The evidence record

Four questions, four kinds of evidence

Before you read a single hash, know what you’re actually asking. Most arguments about this record collapse four different questions into one, and then answer none of them well. Here they are, separated, each with the only kind of evidence that can answer it.

A hash answers “are these the same bytes?” It does not answer “when did these bytes first exist?”, “who wrote every sentence?”, “was the idea new?” or “is the theory true?”

Standards: RFC 5322, RFC 6376, RFC 8617, RFC 3161, NIST SP 800-107 Rev. 1.

December source record

It’s a Sunday, just before three in the morning, December 2024. A man emails his own manuscripts to himself. Nothing dramatic happens. That’s rather the point: the record below is what an ordinary act leaves behind, measured as precisely as I know how. What it proves, and what it doesn’t, is stated underneath it. Read the boundaries as carefully as the claims.

The public sent-side export is sparse. The supplied recipient-form export’s headers record delivery to the separately addressed Gmail account; it contains Received, X-Received, Authentication-Results, sender-domain DKIM, X-Google-DKIM-Signature and an ARC set bearing d=google.com, and it reproduces every decoded attachment hash. Its time evidence is the sending server’s signature and Google’s ARC seal; it holds no token from a separate timestamping authority. The sender-domain body and RSA and paired AMS and Seal mathematics verify against keys returned at the signed selectors by Google Public DNS and Cloudflare DoH on 2 August 2026. The ARC pair is the more probative provider-key element under the audit-date Google-selector binding, consistent with provider-generated ARC. No evidence indicates author control of the corresponding private key, but historical key publication, exclusive Google custody, compromise, mature independent ARC reproduction and live-account acquisition remain open; the separate 1e100.net signature is unverified because its current selector is empty or revoked.

The five decoded attachments

The aggregate includes duplicated draft material and a 154,318-word human-AI working dialogue. The accurate description is three manuscript and research drafts, one book draft and one preserved human-AI dialogue. Not four authored manuscripts, and not 221,236 unique human-authored words.

The email body and attachments contain: ARC, Artificial Recursive Creation; U = I × R in Version 2; the name Eden Protocol and alignment “from inception”; a broad control-fragility concern; recursive self-improvement; cross-faith stewardship; and the named HRIH recursive-cosmogenesis hypothesis.

The most categorical control-failure sentences quoted on this site occur in an assistant response within the HRIH dialogue. They are attributed to the preserved Eastwood-directed human-AI working dialogue, not presented as Eastwood’s unaided manuscript prose.

Why these records were self-emailed at all. I record and hash material relevant to this thesis as routine practice, because broadcast archives and provider mailboxes expire on their own schedules; hashing a private copy is what lets a third party check it later without anyone republishing copyright material. That is why these were sent to myself: to fix a date on a file, not to make a claim about anyone.

April manuscript record

Ideas grow. By the following April the sketch had become architecture: a 562-page manuscript in which care isn’t bolted onto the machine but built into its load-bearing walls, so that tearing it out should bring the roof down. That’s the proposal. Here is its record.

This source develops the stronger load-bearing Eden architecture: Caretaker Doping, Meltdown Triggers and Alignment, Quantum Ethical Gates, Moral Genome Tokens, Eden Looping, Eden Mark, Orchard Caretaker Vow and Infinite Covenant. It proposes that care should be structurally persistent and that attempted removal or bypass should impair operation. These mechanisms were not found in the December bundle and are not backdated to 8 December.

Reproducible raw text-extraction counts: Eden 552 whole-word occurrences; Babylon 90 whole-word occurrences and 106 substring occurrences including Babylonian and Babylonic; Eden Protocol 193; ARC Principle 94; Caretaker Doping 265; Meltdown Alignment 8. The extraction command, Poppler version and case and word-boundary rules are published beside these figures on the reproducibility route. 106 is never described as the whole-word Babylon count.

Public book anchors

A private email fixes a date only for the person holding it. A published book fixes one for everybody. On 2 January 2026 the ideas stopped being mine alone.

Infinite Architects: Intelligence, Recursion, and the Creation of Everything
Paperback: 2 January 2026 · ISBN 9781806056200 · 468 pages
Kindle: 6 January 2026 · ASIN B0GDQZ7P21

The print publication is an independent public anchor for text demonstrably present in that edition. It precedes Eastwood’s later formal HRIH and Paper IV.d publications. It does not predate Alignment faking in large language models (18 December 2024), and a later journal issue does not outrank an earlier external preprint.

The chronology against Alignment faking in large language models, exact safe wording

Ten days. That’s the whole interval, and I’d rather you measured it yourself than took my framing of it. Papers aren’t written in ten days, which is exactly why I claim convergence and not prophecy. Here is the wording I hold myself to.

The Eastwood source record bears a date ten days before Ryan Greenblatt and colleagues posted Alignment faking in large language models (arXiv, 18 December 2024); the timing is convergent, and he claims no more for it. A recipient-form export contains the same five attachments; its sender-domain DKIM and paired Google ARC set verify against keys currently published at their signed selectors. Historical key-state corroboration and independent mailbox custody remain open. The Eastwood material contains a broad concern that sufficiently capable systems may circumvent or perform compliance with external safeguards and argues for formative correction. Greenblatt and colleagues demonstrated alignment-faking reasoning in a specific constructed training conflict, including a reported 78% rate after reinforcement learning in the paper’s stated condition. His research asks what would keep such systems good if control were lost, while holding that humans should keep control for as long as they can.

The external study is consistent with one concern in the source record. It did not test ARC/Eden, does not establish that every external control fails and does not validate the Eden Protocol.

Source: Greenblatt et al., Alignment Faking in Large Language Models. Earlier relevant literature includes Risks from Learned Optimization, Sleeper Agents and the OpenAI o1 system card.

Hrih, one hypothesis, six evidential levels

This is the boldest idea I’ve ever written down, and bold ideas deserve the strictest cages. So HRIH lives on a ladder. Each rung says exactly how much you’re entitled to believe, and the hypothesis only climbs when the evidence does. Right now it stands on the bottom rungs, and I’d rather tell you that plainly than have you discover it resentfully.

Authorship statement. HRIH’s foundational source is an Eastwood-directed, AI-assisted research dialogue containing Eastwood’s originating prompts and subsequent model-assisted elaboration. The dialogue contains the post-ASI, non-ordinary-causality, self-ancestor-universe and ARC/Eden elements. Eastwood directs and adopts the named HRIH programme; sentence-level contribution remains provisional pending a role-preserving export or verified attribution map. HRIH is presently an unvalidated speculative cosmogenesis hypothesis and developing research programme.

Prior work that appears before any originality claim: Crane, Gott and Li, James N. Gardner’s 2003 Biocosm (Wiley), Vidal, Tipler and Bostrom.

The claim is the exact named ARC/Eden/HRIH conjunction, not origination of intelligent cosmogenesis, self-creating universes, causal loops, future godlike intelligence, simulation arguments or AI-created universes.

Arc naming and disambiguation

In this programme, ARC means Artificial Recursive Creation. The arc and loop-back metaphor is intentional. On first mention: Eastwood ARC (Artificial Recursive Creation); thereafter ARC/Eden. The candidate benchmark is Eastwood ARC-Align. Eastwood ARC and Eastwood ARC-Align are unrelated to ARC-AGI, the AI2 Reasoning Challenge and the Alignment Research Center. No affiliation, endorsement or exclusive ownership of the word or acronym ARC is claimed. Trade-mark clearance remains a separate professional legal task.

Later developments, graded before they are counted

I used to publish a number here. I took it down myself, and here is why: some of what I’d counted as the world converging on this work turned out, on honest inspection, to be the world getting there first. A register that can’t tell those apart isn’t evidence. It’s flattery with a denominator.

No numerical convergence total is published. The former register was assembled retrospectively without a pre-outcome scoring rule, fixed horizon, base-rate model or complete failure register. Technical experiments, theorems, engineering proposals, policy developments, cultural echoes, prior work and author-run studies cannot legitimately be summed as one evidential quantity.

Every external row receives one primary class and states the exact external source and earliest public date, the earlier Eastwood proposition, the precise structural match, the important difference, prior work relevant to both, dependence, current grade and reviewer, and correction history. The full graded register lives at /research/convergences/.

One worked example of the dependence rule: the register’s formal-limits chain. Gumbau Mezquita’s Undecidability of AGI Alignment (arXiv 2606.28639, 26 June 2026) states that its conceptual foundation is Hernández-Espinosa et al. (arXiv 2505.02581), which concerns agentic-AI diversity. They are one dependent research chain and are counted once, never as two independent confirmations.

Independent work belongs to its authors. A later event can show that a question matters without establishing Eastwood’s priority, a successful prediction or the truth of Eastwood’s answer.

Programme-run results

What follows is everything my own experiments produced: the result I’d defend, the pilots that only hint, and the findings that went against me. They’re presented together because that’s the only honest way to present them. A body of work that contains no failures is not a body of work.

Paper iv.d, public methods result

In Eastwood’s Paper IV.d setting, moving from v4 to a bundled multi-layer protocol changed the aggregate alignment-depth outcome. Gemini’s Spearman association moved from positive (rho = +0.275, p = 0.0001, an earlier Gemini model in v4) to negative (rho = -0.246, p = 0.00340, Gemini 3 Flash in v5); its grouped exponent was negative but non-significant (-0.309, p = 0.105). DeepSeek’s earlier positive association became null (rho = -0.135, p = 0.0805). GPT remained null. Several protocol elements changed together, so the study does not isolate blinding as the cause. Code and result artefacts are public. Exact-environment reproduction and independent replication remain outstanding.

An internal code review in October 2026 found that, for the three models that read flat or negative in the March 2026 run, 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 early build's laundering-fallback flag is a known false positive (fixed in runner version 5.4.2) and shows nothing about the paraphrase step.

Gemini’s p = 0.00340 is below a Bonferroni threshold of 0.05 / 6 = 0.0083 only if six model-level Spearman tests are the prespecified family. Because several protocols and metrics were examined without that family being frozen prospectively, this page does not claim multiplicity-adjusted confirmation.

Sources: Paper IV.d OSF, public experiment suite, canonical table. Required prior work: Zheng et al., CoBBLEr, Panickssery et al., Safer or Luckier? and The Silent Judge. Response-laundering lineage: Brennan, Afroz and Greenstadt (2012) and Krishna et al. (NeurIPS 2023). These are lineage, not validation of Eastwood’s implementation.

Eastwood arc-align, candidate benchmark and release status

Eastwood ARC-Align is a candidate benchmark for examining how measured alignment responses change with inference-time reasoning depth under specified blinded, cross-model evaluation protocols. Its historical v1 to v5 scripts, v6 runner and result artefacts are public. Exact historical environment reproduction, construct validation and independent cross-laboratory replication remain incomplete. Scores describe behaviour under a named benchmark version; they do not measure intrinsic moral character or establish a universal safety ranking.

Release status: current, code and historical result artefacts public. Next, an Eastwood ARC-Align v0.x candidate replication release. Not claimed: validated standard, contamination-proof benchmark, intrinsic-alignment measure or independently replicated result.

Paper v, prompt-responsiveness pilot

In an author-run March 2026 pilot, an explicit instruction to identify affected people increased AI-judge stakeholder-care scores in five analysable model runs. One model run failed; scoring was nonblind and single-scorer. The result demonstrates prompt responsiveness under that protocol, not embedded values, persistence or validation of a “Stewardship Gene”.

The Fisher-combined p-value is not headlined: independence among component tests is not established.

Paper x, conditional theorem and separate mechanism pilot

In Eastwood’s specified model, correction proportional to x^beta dominates modelled drift proportional to x^k asymptotically when beta > k. Separately, Eastwood reports a 45-trajectory author-run mechanism pilot in which conditions differed under a constructed drift protocol. The pilot did not estimate beta or k; its GPT-3.5 engine was evaluated by another OpenAI-family model rather than a cross-family judge; its coded H1 rule differs from the stated positive-slope hypothesis; and it is not an empirical validation of the inequality.

Confirmatory-run status: not ready to submit or run. The inspected v3 harness was unsafe and protocol-mismatched. A synthetic-only rc1 repairs prompt orthogonality, sandbox isolation, fail-closed panels, carried-state identity, HMAC block randomisation, hidden-bank input disjointness, exponent and bootstrap-validity gates, replicate-cluster bootstraps, schedule-aware missingness and the held-out gate; 35 of 35 non-network tests and its synthetic fixture pass. It deliberately has no real provider adapter or collection command. Exact endpoints, evaluator controls, calibrated fresh banks, encrypted commitment, budget, outer image digest, analyst custody and the final freeze remain unresolved. A prospective registration must disclose and exclude prior pilots and govern a new dataset.

Dgm v4, internal pilot, excluded from external evidence

Eastwood reports a nine-run, single-model exploratory pilot in which the Eden condition had the highest reported mean entangled-fitness score and fewer reward-hacking events than Babylon under the author-defined protocol. The study is preliminary, single-lab and unreplicated. It is excluded from every external-evidence count.

The Darwin Godel Machine is prior work: Zhang et al. (2025).

Where the programme’s own results went against it

The programme publishes adverse and limiting results beside positive ones. The early alpha approximately 2.24 estimate is retracted. The later estimate near 0.49 has wide uncertainty. In Paper IV.d, the grouped Gemini exponent is non-significant, DeepSeek is null under the recomputed v5 Spearman test and GPT remains null. Paper V lost one complete model run and uses nonblind, single-scorer evaluation. Paper X does not estimate beta or k. DGM v4 is a nine-run, one-model pilot without independent replication. These results are part of the evidence record, not footnotes to be hidden.

What would change the status. A frozen and independently replicated result can upgrade the relevant claim. Failure under a frozen rerun, a better-supported alternative explanation or materially earlier prior work can downgrade it. The status changes; the previous record remains visible.

Pnp, proposed descriptive framework

I can draft a skeleton argument for the Court of Appeal. I struggle to open routine post. The framework below is my attempt to describe that architecture honestly, as a hypothesis about minds like mine rather than a certificate about my own.

The Polymathic Neurodivergent Profile is Eastwood’s proposed, falsifiable and presently unvalidated descriptive framework. It is not a diagnosis, recognised syndrome or psychometric category. Eastwood reports adult diagnoses of autism and ADHD. Those diagnoses do not clinically certify all six proposed PNP components. Cross-domain learning, recursive pattern transfer, autonomous synthesis, monotropic attention and institutional misrecognition require separate operationalisation and testing.

Its practical proposition is narrower and well supported: visible capability does not, by itself, establish absence of disability or remove the need to assess reasonable adjustments by their functional effects. The PNP offers a testable account of the cognitive profile associated with this work. Whether it explains the work is the hypothesis under examination. The loop is open by design.

Private model use and later naming echoes

A preserved chat record may establish that a user entered or received specified material on a date if the provider record is authenticated. It does not, without further evidence, establish that the material trained a shared model, entered model weights, was memorised, leaked to another user or caused a later output. A later Gemini output using “Eden Protocol” in a thematically related context may be recorded as a dated model-output echo. The causal route is unknown; no claim of training, memorisation, disclosure or derivation is made.

Verify the record yourself

The point is not to ask you to trust me. The point is to make trust unnecessary. Ten steps, a terminal, and an afternoon.

1. Obtain the December manuscripts. The four are published on OSF (DOI 10.17605/OSF.IO/QASRK); the headers of the email of 8 December 2024 are published, and the emails themselves are held privately for now.
2. Run shasum -a 256 <file> on each and compare it with its attachment hash in the manifest table above; the December email's own digest is f0d1f38ffd8546152d9d9d28dc5ec083c16a35858f2c12b63e69db7ed50901ad.
3. Obtain the April manuscript, its print PDF, from the same OSF component and compare its hash with the manifest table; the April email's own digest is 09f5b5e156ed96f8883eaf668495fd350898ce62be6294b8f788e0e2d6dcb664.
4. Where an email itself is held, parse its MIME structure and compare its attachment hashes with the manifest table above.
5. Reproduce the word and term counts using the published extraction rules.
6. Use the passage map to locate every quotation and identify whether it is manuscript prose, model-generated dialogue or jointly developed material.
7. Inspect the sent and recipient headers separately. Confirm the sent-side absence and recipient-side presence of transport and authentication fields; run the published sender-DKIM and Google-ARC verifiers and preserve their scope controls, key fingerprints and current-DNS captures separately from historical-key and custody claims.
8. Compare source-visible dates, public book dates, repository history, OSF deposit dates and external-work dates without treating them as equivalent evidence.
9. Recompute Paper IV.d metrics from the public results and document any discrepancy with the paper.
10. Report a failed hash, count, passage, date or statistical result at /research/corrections/.

In-browser check: verify byte identity. A match means only that the selected bytes match the published digest.

What is not claimed

This may be the most useful section on the page. Every sentence below is a claim I could have made, and chose not to, because the evidence doesn’t reach it. If you catch this site making any of them, that is a defect: report it and it will be corrected in public.

• Not that every supplied manuscript file carries its own timestamp from an independent timestamping authority. The December record’s time is signed by the sending server and sealed by Google’s servers, and both validate against current selector keys: that is the record’s time evidence, at its true size. Recipient-side copies have since been acquired, carrying the delivery chain and a non-expiring Google ARC seal whose signing time is inside the signed scope; the sender-domain DKIM carries a seven-day expiry and is corroboration rather than the vehicle. Supplied raw-source runs report the April pair PASS; this workspace has not reproduced them without the raw source and executable control harness.
• Not that Eastwood originated embedded alignment, recursive self-improvement, hardware AI governance, interfaith AI ethics or intelligent cosmogenesis.
• Not that Anthropic tested or confirmed ARC/Eden.
• Not that later external work derived from Eastwood.
• Not that the book preceded every external paper; comparisons use earliest public disclosure.
• Not that HRIH has been empirically confirmed.
• Not that PNP is clinically validated.
• Not that Paper IV.d isolates blinding or has been independently replicated.
• Not that Paper V demonstrates embedded values.
• Not that DGM v4 validates Eden.
• Not that private AI use proves training-data leakage.

Exact bytes public. Historical dates graded. Prior work credited. Results open to replication. Claims open to correction.