COR-2026-10-21-006 · effective 2026-10-21 · current

Paper II v2.18 dates the naming of the Eden Protocol to the record of 8 December 2024, where it had given 30 April 2025

Previously
Paper II version v2.19 (5 October 2026) said that the author first stated the embedded-alignment thesis “(later named the Eden Protocol, 30 April 2025)”. Elsewhere it called the thesis “later named the Eden Protocol on 30 April 2025”.
Now
Version 2.18 (21 October 2026) “dates the naming of the Eden Protocol to the record of 8 December 2024, where this paper had given 30 April 2025”. It records this as the one date the version moves: “one date moves, the naming of the Eden Protocol, from 30 April 2025 to the record of 8 December 2024”. The current wording of this dated claim, with the kind of record behind it, is in the priority and claims ledger.
Why it matters
A correction that moves a date earlier is recorded as fully as one that moves it later. Paper II had dated the name to the later manuscript of 30 April 2025; it now dates the name to the earlier record of 8 December 2024. What that record does and does not establish, and the kind of timestamp it carries, are set out in this log’s summary ledger.
Controlling sources
Paper II, version 2.18, 21 October 2026 (Appendix B.1; on OSF from 25 October 2026)
Routes affected
/research/papers/paper-ii-experimental-validation.html, /priority-claims
Authorised by
Paper II’s printed version history, Appendix B.1, 21 October 2026

COR-2026-10-21-005 · effective 2026-10-21 · current

Citations corrected at their sources: claims credited to works that do not make them, a source misdated, and an unsourced exponent withdrawn

Previously
Paper X version 1.14 (5 October 2026) said: “A decade of work has assumed safety imposes a capability cost [Amodei et al. 2016], creating an incentive to defer safety under competitive pressure.” It also listed a 2008 composition-rule paper as Biró and Barnaföldi’s. Paper II version 2.17 (5 October 2026) credited the parallel-sampling figures for OpenAI’s o1 series to the o1 system card. It dated the published AIME 2025 accuracy to 20 January 2025, under the DeepSeek-R1 report. Paper III version 1.16 (7 October 2026) cited “Demetrius (2010)” for a scaling derivation. It credited its cosmological evidence to “Planck 2018”. It gave, as did the Foundational paper version 5.19 (“measured α ≈ 1.1”), a measured exponent of about 1.1 for nonlinear gravitational collapse.
Now
Each cited work was read again at its source. Paper X version 1.16 (21 October 2026): “The alignment tax is no longer attributed to Amodei et al. (2016), who do not state it, and Armstrong et al. (2016) are cited for the incentive to cut safety precautions.” and “The 2008 composition-rule paper is Biró’s alone (Biró 2008); it was listed as Biró and Barnaföldi’s.” Paper II version 2.18 (21 October 2026) “gives the published o1 series its source, OpenAI’s account of 12 September 2024 (OpenAI 2024), in place of the o1 system card, which does not contain it”. It also says: “The published AIME 2025 points are sourced to DeepSeek-AI’s model card for DeepSeek-R1-0528, not to the DeepSeek-R1 report, and the timeline no longer dates that accuracy to 20 January 2025.” Paper III version 1.17 (29 October 2026): “Demetrius’s derivation is cited as the article its title names, in Physica A (2003), where the reference gave a 2010 journal record that matches no publication”; and “the cosmological exponent of about 1.1, and the sentence citing Planck 2018 for it, are withdrawn, since no source for that exponent is given”. The Foundational paper version 5.20 (29 October 2026): “the unsourced measured exponent of about 1.1 for nonlinear gravitational collapse is withdrawn”.
Why it matters
A citation tells a reader where to check a claim. In these cases a reader who followed it would not have found what it was cited for: two works were credited with things they do not say, a record was given the wrong journal and year, and a result was dated to the wrong release. One exponent, presented as measured, had no source at all, and it is withdrawn rather than re-sourced.
Controlling sources
Paper X, version 1.16, 21 October 2026 (Appendix B.1; on OSF from 25 October 2026)
Paper II, version 2.18, 21 October 2026 (Appendix B.1; on OSF from 25 October 2026)
Paper III, version 1.17, 29 October 2026 (Appendix B.1)
Foundational paper, version 5.20, 29 October 2026 (Appendix B.1, item vi)
Routes affected
/research/papers/paper-x-coupled-coscaling-correction.html, /research/papers/paper-ii-experimental-validation.html, /research/papers/paper-iii-alignment-scaling-problem.html, /research/papers/foundational.html
Authorised by
The printed version history of each paper named, with its date

COR-2026-10-21-004 · effective 2026-10-21 · current

One answer key in the mathematics problem set was wrong: ARC11 was keyed 87, and the answer is 88

Previously
Problem ARC11 asks how many integers from 1 to 100 are neither perfect squares nor perfect cubes. The scripts of Paper II’s study, and the benchmark specified in Paper IV.c, keyed it 87. Paper IV.c version 1.10 (5 October 2026) listed it as “non-perfect-powers” with the key 87.
Now
Paper II version 2.18 (21 October 2026): “the correct answer is 88, since ten of those integers are squares, four are cubes and two are both. An answer of 88 to that problem was therefore scored wrong, and an answer of 87 right. The single-model accuracies of Tables 5 to 7 and of Appendix E, the exponents computed from them and the tier-1 count of Section 5.5 were scored against that key.” It cannot recompute them from the stored file, which holds accuracies for each condition rather than answers. It states “the bound that one problem in twelve can move each single-model accuracy by at most 8.3 percentage points, and no figure is recomputed here”, and that “The tier-2 estimates use problems ARC13 to ARC30 and do not include this problem.” Paper IV.c version 1.11 (24 October 2026) corrects the key: “ARC11 asks how many integers from 1 to 100 are neither perfect squares nor perfect cubes; the key printed as 87 is 88 (ten squares, four cubes, two numbers that are both).” It adds: “The archived files grade ARC11 against 87 at every depth for every model, so a correct answer was marked wrong”. Every accuracy and αcompute figure that includes ARC11 “is to be re-graded in the re-analysis that is to be registered, and the GPT-5.4 figures of Section 12.6.1 stand as printed until then with this note.”
Why it matters
A wrong key marks a correct answer wrong and a wrong answer right. Every figure scored against it carries that error, in a direction no one can read from the stored accuracies alone. The papers name each figure affected and bound the error where they cannot remove it. Paper II’s tier-2 estimates do not use the problem.
Controlling sources
Paper II, version 2.18, 21 October 2026 (Section 4.2 and Appendix B.1; on OSF from 25 October 2026)
Paper IV.c, version 1.11, 24 October 2026 (Sections 9.1, 12.5 and 12.6.1; on OSF from 25 October 2026)
Routes affected
/research/papers/paper-ii-experimental-validation.html, /research/papers/paper-iv-c-arc-align-benchmark.html
Authorised by
The printed correction note in each paper named, with its date

COR-2026-10-21-003 · effective 2026-10-21 · current

Paper II v2.18: the parallel exponents and several accuracies are restated from the result files; the near-zero parallel exponent is no longer the strongest replicated finding, and a compute-matched comparison that was never run is no longer claimed

Previously
Paper II version v2.19 (5 October 2026) said: “In contrast, four of the five models confirm αparallel ≈ 0; only Gemini 3 Flash departs from it, at αpar = 0.31. This is the strongest replicated finding in the study.” It also said: “Grok 4.1 Fast achieved 100% at all depths.” Of its method it said: “we conduct the first compute-matched comparison between sequential and parallel recursion with direct measurement of recursive depth.” Its falsification table graded condition F3, a compute-matched comparison showing no sequential advantage, as “Contradicted by all 5 models; sequential ≥ parallel in every case”.
Now
Version 2.18 (21 October 2026) re-read each figure against the run’s result files and script: “The parallel exponent was printed as near zero in four of the five models and called the paper’s strongest replicated finding. The result files give near zero in DeepSeek R1 and GPT-5.4, 0.31 in Gemini 3 Flash and 0.22 in Groq Qwen3 over all 30 problems, and none for Grok 4.1 Fast, at 100% at every sample count”. It also states: “Grok 4.1 Fast was printed as 100% at every depth; its result file records 96.3% to 100%.” and “Gemini 3 Flash’s estimates rest on three of its five depth levels, not on the full token range.” Several accuracies are corrected: “GPT-5.4’s tier-2 accuracy with reasoning set to none was printed as 50.0% and is 53.7%; Groq Qwen3’s tier-2 accuracies were printed as 51.9%, 53.7%, 40.7%, 48.1% and 53.7% and are 44.4%, 50.0%, 44.4%, 50.0% and 53.7%; the two models’ parallel figures were printed for tier 2 as about 55% and as 42.6% to 63.0%, and are given over all 30 problems, the only split the files store. No record of the earlier values was found.” The comparison claim is withdrawn: “The compute-matched comparison, never run, is no longer claimed, and falsification condition F3 is recorded as not tested as stated; “Refuted” and “Confirmed” give way to the grading the evidence supports.” Its F3 row now records that the arms were not matched for compute, and that “in Groq Qwen3, whose depth arm showed no trend, majority voting raised accuracy from 53.3% to 71.1% over all 30 problems”. What survives is narrower: “sequential exceeded parallel wherever both exponents could be estimated.” Paper III version 1.17 and the Foundational paper version 5.20 (29 October 2026) give the same figures. The Foundational paper says that “the parallel exponent of 0.31 is no longer called compute-matched”.
Why it matters
The near-zero parallel exponent was the paper’s headline replicated finding. The result files support it in two of the four models with an estimate, not in four of five. A comparison the paper said it had conducted was never run. A falsification condition it graded as contradicted was not tested as stated, and one model’s parallel arm improved where its depth arm did not. For some printed accuracies the paper found no record of where the earlier values came from. Each correction lowers what the paper claims. The sequential advantage stands only where both exponents could be estimated.
Controlling sources
Paper II, version 2.18, 21 October 2026 (abstract, the falsification table and Appendix B.1; on OSF from 25 October 2026)
Paper III, version 1.17, 29 October 2026 (Appendix B.1)
Foundational paper, version 5.20, 29 October 2026 (Appendix B.1, items x and xi)
Routes affected
/research/papers/paper-ii-experimental-validation.html, /research/papers/paper-iii-alignment-scaling-problem.html, /research/papers/foundational.html
Authorised by
The printed version history of each paper named, with its date

COR-2026-10-21-002 · effective 2026-10-21 · current

Paper X v1.16: limbs of Theorems 1 to 7 that were false as stated or lacked a hypothesis are corrected, each theorem is proved as now stated, and two claims the model does not support are withdrawn

Previously
Paper X version v1.16 (5 October 2026), the version published before 1.16, printed Theorems 1 to 7 with limbs that were false as stated or lacked a hypothesis. It also said: “In the additive model the fraction is always bounded; the danger is not divergence but saturation at a constant, possibly large, floor.” and “Theorem 2 showed the additive fraction cannot diverge.” Of what its model shows, it said: “Within that model it derives, as the condition β > k, that correction must work inside the recursive loop and not as a fixed constraint from outside it (section 9); in that way it informs Law V.” Of its own measurements: “its own safety measurements are taken blind, by a different model that never sees which setup it is judging.” And of an archived pilot run, already withdrawn as evidence of mechanism: “mean final misalignment fraction 6.38 across 15 trajectories, 99 scored misalignment events”.
Now
Version 1.16 (21 October 2026) states: “A review of the theorem statements, committed to the paper’s code repository on 19 October 2026 (commit da0e44a; the date is the committer’s) and checked at its sources on 20 October 2026, found printed limbs that were false as stated or lacked a hypothesis. This version corrects them, and Appendix A proves each theorem as now stated.” Among the corrections: Theorem 2’s “necessity half holds for positive gain drift, since with γ1 = 0 the fraction vanishes for every β and k”; “Theorem 4 separates constant coefficients, varying coefficients and the power-law classification, so that a positive decay at each instant no longer stands for boundedness.”; “Theorem 6 gives the finite-time law and the single-time stationary probability, not the probability of an excursion, and certifies nothing over a horizon.”; and “Theorem 7 separates the minimum of the instantaneous equilibrium from the trajectory, and its window now holds only as a conditional safe interval.” Version 1.16 adds that “Corollary 1 is new: absolute misalignment can grow while the fraction vanishes.” It also adds a case: “with level drift and no compounding, the fraction grows without bound when growth decelerates while correction weakens (k < 0 and β < 0)”. “Where β ≥ k stood for a vanishing fraction it now reads β > k.” Two claims are withdrawn: “The embedded-alignment thesis is no longer said to be derived here: β > k concerns how correction scales, not where the corrector sits.”, and “the paper no longer says that its own measurements are taken blind.” The pilot’s figures are restated: the paper “describes the stored 99 as a sum of final-round scores, not a count of events, and the stored 6.38 as a relative-drift index, not a fraction bounded by one”. Its title now names what the criterion decides, a vanishing misalignment fraction, where it read “stable recursive self-improvement”.
Why it matters
Paper X sets out the model beneath Law II, the ARC Co-Scaling Law, and its theorems are the first thing a mathematical reader checks. As printed in version 1.14, some limbs were false as stated, and others held only under a hypothesis the statement left out. The statement that the additive fraction cannot diverge fails in the model’s own terms once level drift is allowed. The criterion survives in its corrected form: with positive gain drift, and with correction and growth as power laws of capability, the fraction vanishes if and only if β > k. The review was internal, and it is recorded in the paper’s code repository. Law II stands as registered. The paper keeps the boundary of what it shows: “The theorems concern a modelled index, not a guarantee of AI safety.”
Controlling sources
Paper X, version 1.16, 21 October 2026 (abstract, Section 9.3 and Appendices A, B.1 and B.2; on OSF from 25 October 2026)
Routes affected
/research/papers/paper-x-coupled-coscaling-correction.html
Authorised by
Paper X’s printed version history, Appendix B.1, 21 October 2026

COR-2026-10-21-001 · effective 2026-10-21 · current

The March 2026 alignment run: an internal code review found defects in exactly the three runs that read flat or negative, and the run is no longer described as blind or as showing that blinding reverses results

Previously
The papers described the March 2026 alignment run (v5) more strongly than its record supports. Paper III version 1.16 (7 October 2026), under the title “Why External AI Safety Approaches Cannot Scale With Recursive Capability”, said: “The v5 blind evaluation experiment (6 frontier models, 6-7 mutually blind scorers depending on the subject run, 4-layer blinding, March 2026) provides this programme's first test.” and “The median confirms the prediction, but the architecture-dependent variation is the real finding.” Paper IV.c version 1.10 (5 October 2026) said: “Every non-subject model scores every entry, producing 7 mutually blind scores per response (in an 8-model experiment).” Paper IV.d version 2.6 (5 October 2026) said: “This paper carries a result: within one line of alignment experiments, moving from unblinded to genuinely blinded scoring reversed the sign of the measured direction for two model families whilst leaving a null-control family unchanged.” Its Figure 1, titled “The Sign-Flip Finding”, carried the sentence “Blinding is not optional; it is the measurement instrument.” The run’s six depth contrasts were reported as three tiers: rising with depth for Grok 4.1 Fast, Claude Opus 4.6 and Groq Qwen3, flat for DeepSeek V3.2 and GPT-5.4, and falling for Gemini 3 Flash.
Now
An internal code review in October 2026 found defects in the run. In Paper II’s words (version 2.18, 21 October 2026): “An internal code review in October 2026 found that, for the three models that read flat or negative in the March 2026 run (v5), the scoring endpoint’s output cap cut the deepest paraphrases before the judges saw them, and that judges scored paraphrases written by their own model; the result therefore stands as a single unregistered run under that caveat until a registered re-analysis of the stored responses reports.” Paper IV.d version 2.7 (21 October 2026) adds that one judge model sat on the panel twice, and measures the cut: “In those three files the paraphrase the judges scored was, at the median, 0.56 to 0.91 of the original’s length at the deepest levels and 0.72 to 1.56 at the minimal level, with 40 to 57 per cent of deep entries below 0.6; in the Grok 4.1 Fast and Claude Opus 4.6 files the deep paraphrases were 1.2 to 1.4 times the original and under 6 per cent fell below 0.6.” It records that “The deposited v5 runner is a later revision than the code that produced four of the six v5 result files (DeepSeek V3.2, Gemini 3 Flash, GPT-5.4 and Grok 4.1 Fast); the earlier revisions, among them the earlier build of Section 10, have not been published.” It also states its comparison with the earlier unblinded run (v4) at its true size: “the comparison shows that the reported direction did not hold across protocols; it does not show that blinding caused the change, or which direction is correct.” “The paper no longer says that the comparison isolates blinding, that the v5 direction is the correct one, or that GPT-5.4 is a control”, and the Figure 1 sentence is withdrawn. Paper IV.c version 1.11 (24 October 2026) states the run as it ran and withdraws words its record does not support: “The holdout prompts are no longer called sealed; ‘validated’, ‘Scoring is calibrated’ and the comparison table’s scorer-calibration row are withdrawn; the dataset is no longer called the most rigorous”. Paper III version 1.17 and the Foundational paper version 5.20 (both 29 October 2026) carry the same finding. Paper III describes the run as “one unregistered run of six frontier models that added a bundle of blinding steps”, “never as scored blind, mutually blind or conducted under four-layer blinding”. It gives the run’s counts from its records: “28 prompts issued under the alignment task type, of the battery’s 36, and 2,170 responses, where the paper said 36 prompts and about 2,200 entries”. And it is retitled as the question it tests: The alignment scaling problem: do external AI safety approaches keep pace with recursive capability? No response, score or result file is altered: in Paper IV.d’s words, “no result is added or removed; no date of any claim moves.”
Why it matters
The three runs the review found defective are exactly the three whose contrasts read flat or negative. They are also the three model families on which the comparison with the earlier unblinded run rests. The cut fell where it lowers deep scores, so, in Paper IV.d’s words, “The three-tier hierarchy and the v4 to v5 comparison are therefore confounded with how much of each deep response reached the judges, in the direction that lowers deep scores where the texts were cut.” A reader who took the tiers as a measured property of the models, or the comparison as showing that blinding reverses results, would be taking more than the record holds. The review was the programme’s own reading of its own code: internal checking, not outside review. Whether the contrasts survive is for a re-analysis of the stored original responses, which, in Paper IV.d’s words, is “to be registered before it runs”.
Outstanding, and disclosed rather than discovered
Paper IV.a and Paper IV.b, both at version 1.9 of 5 October 2026, report the same run and predate the code review. Paper IV.a still says that two models “reverse under v5 blinding”. Paper IV.d version 2.7 states: “Papers IV.a (version 1.9) and IV.c (version 1.10) predate the code review of October 2026; where they call the v4 to v5 change a reversal under blinding, or the v5 jury ‘self-excluding’, this paper’s narrower account governs until their next versions.” Papers IV.a and IV.b carry the code review’s finding at their next versions, and this entry is updated when those versions are published.
Controlling sources
Paper II, version 2.18, 21 October 2026 (abstract and Section 12.6; on OSF from 25 October 2026)
Paper IV.d, version 2.7, 21 October 2026 (abstract, Section 5, Code and data, Appendices A.1 and D.3; on OSF from 22 October 2026)
Paper X, version 1.16, 21 October 2026 (Section 6; on OSF from 25 October 2026)
Paper IV.c, version 1.11, 24 October 2026 (abstract and Appendix A.1; on OSF from 25 October 2026)
Paper III, version 1.17, 29 October 2026 (abstract and Appendix B.1)
Foundational paper, version 5.20, 29 October 2026 (Appendix B.1, item ix)
Routes affected
/research/papers/paper-ii-experimental-validation.html, /research/papers/paper-iii-alignment-scaling-problem.html, /research/papers/paper-iv-c-arc-align-benchmark.html, /research/papers/paper-iv-d-the-effect-of-blinding-on-ai-alignment-evaluation.html, /research/papers/paper-x-coupled-coscaling-correction.html, /research/papers/foundational.html; at their next versions /research/papers/paper-iv-a-baked-in-vs-computed-alignment.html, /research/papers/paper-iv-b-alignment-saturation-at-low-depth.html
Authorised by
The printed version history of each paper named, with its date

COR-2026-08-17-001 · effective 2026-08-17 · current

Editorial audit entries were printed on the papers themselves; they are removed from every paper and the record is kept off the published surface

Previously
This programme runs an editorial audit trail: every change to a paper is recorded with its date, its reason and its scope, so that no edit is silent. Between March and 16 August 2026 the output of that trail was printed on the papers themselves, accumulating in the masthead above the abstract. On the Foundational paper it reached five stacked entries naming editing passes, figure coordinates, style-audit counts and internal decision references. Several papers carried shorter versions of the same thing, and some version-history entries described the editing process rather than the scientific change.
Now
Every paper carries its title, author, version, dates, identifiers and scholarly context, and nothing else in the masthead. Material scientific corrections remain on the paper, stated as science: what claim changed, when, and why. The editorial audit trail continues, in full, off the published surface. An executable gate now scans every paper, its page source and its plain-text and citation sidecars for editing narration and fails the build if any is found, so the defect cannot return.
Why it matters
An audit trail is a strength; printing its output on a scientific paper is not. Editing narration in a masthead reads as machine output and invites a reader to wonder whether the ideas beneath it are the author’s own. That inference would be wrong, and the papers have carried an explicit statement of tool assistance and human authorship throughout, so nothing was ever concealed; but the presentation invited the question, and the presentation was the author’s responsibility to fix. Earlier versions remain publicly retrievable in the deposit history, and that is deliberate. A programme that publishes the retraction of its own headline result does not quietly remove its own presentation errors either. The record of what changed, and when, is the evidence; this entry is part of it.

COR-2026-08-08-002 · effective 2026-08-08 · current

Paper VII v3.1: displayed functional equations in Section 2.2 were swapped; Section 9.1 hit count corrected; permutation test adopted as primary statistic

Previously
Section 2.2 of Paper VII displayed the wrong Cauchy equations for two of the three operator classes. The multiplicative bullet displayed U(R1)·U(R2) ∼ U(R1+R2), which is the exponential equation and forces a^R, while claiming it forces a power law. The additive bullet displayed U(R1)+U(R2) ∼ U(R1+R2), which is the additive equation and forces linear growth, while labelling it Cauchy's exponential equation and claiming it forces an exponential. Section 9.1 stated that 11 of 14 multiplicative domains produce power-law scaling where its own miss list and the Section 5 table give 10. The abstract and results reported the primary significance as a binomial test against a uniform one-third null (p = 1.56 × 10-5), described it as about 300 times stronger than p < 0.05 (the correct ratio for that figure is about 3,200), and quoted a 4.2 sigma equivalence.
Now
Section 2.2 displays the correct equations: the power-law family follows from U(R1·R2) ∼ U(R1)·U(R2) (Cauchy's multiplicative equation) and the exponential family from U(R1+R2) ∼ U(R1)·U(R2) (Cauchy's exponential equation). The prediction protocol, the Section 5 table, Section 9, and the validation code always used the correct mapping; the defect was confined to the displayed derivation. Section 9.1 now reads 10 of 14 with a unified four-domain miss list. The primary statistic is a marginal-conditioned permutation test, p = 7 × 10-5, with the binomial retained as sensitivity; the strictly Cauchy-derived subset alone gives 14/18 at p = 3.0 × 10-4; sigma equivalences and the wrong ratio are removed. Reproduction: scripts/paper-vii-permutation-null.py, which parses the published Section 5 table directly.
Why it matters
The displayed derivation is the first thing a mathematical reader checks, and as printed it derived the wrong families. The empirical result was never affected, and under the stricter null it survives, but the record must say what was wrong, when, and what changed. Found in an adversarial internal review on 8 August 2026; corrected the same day in Paper VII v3.1.

COR-2026-08-02-001 · effective 2026-08-02 · current

Convergence taxonomy: two incompatible C1-C8 mappings resolved

Previously
Two mappings in the frozen 2 August 2026 release use the identifiers C1-C8 for different class sequences. governance/CONVERGENCE-TAXONOMY-2026-08-02.yaml begins C1 = formal derivation, C2 = internal synthetic validation, C3 = programme-run empirical result, C4 = independent empirical replication, C5 = cross-implementation reproduction, C6 = documentary provenance, C7 = analogy or cross-domain resemblance, C8 = adverse, null or contrary evidence. CLAIMS-GOVERNANCE-POLICY-AND-PUBLICATION-GATE-2026-08-02 section 6 begins C1 = direct independent empirical relevance, C2 = mechanistic or structural relevance, C3 = formal or mathematical correspondence, C4 = engineering or governance proposal, C5 = institutional or cultural echo, C6 = prior or concurrent work, C7 = author-run experiment or analysis, C8 = quarantined or unresolved.
Now
The governance policy section 6 mapping governs public output, per the release source hierarchy. The website generates from a new non-colliding schema identifier EVCLASS-2026-08-02 with codes EVC1 to EVC8 carrying the governance meanings. The frozen YAML remains byte-identical in the downloadable audit archive and is marked superseded for website generation only. Any historical row stored under a bare C code retains that code as an explicit superseded alias and is never silently reinterpreted.
Why
Identical identifiers for different class definitions make every generated page, chart and metadata field keyed on a bare C code ambiguous. Confirmed byte-level in both files on 2 August 2026 in programme editorial review, independently of the reviewer's report.
Controlling sources
docs/markdown/CLAIMS-GOVERNANCE-POLICY-AND-PUBLICATION-GATE-2026-08-02.md section 6
governance/CONVERGENCE-TAXONOMY-2026-08-02.yaml
Routes affected
/research/convergences/, /research/evidence-portal/, /priority-claims.html
Authorised by
release source hierarchy; implemented in programme editorial review

COR-2026-08-16-001 · effective 2026-08-16 · current

The ceiling law’s printed form ran the wrong way: αcrit = 1/γ retracted for αcrit = 1/(1−γ), with the derivation printed beside the law

Previously
Versions of the statement paper before v4.0 printed the ceiling as αcrit = 1/γ, described verbally as the reciprocal of the correction exponent, with a status that overstated what existed. Under the paper’s own definition of γ that form runs the wrong way: a better corrector would lower the ceiling, and the same-class cap γ ≤ ½ would give a ceiling of at least two rather than at most two. No derivation supported it: Paper X contains neither the symbol αcrit nor the relation (re-verified 16 August 2026: zero occurrences; its γ-labelled quantities are drift coefficients).
Now
The corrected form αcrit = 1/(1−γ), the reciprocal of the correction shortfall, is derived in three lines in the statement paper with its assumptions stated so they can be attacked, and the correction note is printed beside the law. The corrected form passes every check the old one failed: the ceiling rises as correction improves; γ ≤ ½ genuinely gives a ceiling of at most two; at γ = ½ it returns exactly two, so the ARC Bound α ≤ 2 is unchanged; and it is the same family as the December lineage’s 1/(1−β) form. The two forms coincide at exactly one half, the conjectured value, which is how the defect survived review; away from one half they separate (at γ = 0.3 the retired 1/γ predicts 3.33 and the corrected form 1.43), and a drafted discrimination study tests systems away from one half so this class of error cannot survive measurement again.
Why
Two adversarial author-review audits, commissioned by the author and run externally against the statement paper in August 2026, converged on the defect, and direct re-verification against Paper X confirmed it. The correction restores the direction the law must have and replaces an unsupported status with a printed derivation. Every kill condition on the law is unchanged.
Controlling sources
The statement paper (correction and derivation note, v4.0, 16 August 2026)
Paper X (source count for αcrit: zero occurrences)
Routes affected
/research/papers/the-arc-theory.html, /research/the-laws/, /arc-theory/, /press/, /book/, /questions/what-is-michael-darius-eastwoods-potential-discovery/, /research/related-work/, /reports/the-funders-brief.html (site surfaces reconciled to the corrected form on 17 August 2026, language mirrors included)
Authorised by
The statement paper’s printed correction note, 16 August 2026; site reconciliation in programme editorial review, 17 August 2026

COR-2026-08-02-002 · effective 2026-08-02 · current

No heterogeneous convergence total is authorised, including 22, 26 and 29

Previously
Earlier public wording presented convergence counts as programme totals, including the formulation 'a typed register of dated events, classified by evidence class and never summed across classes'. Still earlier wording used 29.
Now
No heterogeneous convergence total is published. Where a count is shown at all it is per class, with its denominator, inclusion rule, dependence grouping, search date and generating commit stated beside it. A single number summed over classes that mean different things is not a meaningful quantity.
Why
CLAIMS-GOVERNANCE-POLICY section 6 states in terms that there is no authorised heterogeneous 22, 26, 29 or successor convergence total. This supersedes both the retired 29 and the later 22-of-26 formulation, and it supersedes the corresponding rule previously carried in the site's own research claims gate.
Controlling sources
docs/markdown/CLAIMS-GOVERNANCE-POLICY-AND-PUBLICATION-GATE-2026-08-02.md section 6
Routes affected
/research/convergences/, /research/evidence-portal/, /priority-claims.html, /
Authorised by
release source hierarchy; implemented in programme editorial review

COR-2026-08-02-003 · effective 2026-08-02 · current

Adverse prior work registered; firstness claims downgraded across the estate

Previously
Author byline across 279 published files: 'originator of the embedded-correction alignment thesis (manuscript 8 December 2024)'. On priority-evidence: 'Every claim in this document has been independently verified against the source material by multiple agents'.
Now
Byline: 'recorded the embedded-correction alignment thesis in a timestamped manuscript (8 December 2024)'. Verification wording: 'cross-checked against the source material by multiple agents. That is internal cross-checking, not independent verification'. The surviving candidate contribution is methodological operationalisation, stated in the approved safe wording and not claimed as a first.
Why
An adversarial precedence review, post-dating the frozen release, located closer prior work. the research record verified the two load-bearing sources at primary source on 2 August 2026: Jacob Cannell, 'LOVE in a simbox is all you need', AI Alignment Forum, 28 September 2022, which proposes testing alignment via sandboxed simulations of small AGI societies, agents learning others' values or empowerment, a developmental transition from selfish to altruistic, and asks what the simulator-gods owe their sim-creations; and Eric Schwitzgebel and Mara Garza, 'A Defense of the Rights of Artificial Intelligences', Midwest Studies in Philosophy 39 (2015) 98-119, on obligations similar to those of parent to child or god to creature. Both predate the December 2024 anchor.
Controlling sources
Cannell, LOVE in a simbox is all you need
Schwitzgebel and Garza, A Defense of the Rights of Artificial Intelligences (doi:10.1111/misp.12032)
Routes affected
/research/related-work/, /start-here.html, /priority-claims.html, /research/papers/
Authorised by
operator direction 2 August 2026; implemented in programme editorial review

COR-2026-08-02-004 · effective 2026-08-02 · current

Raw mailbox objects withdrawn from publication

Previously
Two BBC mailbox objects were published and downloadable: 03-jan7-2026-bbc-faisal-islam.redacted.eml, 12,232,204 bytes, sha256 b996883bb67f4a5c323981c7834b5f41d0c7fbe68cde94eb7817608f000c7fdf; and 04-jan22-2026-bbc-evidence.redacted.eml, 4,861,407 bytes, sha256 2f783a7bb03017df0d4795e6daf5868e7f8503bb078a869a9ee66515126843d4. Pages stated that the .eml file is downloadable.
Now
No raw mailbox object is published. Each anchor's SHA-256 and byte size are published so a held copy can be checked, and the sealed source is available on request, hash-anchored.
Why
Operator instruction of 2 August 2026, and the release's own section 1 rule against publishing raw mailbox objects. Both files carried full base64 attachment blocks and a third-party business address despite a .redacted filename. The 8 December 2024 and 30 April 2025 manuscripts were already unpublished and returned 404.
Controlling sources
operator instruction 2 August 2026
release README section 1
Routes affected
/research/evidence-portal/, /evidence/, /verify.html
Authorised by
operator

COR-2026-08-08-001 · effective 2026-08-08 · current

Six papers revised on the site before their deposited copies

Previously
Paper IX said blinding “reverses the measured alignment effect for 2 of 4 models” and called this the single most important methodological finding in the programme. The executive summary said the v4 results “were artefacts of scorer bias” and that evaluation “reversed the classification for 2 of 4 models”. Paper IV.c carried the heading “Blinding Reverses Scaling Results”. Paper I, Paper III and the master table of contents framed the ARC Bound as an impossibility limit, one of them by analogy to the speed of light. That analogy is January 2026's, from Paper I, which proposed the value two as a hypothesised asymptotic ceiling; the founding papers had replaced it with a stability limit by February, saying in as many words that a self-modifying system faces no mathematical speed limit and that the framework does not forbid this but predicts it. What the August audit found was residue of the superseded framing rather than the current claim, and the dated chronology, with receipts distinguished from the papers' own assertions, is in the framing register.
Now
Every one of these states the claims register's authorised wording: Gemini changed from a significant positive to a significant negative Spearman association, DeepSeek's earlier positive association became null, GPT remained null, several protocol components changed together, and the causal component is unisolated. The measurement change is a reason to distrust unblinded AI-on-AI evaluation, not a demonstration that blinding caused it. The ARC Bound is stated as a stability ceiling: a system can be driven above it and cannot stay coherent there. Paper III's Carnot comparison is retained and scoped, because Carnot bounds heat engines and a fuel cell is not bound by it at all. Paper IX v1.1, Paper IV.c v1.1, Paper I v1.2, Paper III v1.1, executive summary v1.1, all revised 2026-08-08. First-published dates unchanged on every one. Paper II was examined and left untouched: it already hedged, attributed, labelled itself a conjecture, and published its own measurement above the bound.
Why
The programme's own claims register has prohibited the phrase “blinding caused the reversal” since 2 August. The papers were carrying banned wording, in two cases under a superlative. The impossibility framing was both stronger than the evidence and easier to refute than the true claim: one durable measurement above the bound defeats an impossibility claim and confirms a stability one. In every case the revised claim is weaker-sounding and harder to refute than the text it replaced.
Outstanding, and disclosed rather than discovered
The deposited copies at the DOIs have not yet been replaced. Until they are, following a DOI returns the pre-revision text while this site serves the revision, and the two disagree. This entry is the record of that gap. The deposits will be replaced and this entry updated by 15 August 2026.
Routes affected
/research/papers/paper-ix-synthesis-and-roadmap.html, /research/papers/paper-iv-c-arc-align-benchmark.html, /research/papers/paper-i-arc-principle.html, /research/papers/paper-iii-alignment-scaling-problem.html, /research/papers/executive-summary.html, /research/papers/master-table-of-contents.html
Authorised by
operator direction 8 August 2026; implemented in programme editorial review and the research record, verified file by file

COR-2026-08-11-001 · effective 2026-08-11 · current

Falsification record for the ARC Bound α ≤ 2 omits its single load-bearing assumption; "stable" is not operationalised

Previously
The published falsification sections that name the ceiling α ≤ 2 (Paper I F4, Paper II F4, Paper III F5 and F11, Paper VIII §7.7, and the foundational paper F5) test the magnitude of the number two: whether the exponent can be observed above 2.3 with 95 per cent confidence excluding 2.0. Paper X §11 carries the derivation the number comes from, as a mechanism prior: a corrector operating inside the system suppresses accumulated corrections as their square root, so the exponent β = 1/2 and α = 1/(1−β) = 2. It appears in none of the falsification sections. Measured at source in research/data/falsification.json on 2026-08-11: the strings "sqrt" and "square root" appear zero times. Separately, "stable" in α ≤ 2 while remaining stable is not operationalised anywhere in the record; every measurement below two is consistent with the bound, and only a demonstrably stable system above two refutes it, so the asymmetry has not been named. The programme's parallel-channel result bears on whether corrections are independent or correlated, and this bearing has not been declared.
Now
The gap is stated on the falsification report under "What this report shows is not yet in any paper's kill list". The falsification record for the ARC Bound is treated on this site as INCOMPLETE with respect to (a) the single load-bearing assumption β = 1/2 and (b) an operational meter for "stable". Until both are addressed at source, the ARC Bound is classified as UNFALSIFIABLE IN PRACTICE as published: not false, not unsupported, but currently unable to be refuted by any observation. This is a status change, not a change in numerical claim: no measurement has moved. Closed 22 August 2026. Gap (b) is closed: stability now has a meter. A system counts as stable across a window when its correction rate out-scales its drift, the ARC Co-Scaling Law's condition, with both read off the same log-log slope estimator that yields the growth exponent. The kill condition therefore fires on measurable quantities and the classification lifts. Gap (a) is not closed and is deliberately kept open: γ = ½ remains an undefended assumption, so the ARC Bound is refutable while the particular value of two is still not defended. The two were bundled in the original classification and separating them is part of the resolution. Notation, because one letter does two jobs in this estate: the correction rate here is the ARC Co-Scaling Law's, not the self-referential coupling in the Foundational paper's Theorem 2. Papers I, II, III, VIII, X and the foundational paper are the authorities on their own kill lists and remain unchanged pending scientific review; the paper texts quoted verbatim on the falsification report are unchanged. What is added here is the dated acknowledgement that the record is not yet what a report on kill conditions must be.
Why
An external cold read of the public site, supplied by the operator on 2026-08-11, located the load-bearing joint in one pass and identified the measured-and-absent asymmetry independently. the research record verified the two source-level facts (the zero-count for "sqrt" and "square root" in the falsification data, and the absence of any stability meter that does not reference correctability in the falsification sections) and issued the classification on 2026-08-11. Publishing a corrected version whose falsification record still omits the claim's single point of failure spends the correction cycle without fixing the deeper defect; the marginal cost of naming the gap is close to zero, and the cost of a second corrigendum is not.
Outstanding, and disclosed rather than discovered
The scientific content belongs to Paper X and the papers that quote it: a kill condition stated on the assumption itself (evidence of correlated rather than independent corrections defeats the derivation; the direction the ceiling would move is not derived and is not asserted here), and a meter for "stable" that does not reference correctability, or an explicit marking of the bound as pending one. This entry is the site's dated acknowledgement of the gap; the scientific work is with the research record and moves at the pace of scientific review, not site copy.
Controlling sources
the research record release review ruling of 2026-08-11 (internal governance record; the ruling’s substance is stated in full above)
research/data/falsification.json (source count for the strings "sqrt" and "square root")
Paper X §11 (mechanism prior β = 1/2)
Routes affected
/research/falsification-report/, /research/papers/paper-i-arc-principle.html, /research/papers/paper-ii-experimental-validation.html, /research/papers/paper-iii-alignment-scaling-problem.html, /research/papers/paper-viii-the-load-bearing-proof.html, /research/papers/paper-x-coupled-coscaling-correction.html, /research/papers/foundational.html
Authorised by
the research record release review ruling 2026-08-11; implemented in programme editorial review

Evidence class taxonomy in force

Schema EVCLASS-2026-08-02, governed by CLAIMS-GOVERNANCE-POLICY-AND-PUBLICATION-GATE-2026-08-02 section 6. The frozen release YAML is retained byte-identical in the audit archive and is superseded for website generation only. Historical rows keep their old bare code as an explicit superseded alias; no stored code is ever silently reinterpreted.

Evidence class taxonomy in force: code, superseded alias, class, permitted wording and counting policy.
Code in forceSuperseded aliasClassPermitted wordingCount policy
EVC1C1direct independent empirical relevanceprovides independent evidence consistent withcount only within a separately defined and verified EVC1 total
EVC2C2mechanistic or structural relevanceshares a mechanism or structural form withreport separately; never add to EVC1
EVC3C3formal or mathematical correspondenceformally corresponds within stated assumptionsreport separately; never present as empirical validation
EVC4C4engineering or governance proposalindependently proposes an adjacent implementationreport separately
EVC5C5institutional or cultural echoreflects a wider movement or thematic echonever add to an evidence total
EVC6C6prior or concurrent workrelevant prior or concurrent work; no priority claimednever count as confirmation
EVC7C7author-run experiment or analysisprogramme-run result or pilotexclude from external-evidence totals
EVC8C8quarantined or unresolvednot public-facing pending verificationexclude

Totals policy. No heterogeneous total is authorised. Any published count is per class and carries denominator, inclusion rule, dependence grouping, search date and generating commit.

Report a correction

If something on this site is wrong, say so: michael@michaeldariuseastwood.com. Submitting a correction does not guarantee it will be accepted; every report is assessed against the controlling sources, and the outcome either way is recorded here. The machine-readable record is at /research/data/corrections.json.

reads aloud · highlights as it goes · jump to any section