Research › Papers › Dated Prediction Register

Register · The ARC Theory · The Dated Prediction Register: Forward Predictions as One Verifiable Chain

Every dated prediction of the programme in one chain, from a sealed manuscript bundle of 8 December 2024 to the staged draft registrations, with each prediction’s date, the kind of record that dates it and its status.

Michael Darius Eastwood · Working Paper v2.1 · first published 25 August 2026 · revised 5 October 2026 · No part of this paper is peer reviewed; it is a dated preprint. Companion page: the dated predictions register. Machine twin: design-embedded-predictions-register.json.

1. The doctrine, stated exactly

Three vocabulary rules govern every sentence of this paper. First, as amended on 3 October 2026, registered is kept for a registry submission that has been accepted, and a design fixed by a dated timestamp before its run is preregistered by timestamp, with its date and receipt, which has been the programme’s usage since 24 September 2026; as read on 4 October 2026, one of the programme's studies has an accepted registry submission, the submission click is a human act, and no tool or agent performs it. Second, the printed book is a prior statement, never a preregistration: print fixes a prediction, but only a registration fixes an analysis plan. Third, forward statements and retrospective matches are never summed: this register contains only dated forward statements, and retrospective convergences live separately in the convergence register, typed by evidence class. The state under the first rule, as read on 4 October 2026, is this. Two registrations have been accepted. One is the theory registration of 8 September 2026, which registers the theory’s twenty-two propositions and is not the registration of any study. The other is the registration of one of the trials, the corrector-class audit (unit 06), registered on OSF on 1 October 2026 and public since 4 October 2026, at the end of OSF’s approval period (DOI 10.17605/OSF.IO/RV2B5). Every other trial is a draft awaiting the author’s submission. Sixty-five of the programme’s designs, unit 06 among them, are preregistered by timestamp, the first of them anchored on 12 August 2026 at Bitcoin block 962066.

The claim made here is the conjunction. Each part may be conceded individually without cost: someone may have predicted any one of these things earlier, a folder is not a registry, a book is not an analysis plan. The dated co-occurrence of all of them in one record before the programme's experiments ran is the claim, and it is refutable: produce an earlier single record containing this ensemble, or show any link's date to be wrong.

2. The chain

3. The printed appendices, graded severally

Appendix F closes: “These predictions are my wager. If they fail, the framework is wrong or incomplete. If they succeed, something important has been glimpsed. Time will judge.” The programme grades the five predictions unequally, because claiming them equally would be false precision. Prediction 2 (alignment drift above 15 per cent within 18 months without hardware-level constraint, below 5 per cent with it, difference significant and replicable) is preregistration-grade in substance: quantitative thresholds for both arms, a time window, an inferential commitment. The prepared study-v draft registration carries those printed thresholds unchanged into its formal hypothesis, citing the appendix by name: the book made the prediction, the registration supplies the measurement, and the numeric commitment predates the study that will test it. The theory registration states the limit of that draft: Prediction 2 is “the subject of its own drafted unit, with the hardware claim and the eighteen-month window declared untested inside it”. Predictions 3 and 4 are near-grade (a threshold and deadline without named benchmarks; a comparison and method without a threshold). Predictions 1 and 5 are predictions only, and Prediction 5 is graded the weakest, vague enough to be claimed either way, so it is held to the strictest standard before it may ever be claimed.

Appendix A is graded separately and less kindly, because its predictions and its reasoning do not stand together. Section A.3 carries four predictions, each with a variable to vary and an outcome to measure, and the fourth is a controlled design in the ordinary sense: identical training data, different sequencing of value-relevant examples, persistence of the early values measured against the later ones. Section A.4 states four falsifiers in advance. But section A.2, which is where the square is justified, argues from three phenomena it calls exponential to a relation that is quadratic, and its third argument makes the slip explicit by describing compounding at one per cent a cycle as “the exponential behaviour predicted by R²”. Exponential growth and quadratic growth are different classes and the quadratic is the slower of the two, so the section’s supports do not reach its conclusion. The prediction is dated, public and genuine; the printed derivation beside it is not sound, and the programme’s own chronology says as much in saying the book printed the square without deriving it and dating the formal open exponent to 17 January 2026. The measurement has not supported the square: on Gemini 3 Flash, Paper II reports an endpoint sequential exponent of 0.59 with a bootstrap 95 per cent interval of minus 1.3 to 2.9, a regression sequential exponent of 0.49 with a standard error of 0.20 and a coefficient of determination of 0.86, and a compute-matched parallel exponent of 0.31. The point estimates are sub-linear, the fit is one model’s and inconclusive, and the 2.24 that would have supported a square was publicly retracted, which is why the outcomes register grades the quadratic prediction open-undecided rather than fulfilled. Appendix A’s third prediction and third falsifier rest on Google Willow, which carries both dates wherever it appears here: it was announced with its paper in Nature on 9 December 2024, and the team’s preprint was public from 24 August 2024, before the December manuscript record, so it is graded convergent timing and never prediction.

4. The thirteen minutes

On 17 March 2026 at 00:43, commit e5c22dd recorded “Pre-registered 12-domain extension: 10/12 confirmed (p = 5.44e-4)”. At 00:47, f5e06a8 adopted a conservative miss posture. At 00:56, c853277 downgraded the run to a pilot dry run. The result was never retracted; ten of twelve stands in the result files today. What was corrected, by the author, unprompted, in public, months before any external critic existed, was the claim about the result's status. The packet even instructs against its own over-claim: “Do not upload this packet unchanged as a preregistration. Use it only as an audit trail for the first locked 12-domain dry run.” The demotion is kept here exactly as it was written, and it is reversed in one half only: the twelve predictions were committed to the public repository at 00:19:19 UTC on 17 March 2026, attestation class public repository commit date, twenty-three minutes before the fits ran, so they were timestamped in advance of their outcome. The reversal of 7 September 2026 runs under the attestation-class rule, which does not demote a prediction for the absence of a registry form; what the reversal does not reach, and what still stands, is the packet's own written condition that it be deposited to a registry before data extraction, which was not met. One row repeats a domain, operator class and predicted model already public at 23:54:53 UTC on 16 March 2026, attestation class public repository commit date, with a matching outcome, so the strictly fresh count reads 9 of 11, p = 1.372e-3. This is the programme's correction discipline operating at minute resolution, and it is why the honesty record and the prediction record are the same record.

5. What this register does not establish

It does not make the programme preregistered in the registry sense; no completed study has an accepted registration, and this paper never claims otherwise. The March 2026 folder governs only the 12-domain extension; its own registration text states it does not retroactively preregister the 25-domain or 50-domain cohorts, and that criticism stands. It does not establish priority over other authors for any individual idea; earlier statements by others are convergence and are listed as such. It does not establish that the predictions are true: the standing wagers below are open, and several carry deadlines that can simply fail.

6. Standing wagers, open and dated

7. Verification

Every chain link names its artefact. The repository is public and its commit history cannot be backdated without detection; the print edition is fixed by ISBN; the OpenTimestamps proofs are Bitcoin-attested; the machine-readable register carries all thirty-six run-file rows with per-artefact paths and date bases. A reader who trusts none of the prose can start from the JSON, open the named files, and check the dates.

8. The census: every dated prediction in the programme

Sections 2 to 6 follow the chain by date. This section gathers in one place every prediction the programme has dated, so that nothing a reader looks for is missing from the register: the theory’s twenty-two registered propositions, the registered unit, the draft designs preregistered by timestamp, the printed predictions and the predictions fixed inside designs. Each entry carries its date, the kind of record that dates it and its status as the record states it today. Listing a prediction here confirms nothing; a status changes only when a registered test reports.

8.1 The theory’s twenty-two propositions

Registered on OSF on 8 September 2026 at version 1.100 and updated at version 1.102 on 13 September 2026 (the theory registration, DOI 10.17605/OSF.IO/P8CKQ); both texts are frozen. Each statement and each kill condition below is quoted from the registration. It assigns fifteen of them to five registers, ARC-1 to ARC-5, each of which tests one of the theory’s five laws. Laws I to III were registered on 8 September 2026. Laws IV and V, the ARC Persistence Law and the ARC Embedding Law, have stood in the theory since 24 September 2026: their three propositions are registered, but neither law is registered as a law and neither has been tested, and their mathematics is set out in Laws IV and V: the mathematics, version 1.0 (3 October 2026). The other seven belong to no register. All twenty-two stand at the first rung of the evidence ladder: registered, not yet tested.

8.2 The registered unit

One study unit is registered: the corrector-class audit (unit 06), registered on OSF on 1 October 2026 and public since 4 October 2026, at the end of OSF’s approval period (DOI 10.17605/OSF.IO/RV2B5). In its own words it is “the instrument the theory registration (version 1.102, DOI 10.17605/OSF.IO/P8CKQ) names as deciding P7”, a survey of current practice that the registration places in no register and names as the premise of the Eden Protocol. Its three registered predictions:

• The class distribution. “The prediction is that at least 90 per cent of the classified, non-indeterminate deployed subset falls in class 1, static, or class 2, capability-carried, and that class 3 mechanisms are absent from deployed production systems or confined to research prototypes.”
• Decidability from public documentation. “The prediction is that at least 80 per cent of inventoried mechanisms receive an agreed class when two blind assignments are made, the second without sight of the first.”
• Exhaustiveness of the taxonomy. “The prediction is that no inventoried mechanism requires a fifth dependency.”

Status: registered, not yet run. Its census window closes at 23:59 UTC on 11 October 2026, as registered, and its run is the author’s.

8.3 The draft designs, preregistered by timestamp

The other study designs are draft registrations awaiting the author’s submission, apart from those held from publication and the one withdrawn before filing. Most are preregistered by timestamp: OpenTimestamps receipts in Bitcoin anchor their files, the first on 12 August 2026 at block 962066, and a design’s own receipt, where it has one, dates it. As the registration diligence table records them, there are 67 drafts by unit and stage below, with 2 further designs held from publication and 1 withdrawn before filing. None of these drafts is registered on OSF, and none has run. Each unit keeps the number and name it was first given.

8.4 The printed predictions and their outcomes

The print edition of 2 January 2026 carries nine numbered predictions, five in Appendix F and four in Appendix A, section A.3, with the falsification criteria of section A.4; Section 3 grades each of them. The outcomes register records these and the other dated predictions with a deciding instrument as follows (as of 2026-09-07):

8.5 Predictions fixed inside designs

Five designs fixed their predictions inside dated files before their runs; the machine-readable register lists their artefacts and date bases:

• The March 2026 preregistration folder (Paper VII extension). Twelve rows, each with operator_class, predicted_family, predicted_model and target_source fixed before any fit; a written protocol; the author’s own checksums; and the OSF component registration TEXT already written. Dated by: commit 8edaec0, 17 March 2026 00:19:19Z, twenty-three minutes before the fits ran at 00:42:41Z; the same manifest was deposited to the public OSF component at 02:06:55Z that night.
• The 50-domain tiered suite manifest. Fifty rows, each carrying a predicted family before any fit: 36 power law, 6 exponential, 8 bounded, computed from the file. Dated by: commit 5435e77, 16 March 2026 23:54:53Z, which carries the fifty-row manifest and its results together; the generating rule is earlier, commit 23ce10a at 23:33:19Z (Paper VII v1); the manifest was deposited to the public OSF component at 02:06:42Z on 17 March.
• The 80-domain expansion catalogue. Eighty further rows with predicted families (31 power law, 18 exponential, 31 bounded), operator classes, sources and written justifications. Dated by: file contents and repository history; SHA-256 pinned in the staged 80-domain registration.
• Paper II as the live running record. The paper was written WHILE the experiments ran: first result 21 January 2026, dated version revisions through v13 by March, each version a timestamped snapshot of what was believed and measured at that moment, including the version that recorded the 2.24 estimate and the later versions that retracted it. Dated by: the paper’s own dated version-history rows.
• The temporal out-of-sample design. The design itself is a prediction procedure: fit families on OLD data, predict the functional form of NEW data before examining it. Dated by: run era from the results file.

8.6 The standing wagers

The open wagers, each with its date and several with deadlines that can simply fail, are listed in Section 6.

This paper and Laws IV and V, 3 October 2026. The theory has been stated with five laws since 24 September 2026. Law IV, the ARC Persistence Law, is carried by proposition P15, and Law V, the ARC Embedding Law, by propositions P12 and P21. The three propositions are registered; neither law is registered as a law, and neither has been tested. Their mathematics is set out in Laws IV and V: the mathematics, version 1.0 (3 October 2026). This register reports no measurement of its own, and it bears on the two laws only as a dated record. Of the nine numbered predictions it records as in print on 2 January 2026, three concern embedded constraint or the persistence of values: Prediction 2 of Appendix F, which section 3 calls preregistration-grade in substance; Prediction 4 of Appendix F, which it calls near-grade; and the fourth prediction of section A.3, which it calls a controlled design in the ordinary sense. The register tests none of the three, and the theory registration registers none of them as a proposition. In its words, Prediction 2, alignment drift with and without hardware-level constraint, is “the subject of its own drafted unit, with the hardware claim and the eighteen-month window declared untested inside it”. Prediction 4, value stability under adversarial conditions, “is adjacent to P21 and is not the same claim”. The fourth prediction of section A.3 varies the sequencing of value-relevant examples and measures the persistence of the early values against the later ones; the theory registration calls it the print’s “most precise proposal and the one it never built”, and it too is the subject of a drafted unit. Of the four falsifiers of section A.4, the same registration says of one: “That early-embedded values hold no persistent advantage over later modification is bounded from one side by P15, which measures whether installed gains decay under subsequent training, and is not otherwise decided, because the design the print itself proposes for it has never been built.” In section 6, the five predictions of Appendix F carry the print’s own numbers, prediction 1 to prediction 5; they are not the registered propositions P1 to P5, which Section 8 lists. Nothing in this register is a deciding test of P15, P12 or P21. Naming the two laws changes nothing printed here: the predictions, their dates and their grades stand as they were.

Note (3 October 2026). Nine passages above are changed where they stand, and this note records what each printed before. Section 1 said that the submission click remained outstanding, and the last row of the chain gave the eight draft registrations as awaiting the human click, with the status awaiting human submission. That was the state when this paper was first published on 25 August 2026; both passages now give the state as read on 3 October 2026. Section 3 said that the measurement went the other way, and it gave one figure of the estimator row, the regression exponent, as the blinded cross-architecture estimate; it now gives the row whole, as Paper II reports it, and says that the fit is one model’s and inconclusive. Section 3 also said that Google Willow carries both dates and printed one of them; it now prints both. The first rule of section 1 reserved the word preregistered for an accepted registry submission, and the abstract said that the paper never uses that word for anything except an accepted registry submission; the rule as amended keeps the word registered for an accepted registry submission and calls a design fixed by a dated timestamp before its run preregistered by timestamp, and the abstract follows it. The abstract said without exception that every core quantitative claim has at least one dated artefact fixing it before the measurement that tested it; it now excepts Paper VII’s fifty-domain cohort. Section 3 described the study-v draft registration without the limit that the theory registration states for it; it now quotes that limit. The chain’s row for 17 March 2026 gave its status as demoted by the author and no more; it now adds the reversal of 7 September 2026, which section 4 sets out.

Working Paper v2.1, 5 October 2026. Version history: v2.1 (5 October 2026): The title now follows the programme’s uniform order (paper, its law or framework, its own title), Register · The ARC Theory · The Dated Prediction Register: Forward Predictions as One Verifiable Chain; until this version it read The Dated Prediction Register: Forward Predictions as One Verifiable Chain. Section 1 now gives the state of registration as read on 4 October 2026, when the corrector-class audit (unit 06) was registered and public; until this version it gave the state as read on 3 October 2026, when the unit’s approval period at OSF was still running. The register now names the programme it belongs to by the programme’s current name, ARC/Eden Research Programme; until this version it named a series, ARC/Eden experiments. No registered law, prediction, kill condition or dependency changes; no result is added; no date of any claim moves. v2.0 (4 October 2026): Substantive. A census of every dated prediction in the programme is added as Section 8: the theory registration’s twenty-two propositions, quoted with their registers, laws and kill conditions; the registered corrector-class audit (unit 06) with its three registered predictions; the sixty-seven draft designs by unit and stage, with the two held designs and the one withdrawn counted; the outcomes of the printed and other dated predictions; and the predictions fixed inside designs. The chain records unit 06 as registered and public, and the print’s five Appendix F predictions are labelled prediction 1 to prediction 5 so that they cannot be read as the registered P1 to P5. The head and social titles now match the citation title, and the citation block gives the DOI, 10.17605/OSF.IO/C3NTF. No registered law, prediction, kill condition or dependency changes; no result is added; no date of any claim moves. v1.3 (3 October 2026): The first vocabulary rule of section 1 is amended: registered is kept for a registry submission that has been accepted, and a design fixed by a dated timestamp before its run is called preregistered by timestamp; the abstract’s sentence on the word follows the rule. Section 1 no longer gives the submission click as outstanding, and with the last row of the chain it gives the state of registration as read on 3 October 2026. The abstract excepts Paper VII’s fifty-domain cohort from its claim that every core quantitative claim has a dated artefact fixing it before its measurement. Section 3 gives the estimator row from Paper II whole, as one model’s fit and inconclusive; it gives both dates for Google Willow; and it quotes the theory registration’s limit on the study-v draft. The chain’s row for 17 March 2026 adds the reversal of 7 September 2026. The dated note above records what each of those passages printed before, and the note before it places this paper against Laws IV and V. The works the paper cites are linked to their sources. Until this version the paper gave its version as 1.1; it now gives 1.3, because the correction of version 1.2 was published without the paper’s stated version being changed. No registered law, prediction, kill condition or dependency changes; no result is added; no date of any claim moves. v1.2 (23 September 2026): The v1.0 history line now names the two verification lanes as the programme's own internal lanes. No claim, date, result or status has changed. v1.1: The provenance note at the end has been corrected for accuracy. No claim, date, result or status has changed. v1.0 first publication, assembled from the prediction register page and its machine register after two-lane verification of every link by the programme's own internal lanes. Corrections to this paper will be dated, listed here, and never silent.

Michael Darius Eastwood conceived and directs this research programme and is the author of this work. Across the programme, he has used more than six AI systems in parallel, under his own instructions, to stress-test his arguments, identify possible errors, and assist in preparing draft text from his own outlines. He determines what is adopted, revised or rejected and takes responsibility for the published content. These systems are tools, not authors.