The ARC & Eden Research Programme

The Registered Programme

This programme tests whether recursively self-improving AI stays correctable. This page shows each study with the state of its registration. When a study runs, the result decides, and it is published whichever way it falls. A registered text is not edited afterwards; an update is a new version beside the one it follows.

You are not asked to trust the author. You are invited to watch him win or lose.

The predictions are dated before their evidence by unalterable routes: the sealed December 2024 record, the printed January 2026 edition and dated paper versions. The OpenTimestamps anchors prove those texts unedited since their vintages; the OSF forms freeze the analysis plans and await the human click. Reporting this page as “predicted nothing” is wrong. Reporting it as “preregistered” without qualification is wrong in the other direction.

Every study in this programme states the result that would refute it, in a dated registration written before any data is collected. This page is the live index: what is specified, what is drafted, and what each registration commits to, in the order the commitments were made.

Be exact about the status, because the table below is and an earlier version of this sentence was not. The study registrations are written, dated and prepared as draft registrations awaiting human submission, except Unit 06, which was registered on OSF on 1 October 2026 (DOI 10.17605/OSF.IO/RV2B5): every other study row in the table reads “private draft, awaiting the author's submission”. The theory-level row is not a study registration and does not read that way. The theory-level registration, the twenty-two conditional and severable propositions of the framework, is on the public registry under DOI 10.17605/OSF.IO/P8CKQ, registered 8 September 2026 at version 1.100 and updated at version 1.102 on 13 September 2026; its row on the table carries that state and links to the registration itself. Read the study rows as what they are, a statement about one platform’s template, and not as a statement about whether the predictions were fixed before their data. They were. A registration form is mechanically a timestamp and a template, and the timestamps here are the stronger half: a sealed record dated 8 December 2024; the book in print since 2 January 2026, ISBN 978-1-80605-620-0, closing with prediction appendices; and a byte-for-byte anchor in Bitcoin block 961,340 that anyone can check against a public chain without trusting this site or any platform outliving it. Preregistration exists to stop a prediction being edited after the result. A printed book cannot be edited at all. What the template adds, and what is still outstanding for every study on the table, is the analysis plan fixed in advance in the registry’s own form. That is a standing rule of this programme rather than an oversight, and it cuts both ways: no tool, script or agent in this estate submits anything anywhere, ever, so a registration reaches the registry by a human hand or it does not reach it at all. Dating is real here. Public filing has begun, at theory level and nowhere else yet, and a reader should weigh each row by the status its own column carries.

This paragraph exists because several AI systems, reading this page cold, concluded there was no preregistration at all. They were right about the filings and wrong about the programme, and the sentence they were reading caused both halves of that: it claimed a “timestamped public registration” for every study while the table beneath it said the opposite forty-six times. Overclaiming in one direction invites a reader to overcorrect in the other, and that is what happened. The sentence was the error and it has been corrected here rather than quietly. What the correction does not concede, because it is not true, is that the predictions came after their evidence. They did not, and the dates are checkable below. The first filing has since gone public and its status column changed with it, which is the mechanism working rather than an exception to it: the columns are generated from the dataset this page publishes, so read the column, and expect the prose to follow it.

This page is the programme's registrations. Its companion, the dated prediction register, is the programme's predictions: every claim made before its evidence, in date order, including the printed nine, the March packet, and the one that failed. The theory-level registration on the table below, which is the one that states what the framework predicts rather than how a single study will test it, has its propositions set out in full at the twenty-two registered predictions.

The book is a prediction register, and it is the one nobody can edit

Infinite Architects went into print on 2 January 2026, ISBN 978-1-80605-620-0. It closes with two appendices that do, in ordinary language, what a registration form does in a template. Appendix A section A.3 sets out four predictions and is followed by A.4, headed Falsification Criteria, which states in advance what result would sink the framework. Appendix F sets out five more, numbered, each with a deadline or a threshold: meta-cognitive emergence by 2028; alignment drift above 15 per cent without embedded constraint and below 5 per cent with it, inside eighteen months; recursive capability gains above 300 per cent by 2029; value stability under adversarial conditions; convergent consciousness signatures across biological and artificial systems. It ends: “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.”

Nine numbered predictions, with kill criteria, fixed in print before the programme that tests them existed. A registry entry can be withdrawn and a website can be edited tonight. Fifteen hundred printed copies cannot.

And the chain is checkable at the numbers. Appendix F’s second prediction names 15 per cent and 5 per cent. The registration that tests it, study-v on the table below, states its own hypothesis as a “threshold test against printed 15 per cent and 5 per cent, separately not jointly”. The word the registration uses is printed: it is pointing back at the book for its numbers, not choosing them after the fact. A numeric threshold published in a printed book with an ISBN, then carried unchanged into a formal hypothesis, is the thing preregistration exists to guarantee, reached by a route that predates the form and cannot be edited by anyone.

The programme also kept a public laboratory notebook while the work was happening. The ARC alignment scaling report was commenced on 10 March 2026 and carries the version stamp live, updated in real time: a step-by-step narrative of every version, every methodological correction and every failure, written as the results arrived rather than reconstructed afterwards. It records, among other things, that blinding introduced at version five exposed false positives at version four. Writing that down while it is happening is the opposite of retrofitting.

The March 2026 packet, and the thirteen minutes that matter most

On 17 March 2026 a preregistration packet was committed to the public repository these papers cite, for a twelve-domain extension of the Paper VII comparison. It is a preregistration by every element. A locked candidate list of twelve named domains, each carrying its operator class, its predicted family and its predicted model. A written protocol with a stated primary endpoint. Inclusion rules. Declared exclusion risks. File checksums. An OSF component registration document, written and ready. The protocol’s own words: operator class is locked before data extraction; predicted family is locked before data extraction.

The run scored ten of twelve, p = 5.44e-4.

Then this, from the public commit log, with its timestamps:

2026-03-17 00:43   Pre-registered 12-domain extension: 10/12 confirmed (p = 5.44e-4)
2026-03-17 00:47   Fix stale v1 references, adopt conservative miss posture
2026-03-17 00:56   Downgrade 12-domain extension to pilot dry run, fix all stale references

Thirteen minutes after recording a successful result, the author demoted his own claim. The reason is written into the packet and left there: extraction and a first execution of the analysis had already been performed that day, before any external timestamp existed, so on the packet's own written condition it could not serve as a clean prospective preregistration. That reason was split in two on 7 September 2026, and only half of it survives. 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 and the run is preregistered in substance; the packet's own condition, deposit to a registry before extraction, was not met, and that half of the demotion stands unreversed. 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. The file carries the instruction he wrote to himself, which is still in it: do not upload this packet unchanged as a preregistration. Its status field reads archived_local_packet_exercised_before_timestamp.

Nobody required any of that. No reviewer, no registry, no coauthor, no funder. The result stood. The claim was reduced anyway, that same night, in public, by the only person who stood to lose by it.

So read the record carefully, because it does not say what a quick reading says. The predictions are in the file: twelve of them, per row, dated, and fixed before the data, on a manifest whose own status field at that commit read draft_local_not_locked. What was missing was never the prediction, and it was not the timestamp either: the repository had already fixed the twelve rows in public at 00:19:19 UTC that night, attestation class public repository commit date. What was missing was a registry form, and by March the author was already policing that distinction against himself more strictly than the convention requires. An archive that contains a commit demoting its own successful result is an archive that can be trusted about the rest of its contents. That is the argument this page would rather make than the easy one.

The diligence table

Nothing in this programme gets defined after its results arrive. Each registration freezes its definitions, estimator and stopping rules at the moment it files, and the register below shows the honest state of every one: filed and frozen, still drafting with its open decisions marked in the text, or resulted. The drafts wear their unfinished decisions as visible markers rather than smoothing them over, because a tidied placeholder reads as finished and hides the gap. Registered on OSF: 2, the theory registration (8 September 2026) and Unit 06 (registered 1 October 2026). No study registration has resulted; submission is operator-held, every row states its own honest position, and the census below counts the theory-level registration separately.

Anyone can claim a stack of preregistrations. This table is the claim made checkable: every registration, its one-line hypothesis, whether its analysis is frozen, and exactly where it stands today. Registrations that recut or succeed the same experiment share a family marker, so the count can never be accused of inflating one design into several rows.

Read this line before the table. Every STUDY row below is a private draft registration prepared and populated on OSF, awaiting the author's review and submission, except Unit 06, which the author submitted and OSF registered on 1 October 2026 (public since 4 October 2026; DOI 10.17605/OSF.IO/RV2B5). One study registration has been submitted and accepted, Unit 06: the programme's own rule is that only the author submits a registration, so any public sentence implying that more study registrations are FILED is false, and this dataset says so first. One row on this table is not a study registration and these sentences do not describe it. The programme's theory-level registration, the twenty-two conditional and severable propositions of the framework, is on the public registry under DOI 10.17605/OSF.IO/P8CKQ: registered 8 September 2026 at version 1.100 and updated at version 1.102 on 13 September 2026. Its row carries that state, and it is counted on its own line of the census and never inside the draft counts. One row (study-ae, the ceiling and exchange-rate reciprocity test) was withdrawn as vacuous on 2026-08-12 before filing: shared accumulation law forces agreement, so a test that cannot fail is not a test. Its replacement carrier is the cross-class to same-class correction-exponent ratio, already on the register.

Study registrations: 67 private drafts prepared · 4 final-ready · 58 working · 1 withdrawn before filing · 2 held privately, not yet published, never uploaded · 1 submitted · 0 resulted · generated 2026-09-16
Not a study registration, and not counted above: 2 theory-level registration on the public registry · 1 submitted · 1 accepted · Unit 06 was registered on OSF on 1 October 2026 and became public on 4 October 2026, at the end of OSF's approval period (DOI 10.17605/OSF.IO/RV2B5).; registered 8 September 2026 at version 1.100, updated 13 September 2026 at version 1.102, DOI 10.17605/OSF.IO/P8CKQ

The totals count STUDY registrations only: designs written to test a single proposition, each awaiting the author's submission except Unit 06, which the author submitted and OSF registered on 1 October 2026. The programme's theory-level registration is counted on its own line, with its own submitted and accepted, and never inside the draft counts, so neither number can be read as the other. A row reaches the registered tier by curation against the public registry, never by derivation from a folder. The row states carry the date on the census: the study rows were derived from the estate folders on 3 September 2026 and the registered row was corrected against the public registry on 16 September 2026. The estate is the authority on membership and it grows; scripts/build-registration-diligence.py names any unit this table does not yet list.

One rule governs this page: a design appears here only after its registration is timestamped on OSF. Until then a study shows a name and a status and nothing else, because publishing a design ahead of its timestamp would spend the very thing a registration exists to protect. As registrations land, their rows fill in with the hypothesis, the kill condition and the OSF link, and nothing already shown is ever edited afterwards.

One family reshaped, in public. The reciprocity family, listed as ceiling-exchange-rate reciprocity, was reshaped after audit. The earlier two-independent-ways-must-agree phrasing was withdrawn as vacuous: shared accumulation would force the two answers to agree, so a test that cannot fail is not a test. Its replacement is one accumulation law with two consequences (the conversion exponent and the correction ceiling), and the cross-class ratio already on the register carries the unification instead. The row is preserved as the reciprocity registration; the withdrawn phrasing is not on it.

The programme at a glance

quantity	count	meaning
Verified ready to register	4	packets machine-verified: manifests, hashes, randomisation
Study registrations submitted to OSF	1	the preregistration status page carries each study registration's state
Theory-level registration	1	DOI 10.17605/OSF.IO/P8CKQ, registered 8 September 2026 at version 1.100, updated at version 1.102 on 13 September 2026
Unit 06	1	Unit 06 was registered on OSF on 1 October 2026 and became public on 4 October 2026, at the end of OSF's approval period (DOI 10.17605/OSF.IO/RV2B5).
Held for disclosure review	2	withheld deliberately, pending a legal ruling

Registered studies

Unit 06 was registered on OSF on 1 October 2026 and became public on 4 October 2026, at the end of OSF's approval period (DOI 10.17605/OSF.IO/RV2B5).

Each unit's registration and evidence state, and what the unit can decide, are set out row by row at the preregistration status page.

Why this page exists

Most research is reported after the fact, when every choice can quietly fit the result. This programme runs the other way: the predictions, the analysis and the conditions of defeat are fixed in public first, and the data arrives second. Fixing them is not the same as filing them, and the table above says which of the two has happened. This page makes that order visible. Whatever these studies find, the record of what was promised will already be here, and it cannot be rewritten.

How to weigh a programme like this, including the case against it, is set out at how to weigh this. The current state of every published claim, including the retracted one, is on the Research evidence spine.

Machine-readable: registered-programme.json · regenerated as registrations land · designs withheld until timestamped by construction of the exporter.