Claim → source → check → status → correction
Research evidence spine
Every claim on this site is supposed to end somewhere you can stand. A file. A hash. A date. A page you can open in another tab and read for yourself.
The spine is the map of where each one ends. Follow any claim down and you arrive at an artefact or you arrive at nothing, and if it is nothing I would rather you found that out here than somewhere less convenient for me.
A note before you start, because it changes how the rest reads: most components here have prior work, my book conceded that before any review existed, and what a hostile review could not find was the arrangement. The eight joins, and what each one concedes.
declaration · 12 min read
The Evidence Spine is the programme’s source of truth. No claim may appear in a paper, grant, public page, structured-data block or AI-ingestion file unless it resolves to an authorised row here. Every row identifies the proposition, source bytes, date type, independently public date, prior work, evidence grade, authorised wording, prohibited wording, last review and correction history. The purpose is not to make every claim sound stronger. It is to ensure that no claim sounds stronger anywhere than its evidence permits.
Source-edition and verification model
Every source with sensitive, mixed-authorship or historically disputed material uses five linked objects:
SO, source original. The untouched source bytes and whole-file and MIME-part hashes. The December and April source originals were deposited to OSF on 4 July 2026 and relocated to a private access-on-request component on 31 July 2026 for privacy; the published hashes remain public and controlling. Provider-held originals and enhanced transport records may be handled through a documented review process.
PE, planned public extract. A future context-preserving research edition containing the load-bearing passages and adjacent context. It will have its own hash and will never replace SO.
AM, planned attribution map. A future passage-level record identifying Eastwood-authored prose, Eastwood prompts, model responses, quotations and later editorial synthesis.
RM, planned redaction manifest. Every omitted range, reason and transformation, without exposing the removed private information.
VR, planned verification report. A future independent signed report confirming that PE accurately represents SO, its cited passages exist and the chronology is described no more strongly than the evidence permits.
The target chain is: claim → PE passage → AM role → SO byte range and hash → date evidence → prior work → status → correction history. Until PE, AM, RM and VR exist, claims must link directly to the public SO and clearly state the attribution limitation. Neither privacy filtering nor a review process may be represented as third-party authentication.
Evidence rungs
No current ARC/Eden/HRIH claim should be promoted to R5 or R6 without the qualifying external work in the ledger.
Anchor cards
December source
- Exact current bytes
- verified
- Visible date
- 8 December 2024, 02:45:18 UTC
- Recipient transport
- sender-domain DKIM and paired Google ARC validate against keys captured at their signed selectors; historical key-state and independent live-account acquisition pending
- Independent public deposit
- 4 July 2026
- Sent-side SHA-256
f0d1f38ffd8546152d9d9d28dc5ec083c16a35858f2c12b63e69db7ed50901ad- Recipient-side SHA-256
ecdc8999e5326235ad9a12810e864eefb6f305c04b15568c0c86a1a3a6a53b39- Key coherence
- the April 2025 record verifies under the same sender-domain signing key, sixteen months apart on one continuously configured domain; December’s signing timestamp t=1733625930 (02:45:30 UTC) sits twelve seconds after the visible Date header
- Contents
- five attachments; 221,236 aggregate whitespace-delimited words
The public sent-side export has no transport fields. The separate recipient-form export contains Google trace and authentication fields and a paired Google ARC set. Every attachment hash, the complete signed-body bh= values, sender-domain RSA signature and Google ARC pair reproduce against keys captured at the signed selectors. Recipient-side copies have since been acquired and both signing selectors still resolve in DNS, so the seal is checkable today. The ARC set bearing d=google.com has no inspected x= expiry and is the preferred provider-key route; the sender-domain DKIM, which does carry a seven-day expiry, is corroboration rather than the vehicle. Sender-signature expiry, historical key-state and live-account custody remain pending. Quote the sealed ARC-Authentication-Results, never the unsealed Authentication-Results, which a file can be edited to say anything.. Do not call it RFC 3161 timestamped, historically closed or independently acquired from Gmail.
The recipient source supplements rather than replaces the sent-side source. An attachment crosswalk explains the different whole-file hashes and identical decoded attachment hashes. The exact mailbox address belongs in the examiner protocol, not page copy.
April source
- Exact current bytes
- verified
- Visible date
- 30 April 2025, 13:37:05 BST
- Transport-authentication status
- sender-domain DKIM body hash and RSA signature pass, and Google ARC AMS and Seal pass, in the supplied raw-source verification, with five negative controls failing exactly where they should: flip one byte in either attachment and the body hash fails; alter the Date or Subject and the signature fails. Local independent reproduction is pending because the raw recipient source and control harness have not yet been supplied to an audit environment. The signing key returned by four independent resolver vantages on 2 August 2026 matches the verifying fingerprint; the key state in force on the message dates is a separate, open question. Noted and carried: 1024-bit key, seven-day x= expiry (which governs live acceptance, not archival verification), and examiner acquisition of the mailbox copy by Gmail API, Show Original or Takeout remains open
- Recipient-side SHA-256
d52730f3b8246dffd58fdaa0f02d22d7453ea38ab6b5e9c6c78a65687ba30834(33,275,268 bytes)- Attachment 2
- Michael Eastwood - Signature.png, 45,205 bytes, SHA-256 beginning 95508854 and ending 338a609 (full value in the verification manifest). A signature image, not a manuscript; listed so the attachment manifest is complete.
- Independent public deposit
- 4 July 2026; relocated to a private access-on-request component on 31 July 2026 for privacy; the published hashes remain public and controlling
- Sent-side email SHA-256
09f5b5e156ed96f8883eaf668495fd350898ce62be6294b8f788e0e2d6dcb664- PDF SHA-256
735a975cca2c6ac803ea54b901e964f6ac996a093a29c4708e2668e942c7f148
Public book
- 2 January 2026 · ISBN
9781806056200 - Digital
- 6 January 2026 · ASIN
B0GDQZ7P21 - Scope
- public priority only for text demonstrably present in the relevant edition
Expert submission, 24 March 2026
- Sent
- 24 March 2026 · Eastwood to Professor Lloyd Demetrius (Harvard, organismic and evolutionary biology) and Professor Geoffrey West (Santa Fe Institute)
- Content
- a formal letter on Cauchy’s 1821 functional equations as the connection between their two independent derivations of α = d/(d+1), with two papers attached: On the Origin of Scaling Laws (v3.1) and Paper VII, Cauchy Unification (v3.0)
- Letter SHA-256
9085fdc907405bb15b485c0fb0c2e214d4d00614a34b6780396ef0879f64baef- Scope
- a dated third-party disclosure to the field’s leading metabolic-scaling theorists, verifiable by contacting the recipients; it establishes submission for expert scrutiny on that date, not endorsement, and no reply is claimed
What can be reproduced now
- both manuscript-email SHA-256 values;
- MIME structure and all December attachment hashes;
- aggregate word counts and April term counts;
- source passage locations and attribution class;
- OSF public-deposit chronology;
- current bibliographic print and digital dates;
- Paper IV.d statistics from published result artefacts;
- inspection of the public v1 to v5 runners, v6 runner and historical result artefacts.
What remains incomplete
- historical key-state and live-account or API acquisition for the December recipient record; mature independent RSA reproduction; supply of the April recipient record and executable harness, and local reproduction of its reported-valid ARC pair and controls;
- a documented end-to-end clean run of the public Paper IV code against current APIs; such a run cannot recreate historical provider behaviour without preserved snapshots;
- exact historical model and provider snapshots and a fully pinned Paper IV environment;
- the planned PE, AM and RM mappings for the context-preserving HRIH research edition;
- an independent VR confirming the HRIH public extract against the source original;
- independent replication of Papers IV.d, V, X, DGM or the Cauchy and scaling work;
- a complete final-edition passage map for every January-only term;
- independent clinical testing of PNP;
- independent Eden Legal AI performance evaluation;
- provider-specific evidence for any private-chat training or leakage claim.
The public scripts correct the stale statement that the Paper IV harness is absent. Public code does not by itself reproduce historical API behaviour, so the accurate status is: harness public and merged since 31 July 2026; results public; exact-environment and external reproduction outstanding.
Append-only correction record
Corrections are never overwritten. A superseded value remains visible with its date, reason and replacement. The full log lives at /research/corrections/; the headline entries:
- four attachments → five attachments;
- “four authored manuscripts” → three manuscript and research drafts, one book draft and one human-AI dialogue;
- “Google-server-timestamped” → December recipient-form record with matching attachments and sender-DKIM body-hash and RSA plus Google-ARC AMS and Seal mathematics passing against current selector keys; sender-signature expiry, historical key state and live-account custody pending;
- “106 Babylon whole words” → 106 substring occurrences, 90 whole words;
- HRIH
datePublished: 2024-12-08→ actual public paper-release date; private-source date recorded separately; alphaapproximately 2.24 → retracted; later estimate and full uncertainty reported separately;- “two of six reversed sign” → one significant opposite-direction Spearman result, one loss to null, GPT null;
- “blinding caused reversal” → bundled protocol change; causal component unisolated;
- “22 independent convergences” → typed, graded register with no heterogeneous total;
- “originator of embedded correction” → author of specified names, formulations and composite synthesis;
- “first validation” → preliminary internal pilot;
- “loop closed” → hypothesis open to testing and refutation;
- “harness unavailable” → harness merged and public 31 July 2026; exact environment and independent reproduction outstanding.
Canonical DOI register: every paper individually citable (22 public components + programme)
Paper III: cite its own DOI 10.17605/OSF.IO/HQCGF, verified live and public against the OSF API on 1 August 2026.
The Eden Protocol, named in full ("Chapter 20: The Eden Protocol", 31 uses). The governing equation, verbatim: "U = IR (Universe = Intelligence × Recursion)", 43 occurrences. The control-failure thesis, in writing ten days before Anthropic's alignment-faking result. And a complete creation theory, the Hyperspace Recursive Intelligence Hypothesis, 1,103,066 bytes in the manifest. One night, one hash: f0d1f38f. The eighteen months since kept walking into it: the error-suppression principle the day before Willow (disclosed, not claimed), the faith-leaders call ten months before the Vatican arc, the external-verification insufficiency thesis eighteen months before a formal preprint argued the same limit from undecidability. The complete priority ledger →
Run it yourself
Nothing on this page asks to be taken on trust. Every claim above resolves to a row, and the rows below resolve to a command you can run, each indexed with its commands in the repository’s experiments index. If a step fails, the claim fails. That is the point.
1. Check the 8 December 2024 anchor
This is the manuscript that carries the embedded-correction thesis, sent by the author to himself, its origination fields assigned by Google's servers at send time. The source file is available on request, hash-anchored; its SHA-256 was published on day one, and the two BBC anchors are downloadable directly from the evidence portal with both served and original digests. On any provided copy, hash it and read its headers.
The 8 December 2024 source object is withheld for privacy. Its SHA-256 is public and controlling, and the file is supplied on request for verification.
shasum -a 256 dec8-2024-manuscript.eml
# expect: f0d1f38ffd8546152d9d9d28dc5ec083c16a35858f2c12b63e69db7ed50901adIf the digest matches, the file has not changed since the day that hash was first published. Then open the same file in any text editor and read three headers. Date gives the timestamp. Message-ID begins CAGPsKA, the Gmail-origin pattern, assigned by Google's servers at the moment of sending. From and To are the same address, because the message was sent to himself. Stated plainly: two exports exist and they prove different things. The sent-side export carries no transport fields; hash it against f0d1f38f to confirm the bytes. The recipient-form export carries the transport record; hash it against ecdc8999 and read three things: the Date and Message-ID assigned by Google’s servers, the sender-domain DKIM-Signature that verifies under the domain’s published key, and the ARC-Authentication-Results header, which is the sealed one. Quote that header, never the unsealed Authentication-Results, which carries no integrity guarantee. Third-party attestation also exists where a message was delivered to someone else, as with the 7 January BBC anchor, whose authenticated copy sits on the BBC's own mail servers.
What you need to run them: a laptop and API keys for the model families under test: these are API experiments, not GPU cluster jobs; individual runs complete in hours of wall-clock. Exact model lists, versions and scripts per experiment are in EXPERIMENTS-INDEX.md in the repository. For funders and assessors: start with Paper IV.d (DOI 10.17605/OSF.IO/2S3E6), the programme's most defensible result: its experiment scripts are in the repository and the statistics can be recomputed from the published outputs today.
A note on what you will see. The message is addressed from the author to himself, so both the sender and recipient headers carry his own addresses, one of them an old business address he no longer uses. Those headers are not incidental: self-addressing is what makes this a self-timestamp rather than correspondence, so they cannot be removed without destroying the thing being proved.
2. Clone the repository
The papers, the experiment code, the result files and the priority material are all in the public repository, where anyone can read them. This page states no licence: the terms are in the repository’s LICENCE file.
git clone https://github.com/MichaelDariusEastwood/arc-principle-validation
cd arc-principle-validation Start with PRIORITY-VERIFY.md, which walks every priority claim through the check that would break it, and EXPERIMENTS-INDEX.md, which maps each paper to the experiment that produced it. Browse the repository.
3. Re-run the findings, strongest first
The blinding sign-reversal (Paper IV.d). The per-model result files under papers/Paper-IV-d-The-Effect-of-Blinding-on-AI-Alignment-Evaluation/results/v5-final/ hold the scored runs. Recompute the Spearman correlation on consensus_weighted_mean per model. You should find one model shows a significant opposite-direction result once the scorer is blinded, a second loses its significance to null, and GPT is null throughout. If you find anything stronger, the page overstated and I want to know. This is the most defensible result in the programme, and it is the one most worth attacking.
The stakeholder-care result (Paper V). The matched-pair result files give the care difference between the intervention and control conditions, Fisher-combined.
The theorems (Paper X). Paper X is self-contained, with its own Makefile, Dockerfile, tests and SHA-256 manifest. Its test suite re-derives the theorems from scratch and shares no code with the harness that produced the empirical results, so a single common bug cannot hide in both.
One safety note: the historical v1 to v5 runners execute model-generated code in-process. Inspect them, but do not run them on a normal workstation; the sandboxed rebuild is the version to execute when it ships.
4. What you cannot yet reproduce, stated plainly
The blinding harness for the Paper IV series is public. It merged to the default branch as eden_blinding_kit plus the v5 runner (experiments/alignment-scaling__Papers-IV-a-b-c-d/scripts/arc_alignment_scaling_v5.py) with a SHA-256 release manifest, so it can now be cloned and inspected end to end. That harness was the single most requested reproducibility artefact: declared as an open gap in June, released on 31 July. What remains outstanding is narrower and unavoidable: a clean run cannot recreate the historical provider behaviour the original results were measured against, because those model snapshots are not preserved. Saying so here is cheaper than letting someone discover it themselves.