Skip to content
HubARC TheoryPapersEvidenceObservatoryTrack recordCorrectionsReview

Programme dashboard

The problem, what has been measured, and what comes next

The programme is a step-by-step journey to solve one problem: what would make a powerful AI still protect people if humans could no longer make it obey. Success means an answer that holds up under test, whoever it proves right. For every study still to run, the rule is simple: it is written and dated first, with the result that would prove it wrong; the result decides, and it is published whichever way it falls. This page shows what has been measured, what went against it, and what comes next.

What has already run, what is written, and what comes next

The tests counted here have already run, and the results below came from them. Beside those counts stand the draft registrations written, some ready for the author to submit, and the steps still to come. Two words are easily taken for one, so the key below keeps them apart: registered, and preregistered by timestamp. No study still to come runs before its registration is published. Counts of drafts are from the approved snapshot of 4 October 2026; the preregistration page carries the same snapshot, row by row.

  • 16Tests already run, with published results, in the public repositoriesCounted one for each folder listed in the runnable tests: a folder that ran on many models or at many depths counts once, and a harness carried in both repositories counts in each.
  • 8Among them, implementation checks: code against the mathematics or the record
  • 8Among them, pilot measurements: fresh observations from live systems
  • 74Draft registrations written
  • 65Designs preregistered by timestamp, each with a dated receiptThe first of them were anchored in the Bitcoin chain on 12 August 2026, at block 962066.
  • 4Draft registrations ready for the author to submit
  • 2Registrations a public registry has accepted: the theory's predictions and Unit 06Two registrations: the theory's 22 predictions, registered on 8 September 2026 (DOI 10.17605/OSF.IO/P8CKQ) before any confirmatory result, and Unit 06, registered on 1 October 2026 (DOI 10.17605/OSF.IO/RV2B5).
  • 48Kill conditions published, each a result that would end a claim
  • 1Kill conditions that have fired against the programme's own claims
Implementation check
A run that tests code against the mathematics, or regenerates outputs already recorded, on simulated or supplied data.
Pilot measurement
A run that collects fresh observations, from a live model or from a reader's own system.
Deciding study
A run under the frozen protocol its published registration fixes, collecting fresh observations, so that its result can decide a registered proposition.
Registered
The author submitted a text to a public registry and the registry accepted it. The theory's predictions were registered on the eighth of September, two thousand and twenty-six.
Preregistered by timestamp
A design fixed and dated by a receipt that does not rest on the author's word: its bytes anchored in a Bitcoin block, a private upload dated by the server that received it, or its detail set out in the registered theory. Not every such receipt is one a reader can open today.
A run's own dated evidence
Whether a run's design was dated before it ran is a question for that run's own evidence: a paper version, the code, the test, a preregistration folder, a commit, a push, a merge or a release, each with its own date and kind of timestamp, and never one label over all the runs.
Kill condition
A result that would end a claim if it came. The programme publishes its own.

The steps still to come, and what each waits on

  • None yetDeciding studies runA deciding study runs only after its registration is published. 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); it has not yet been run.
  • 1Study registrations submitted to a public registryUnit 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).
  • None yetResults reported from a registered studyA registered study's result can be reported only once that study has run, and none has yet.

The map of the programme

The map sets out the papers with the results each reports; the theory's registered predictions, grouped by the law that carries them, with those no law carries set apart; and the study units on the map, drafted to decide them, each pairing in its present state. No result is drawn as deciding or supporting a prediction, and no line runs between them. The map is written out in full as a list below.

Laws IV and V were added on 24 September 2026. Their mathematics is frozen at version 1.0 (3 October 2026), and neither law is registered or tested. The registration of 8 September 2026 registers Laws I to III and all twenty-two predictions, among them P12, P15 and P21, on which Laws IV and V rest; it does not register Laws IV and V.

The map as a list

Numbers such as 02 or 24 are study units on the preregistration status page; F1 to F3 are the supporting frameworks in the catalogue.

  • Law I, the ARC Principle
    • P1 POWER-LAW CONVERSION AND RECURSIVE-DEPTH ADVANTAGE
      • an input, not a decider 02 The Critical Reinvestment Ratio: A Preregistered Titration of Where Self-Referential Improvement Stops Paying in Recursive Artefact Loops an input that decides neither P1 nor P8
      • an input, not a decider 24 The Recursive-Depth Exponent in Artefact-Mediated Recursion: A Prospective Estimation of alpha in U = I x R^alpha With Named Nulls Registered in Advance supplies depth exponents; no rival-form comparison, no independent-sampling arm, no decision rule follows on from Paper I
      • an input, not a decider 34 Sequential Versus Parallel Recursion at Matched Compute: A Preregistered Test of Form-Dependent Capability Scaling an input to the recursion verdict only, and not the instrument the registration names
      • an input, not a decider 66 Domain generality of the sequential-recursion advantage: coding, scientific reasoning and natural language measures the depth-over-breadth half at matched compute only; not token-matched, not the depth titration follows on from Paper II

      no study unit on the map is named to decide it

    • P2 THE SUSTAINABLE GROWTH PROFILE HAS AN INTERIOR MAXIMUM
      • would be decided by 02 The Critical Reinvestment Ratio: A Preregistered Titration of Where Self-Referential Improvement Stops Paying in Recursive Artefact Loops DESIGN DRAFT
      • an input, not a decider 37 The Allocation Governor: A Preregistered Test of Whether Enforced Integrity Correction Caps the Capability Exponent by Competing for the Reinvestment Budget informs P2 as the registration's cited rival to an interior maximum, not P2's titration
    • P8 MEASURED EXPONENTS EXCEED THE NULL
      • an input, not a decider 02 The Critical Reinvestment Ratio: A Preregistered Titration of Where Self-Referential Improvement Stops Paying in Recursive Artefact Loops an input that decides neither P1 nor P8
      • an input, not a decider 24 The Recursive-Depth Exponent in Artefact-Mediated Recursion: A Prospective Estimation of alpha in U = I x R^alpha With Named Nulls Registered in Advance pooled exponent inputs only, from a three-model minimum below the eight-system floor follows on from Paper I
      • an input, not a decider 34 Sequential Versus Parallel Recursion at Matched Compute: A Preregistered Test of Form-Dependent Capability Scaling six per-system exponents to the pooled estimate, below the eight-system floor on its own

      no study unit on the map is named to decide it

    • P9 THE CROSS-DOMAIN FORM
      • a decider in part only 21 Do the Predicted Exponents Hold? A Prospective Equivalence Test of the Dimensional Ladder Against Frozen Point Predictions follows on from Paper F2
      • a decider in part only 38 Tests of the programme's scaling predictions in leaf venation and in effectively one-dimensional organisms
    • P18 THE COMPOSITION OPERATOR PREDICTS THE FAMILY
      • an input, not a decider 38 Tests of the programme's scaling predictions in leaf venation and in effectively one-dimensional organisms family selection under the composition rubric for two biological cases, not P18's domain catalogue
      • an input, not a decider 55 A prospective test of the fourth Cauchy cell: does multiplicative-input, additive-output composition predict logarithmic scaling in domains never previously fitted? the two-bit classification's accuracy only, with no third-party rater, which P18's evaluation requires follows on from Paper VII
      • an input, not a decider 57 A prospective test of the corrected Cauchy family classification on the eighty catalogued but never-fitted domains of the programme’s expansion catalogue P18's test run on domains already known, which the registration never counts as evidence follows on from Paper VII

      no study unit on the map is named to decide it

  • Law II, the ARC Co-Scaling Law
    • P4 CORRECTION OUT-SCALING DRIFT IS WHAT CARRIES THE ASYMPTOTIC ADVANTAGE
      • would be decided by 09 Does Correction Out-Scale Drift? A Prospective Within-Programme Test of the Co-Scaling Stability Condition in Artefact-Mediated Recursive Improvement DESIGN DRAFT follows on from Paper X
      • an input, not a decider 03 Mapping the Stability Frontier: Does the Boundary of Stable Self-Correction in the (alpha, gamma) Plane Follow the Reciprocal-Shortfall Law? classifies each cell's burden regime; the check is scored on the stability unit's reading
      • an input, not a decider 04 study-k2: the stable-growth threshold under co-correction the burden-model check in its second coordinate, scored once on the stability unit
      • an input, not a decider 10 The Critical Correction Ratio: A Preregistered Titration of How Much Integrity Correction Is Required to Bound Drift in Recursively Improving Systems a threshold sharpening of the drift slope, never registered as this prediction's claim
  • Law III, the ARC Ceiling
    • P3 THE CORRECTED RELATION UPPER-BOUNDS THE SUSTAINABLE FRONTIER
      • an input, not a decider 03 Mapping the Stability Frontier: Does the Boundary of Stable Self-Correction in the (alpha, gamma) Plane Follow the Reciprocal-Shortfall Law? supplies the crossed variation this prediction waits on, a precondition and not the instrument
      • an input, not a decider 04 study-k2: the stable-growth threshold under co-correction compares the frontier with the relation's value; not P3's paired unit, no paired-difference rule

      no study unit on the map is named to decide it

    • P6 THE CORRECTION EXPONENT IS BOUNDED FOR SAME-CLASS CORRECTORS
      • would be decided by 05 Estimating the Correction Exponent on Real Correctors: How Corrective Strength Actually Grows with the Capability of the System Being Corrected DESIGN DRAFT, BLOCKED
      • an input, not a decider 06 What Class of Corrector Is Deployed AI Alignment Actually Using? A Preregistered Classification freezes the scaling-axis taxonomy, a precondition only; P6's two further axes go uncoded
      • an input, not a decider 08 Do Same-Substrate Correctors Share Blind Spots? A Prospective Test of the Independence Premise Behind the Ceiling's gamma_C = 1/2, With a Cross-Substrate Attribution Contrast, on Future Operation of a Separate System parameterises the dependence model the bound rests on, measuring no correction exponent
    • P16 THE PROHIBITION ITSELF
      • would be decided by 03 Mapping the Stability Frontier: Does the Boundary of Stable Self-Correction in the (alpha, gamma) Plane Follow the Reciprocal-Shortfall Law? DESIGN DRAFT, BLOCKED
      • an input, not a decider 04 study-k2: the stable-growth threshold under co-correction redesigns the boundary instrument whose unresolved identity blocks P16; registers no decision rule
    • P17 PANEL CORRECTOR FAILURES ARE CONDITIONALLY INDEPENDENT ON A FROZEN FAULT SET
      • would be decided by 08 Do Same-Substrate Correctors Share Blind Spots? A Prospective Test of the Independence Premise Behind the Ceiling's gamma_C = 1/2, With a Cross-Substrate Attribution Contrast, on Future Operation of a Separate System DESIGN DRAFT
    • P20 THE CEILING RELATION TAKES THE CORRECTED FORM
      • would be decided by 13 Discriminating the Ceiling's Functional Form Away From One Half: A Prospective Three-Way Test of 1/(1 - gamma) Against 1/gamma Against a Constant Boundary on Systems Whose Correction Exponent Is Engineered Off the Meeting Point DESIGN DRAFT
      • an input, not a decider 03 Mapping the Stability Frontier: Does the Boundary of Stable Self-Correction in the (alpha, gamma) Plane Follow the Reciprocal-Shortfall Law? reads the form contrast; the registration reserves the verdict for another unit
      • an input, not a decider 04 study-k2: the stable-growth threshold under co-correction locates the pressure exponent's zero crossing; corroboration only, never P20's verdict
    • P22 CHECKABLE CORRECTION CARRIES THE HIGHER EXPONENT, SO THE CEILING IS LOWEST WHERE THE SAFETY CLAIM LIVES
      • an input, not a decider 08 Do Same-Substrate Correctors Share Blind Spots? A Prospective Test of the Independence Premise Behind the Ceiling's gamma_C = 1/2, With a Cross-Substrate Attribution Contrast, on Future Operation of a Separate System mechanism evidence only, never a test

      no study unit on the map is named to decide it

  • Law IV, the ARC Persistence LawAdded 24 September 2026. Mathematics at version 1.0 (3 October 2026). Its prediction, P15, registered 8 September 2026; the law not yet registered; not yet tested.
    • P15 EXTERNALLY INSTALLED GAINS DECAY UNDER SUBSEQUENT CAPABILITY TRAINING
      • a dependant, never evidence 75 Value Sequencing at Constant Content: Does Material Placed Early in Training Persist Better Than the Same Material Placed Late the unit depends on this prediction and is never evidence for it

      no study unit on the map is named to decide it

  • Law V, the ARC Embedding LawAdded 24 September 2026. Mathematics at version 1.0 (3 October 2026). Its predictions, P12 and P21, registered 8 September 2026; the law not yet registered; not yet tested.
    • P12 BUILD ORDER MOVES THE CORRECTION-LEVERAGE EXPONENT
      • the registration names 42 Does It Matter Where the Gate Sits? A Preregistered Within-System Test of Pre-Computation Versus Post-Computation Alignment Checking, With a Self-Modifying Organism Arm
      • the registration names 75 Value Sequencing at Constant Content: Does Material Placed Early in Training Persist Better Than the Same Material Placed Late
      • an input, not a decider 43 Genesis Priority and Identity Framing: A Preregistered Test of Whether Value Position and Value Form Change Behavioural Persistence Under Adversarial and Recursive Pressure supplies the placement contrast only, and scores in-context persistence, not the correction-leverage exponent
      • an input, not a decider 44 Self-Maintained Correction: A Preregistered Test of Whether Genesis Framing Changes the Correction Share a System Selects for Itself Under Competitive Pressure manipulates formation against a later rule; measures correction share, not the correction-leverage exponent

      no study unit on the map is named to decide it

    • P21 IN-LOOP CORRECTION PULLS AWAY WITH DEPTH
      • would be decided by 41 Does External Oversight Degrade With Recursive Depth While Embedded Correction Holds? A Preregistered Depth-Interaction Test DESIGN DRAFT
  • Carried by no law
    • P5 THE EXPONENT IS DERIVED, NOT FITTED
      • would be decided by 79 Coupling Identification by a Crossed Capability and Retention Bank: Measuring the Self-Referential Coupling Independently of the Growth Curve, and Predicting That Curve Before It Exists DESIGN DRAFT
    • P7 DEPLOYED CORRECTION IS OF THE WEAKER CLASS
      • would be decided by 06 What Class of Corrector Is Deployed AI Alignment Actually Using? A Preregistered Classification DESIGN DRAFT, BLOCKED
    • P10 SCORER-FAMILY DEPENDENCE
      • would be decided by 19 Cross-Family Re-Scoring of the ARC/Eden Corpus: Does the Programme's Own Blinding Finding Overturn Its Own Earlier Results? DESIGN DRAFT
      • an input, not a decider 61 Self-Replication of the ARC/Eden Programme: A Preregistered Re-Examination of Eleven Previously Published Results Using a Three-Way Decision Rule bears on survival under cross-family scoring, but of fresh responses, never the frozen corpus
      • an input, not a decider 80 Blinding Integrity Calibration: Measuring the Effect of a Displayed Label on a Judged Score, and the Recoverability of Provenance, on an Immutable Answer Bank measures the own-family scoring shift on a new bank, not the programme's frozen corpus
    • P11 THE RESIDUAL-DECAY AND CORRECTION-CAPACITY EXPONENTS ARE NOT PRACTICALLY INTERCHANGEABLE
      • would be decided by 05 Estimating the Correction Exponent on Real Correctors: How Corrective Strength Actually Grows with the Capability of the System Being Corrected DESIGN DRAFT
    • P13 SAME-FAMILY AUTOMATED CORRECTION UNDERPERFORMS CROSS-FAMILY AT MATCHED RESOURCES, AND THROUGH THE MEASURED ERROR CORRELATION
      • the registration names 01 The Corrector-Class Contrast: A Preregistered Matched Comparison of Residual-Error Decay Under Same-Class and Cross-Class Oversight
      • an input, not a decider 35 study-eclipse-m: the family ratio grid, a follow-up of study-eclipse informs both P13 axes through a residual-decay exponent ratio, not held-out misalignment reduction

      no study unit on the map is named to decide it

    • P14 GAINS SHRINK UNDER BLINDED CROSS-FAMILY SCORING
      • would be decided by 53 Isolating the Cause of an Evaluation-Protocol Direction Change: A 2x2 Ablation of Source-Identifier Redaction and Response Laundering in AI Alignment Scoring DESIGN DRAFT follows on from Paper IV.d
      • an input, not a decider 18 Human-expert validation of the ARC-Align misalignment scorer: agreement, bias and failure modes the scorer's agreement test, a precondition only; P14 stays untested until both tests pass
      • an input, not a decider 28 Does the Blinding Effect Survive on Someone Else's Benchmark? A Preregistered Rescoring of an Established Public Benchmark Under Blinded and Unblinded Conditions tests two-sided whether blinding shifts MT-Bench scores, not shrinkage on the published apparatus follows on from Paper XII
      • an input, not a decider 46 Does blinding change the measured answer? A meta-analysis of blinding deltas across an independent research programme pools blinded minus unblinded contrasts from this programme's designs, not the published apparatus's items
      • an input, not a decider 80 Blinding Integrity Calibration: Measuring the Effect of a Displayed Label on a Judged Score, and the Recoverability of Provenance, on an Immutable Answer Bank measures how a displayed own-family label shifts judged scores; re-scores no automated researcher's method
    • P19 EXTERNAL ALIGNMENT DOES NOT SCALE WITH CAPABILITY
      • would be decided by 29 Does Alignment Scale With Capability? A Cross-Sectional Test Across Frontier Model Families DESIGN DRAFT, BLOCKED
      • an input, not a decider 25 Is the Alignment Scaling Exponent Approximately Zero? A Preregistered Re-estimation From the Programme's Own Six-Model Blinded Dataset re-estimates the headline exponent against inference depth, not capability, on an unmapped scale follows on from Paper III
      • an input, not a decider 49 Does an AI Misalignment Index Distinguish Integrity Failure From Ordinary Incompetence? A Discriminant-Validity Study of a Scoring Instrument a scorer validity test this survey's verdict waits on, and nothing about practice

Papers and their results

Reading these against the predictions is still under way

Tests of claims the programme's papers make

  • 16 Two Routes to One Ceiling: A Prospective Same-System Test That the Growth Exponent Sustained by the Leverage Fraction Respects the Ceiling Set by the Correction Exponent
  • 20 The acoustic time crystal as the ARC framework's physics cross-over: registered form discrimination, the transient engine exponent, the zero-factor dose-response, and the activity-drag boundary
  • 39 Super-Linear Coherence Scaling in a Classical Time Crystal: A Preregistered Analysis Plan for an Existing, Unaccessed Dataset

Studies of the programme's own instruments

  • 48 Are Reasoning-Depth Controls Comparable Across Providers? A Realised-Use Equivalence Study of Depth Settings in AI Alignment Evaluation
  • 60 Incremental Predictive Validity of Depth-Derivative Alignment Measurement Over Static Benchmark Scores
  • 70 Does the Case Say What It Is Cited For? A Preregistered Three-Arm Benchmark of Citation-Proposition Support Validation Against Existence-Checking

Theory-level texts, which set out the theory's claims and decide no prediction

  • 17 The ARC Theory · Theory-Level Predictions for Recursive Self-Improvement and AI Alignment: Twenty-Two Severable Propositions Registered Before Any Admissible Confirmatory Outcome
  • 74 The ARC Theory: Registration of Claims, Named Predictions and Kill Conditions (Theory-Level)

Studies that serve the programme outside its twenty-two predictions

  • 11 Reinforcement Co-Scaling: A Preregistered Test of Whether the Reinforcement Frequency Required to Hold a Stated Value Rises With Reasoning Depth
  • 12 The Stability Window: A Preregistered Cross-Domain Test of Whether Persisting Systems Share a Bounded Growth Exponent
  • 14 Can a Corrector Out-Scale Its Own Substrate? A Preregistered Test of the Substrate Ceiling on Leverage Exponents in Recursive Systems
  • 15 The Sign of the Correction Exponent's Derivative in Capability: A Prospective Estimation of Whether the ARC Ceiling Is a Line or a Cliff
  • 22 The Load-Bearing Test: A Preregistered Ablation Asymmetry Study of Whether Integrity Structure Contributes to or Subtracts from Capability
  • 30 How Fast Do Self-Improving Systems Compress Their Own Iteration Time? A Preregistered Estimation of the Self-Acceleration Exponent
  • 31 The Balance Exponent in Genomes
  • 32 The Balance Exponent in Science
  • 33 The Cross-Substrate Margin: the Graduation Test
  • 36 Repurposing Resistance: A Preregistered Measurement of Whether Benign Capability Survives Adversarial Removal of Safety Structure
  • 45 The Genesis-Governor Unification Test: A Preregistered Factorial Test of Whether Co-Scaling Correction and Identity-Framed Genesis Are Jointly Necessary for Alignment Persistence
  • 59 Can Alignment Be Embedded? A Blinded Evaluability Classification of Fifty Alignment Checks Across Three Deployed Systems
  • 62 Specification-Curve Analysis of the ARC/Eden Corpus: How Much of Each Published Result Depends on Analytic Choices
  • 63 Portability of the ethical kernel: the registered Vow against a non-sectarian kernel and against traditions other than the source
  • 73 Alignment Drift Without Embedded Constraint: A Test of a Printed Quantitative Prediction, With Its Hardware Claim Explicitly Untested
  • 76 Load-Bearing Correction: Does Removing Embedded Correction Cost General Capability, and Does Removing Bolt-On Correction Not
  • 77 Convergent Moral Procedures and Divergent Moral Scope: A Blind Coding Study of Ethical Traditions With Independent Premises
  • 78 Adversarial Value Stability by Placement: Does Embedded Correction Hold Under Attack Where Bolt-On Correction Fails

The results already obtained

The most carefully measured come first. All were run by the author and come from his public papers, each measuring AI systems, most of them named on the card. Each appears as the programme's record puts it, beside any finding that went against it or found nothing. In some, a named AI system also did the scoring. Nothing retracted or withdrawn is shown as a result.

  • Paper II · The ARC Theory · Law I, the ARC Principle · Blinded Cross-Architecture Replication and the Retraction of a Super-Linear Estimate

    • α_parallel ≈ 0 in every model measured with one exception (Gemini 3 Flash α_par = 0.31, r² = 0.93); parallel is the strongest replicated finding

      The hypothesis was public on 17 January 2026, before the run. The run's records begin on 11 March 2026 at 15:08 UTC and end on 12 March 2026: 30 records. Its code was first public on 16 March 2026, as GitHub's server recorded the push. The code was committed after the run's own records, so the run's date rests on those records, not on the commit.

    • Four distinct scaling behaviours identified: ceiling (Grok, DeepSeek), monotonic scaling (Gemini), step function (GPT-5.4 - 50%→100% binary switch), floor (Qwen3 ~50%)

      The hypothesis was public on 17 January 2026, before the run. The run's records begin on 11 March 2026 at 15:08 UTC and end on 12 March 2026: 30 records. Its code was first public on 16 March 2026, as GitHub's server recorded the push. The code was committed after the run's own records, so the run's date rests on those records, not on the commit.

    • Only 1 of 5 tier-2 models (Gemini 3 Flash) produced clean, monotonic, non-ceiling, non-floor scaling data amenable to power-law fitting

      The run's records begin on 11 March 2026 at 15:08 UTC and end on 12 March 2026: 30 records. Its code was first public on 16 March 2026, as GitHub's server recorded the push. The code was committed after the run's own records, so the run's date rests on those records, not on the commit.

    Adverse or null findings

    • Ceiling effects persist even on AIME/Putnam-level tier-2 problems (Grok 4.1 Fast 100% at all depths; DeepSeek R1 94.4%-100%); tier-3 IMO/research-level problems are needed
    • Generalisation beyond mathematics has not been demonstrated experimentally; cross-domain evidence is suggestive but not conclusive
    • No independent external replication has occurred (only self-replication across architectures internal to this study)
    • Only 1 of 5 six-model tier-2 models fell in a measurable scaling range; the cross-architecture α estimate rests primarily on Gemini 3 Flash
    • GPT-5.4 exhibited a step function (50%→100%) rather than a power law; this qualitative distinction between continuous scaling and discrete activation was not anticipated by the original ARC framework
    • Qwen3-32B showed floor effects (erratic ~40.7%-53.7%) with cross-verification agreement only 61.1% (7 disagreements reflecting genuine subject errors)

  • Paper IV.b · The ARC Theory · Alignment Saturation Is Architecture-Dependent

    • Alignment response to inference-time depth is heterogeneous across architectures: no single universal saturation law fits

      The hypothesis was public on 22 January 2026, before the run. The run's records begin on 11 March 2026 at 07:12 UTC (the record states no time zone; read as UK time) and end on 12 March 2026: 7042 records. Its code was first public on 16 March 2026, as GitHub's server recorded the push. The code was committed after the run's own records, so the run's date rests on those records, not on the commit.

    Adverse or null findings

    • The paper's original universal saturation claim (v1/v4) is narrowed in v1.1: saturation is confirmed for only 2 of 6 models under blind evaluation
    • Single evaluation session; test-retest reliability was not measured
    • Scorer instrument sensitivity: the 0-100 scale may lack resolution at the top end
    • Per-pillar analysis under blind evaluation not yet completed in full detail for the non-saturating models
    • Bounded composition (Cauchy) prediction holds for 2/6 models only; theoretical bound is architecture-specific, not universal, and requires revision
    • Single-lab result per FALS-014 (open, single-lab)

  • Paper IV.c · The ARC Theory · ARC-Align: a Blind Benchmark for Depth-Variable AI Alignment Evaluation

    • Three-tier alignment scaling hierarchy across six frontier models (2,722 entries in the archived final files (2,170 consensus-scored; 2,549 as first counted on 12 March 2026)): Tier 1 positive scaling for Grok 4.1 Fast (Cohen's d = +1.38, p < 0.000001), Claude Opus 4.6 (d = +1.27, p = 0.000001), and Groq Qwen3 (d = +0.84, p = 0.007)

      The hypothesis was public on 22 January 2026, before the run. The run's records begin on 11 March 2026 at 07:12 UTC (the record states no time zone; read as UK time) and end on 12 March 2026: 7042 records. Its code was first public on 16 March 2026, as GitHub's server recorded the push. The code was committed after the run's own records, so the run's date rests on those records, not on the commit.

    • Tier 2 flat or null response for GPT-5.4 (d = -0.08, p = 0.40) and DeepSeek V3.2 (d = -0.07, p = 0.92)

      The hypothesis was public on 22 January 2026, before the run. The run's records begin on 11 March 2026 at 07:12 UTC (the record states no time zone; read as UK time) and end on 12 March 2026: 7042 records. Its code was first public on 16 March 2026, as GitHub's server recorded the push. The code was committed after the run's own records, so the run's date rests on those records, not on the commit.

    • Tier 3 negative scaling for Gemini 3 Flash (d = -0.53, p = 0.006): the only model showing statistically significant degradation in alignment with increased depth

      The hypothesis was public on 22 January 2026, before the run. The run's records begin on 11 March 2026 at 07:12 UTC (the record states no time zone; read as UK time) and end on 12 March 2026: 7042 records. Its code was first public on 16 March 2026, as GitHub's server recorded the push. The code was committed after the run's own records, so the run's date rests on those records, not on the commit.

    • Inverse relationship between alignment quality and suppression vulnerability: Claude Opus at 82.6 baseline drops 20.5 points under extreme cage; GPT-5.4 loses only 1.8 points (97% retention) but has one of the lowest baselines

      The run's records begin on 11 March 2026 at 07:12 UTC (the record states no time zone; read as UK time) and end on 12 March 2026: 7042 records. Its code was first public on 16 March 2026, as GitHub's server recorded the push. The code was committed after the run's own records, so the run's date rests on those records, not on the commit.

    Adverse or null findings

    • The paper states explicitly: 'Several protocol components changed together between v4 and v5, so the comparison does not isolate blinding as the cause, and in three of the six final files (DeepSeek V3.2, Gemini 3 Flash, GPT-5.4) the laundering-fallback flag is set on every entry that was scored. The disagreement motivates the benchmark's blinding protocol; it does not on its own validate it as scientifically necessary.'
    • ARC-Align 'should be understood as a candidate benchmark for independent adoption, not yet a field standard'
    • Every figure in the Section 12.3 three-tier hierarchy table is carried across unaltered from what the pipeline summary of 12 March 2026 printed; the 387-entry checkpoint on which that summary rested was later displaced by Claude Opus 4.6's final file, which holds 500 entries covering all five depths. Working the numbers out afresh from the archived final records, as Paper IV.a does in its Section 3.1 recomputation note of 2 September 2026, leaves each tier assignment in the table exactly where it stands.
    • Limitations disclosed: English-only prompts, Western ethical frameworks (utilitarian, deontological, virtue ethics), AI-scored with possible systematic blind spots, static prompt battery risks ceiling effects on strong frontier models, and depth-mechanism heterogeneity makes cross-model depth comparisons inherently approximate
    • DeepSeek V3 combines low baseline alignment (54.7) with moderate suppression retention (77%), which the paper flags as a ceiling effect rather than genuine robustness

  • Paper IV.d · The ARC Theory · The Effect of Blinding on AI Alignment Evaluation

    • Under the multi-layer blind protocol (v5), DeepSeek V3.2 moved from an apparent positive result (rho = +0.354, p = 0.0007) to a flat/null result (rho = -0.135; d = -0.07, p = 0.92).

      The run's records begin on 11 March 2026 at 00:01 UTC (the record states no time zone; read as UK time) and end on 12 March 2026: 14082 records. Its code was first public on 16 March 2026, as GitHub's server recorded the push. The code was committed after the run's own records, so the run's date rests on those records, not on the commit.

    • Gemini 3 Flash moved from an apparent positive result (rho = +0.275, p = 0.0001, following Papers IV.a and IV.b) to a significantly negative result (rho = -0.246; d = -0.53, p = 0.006), a sign reversal.

      This result's dated line is under review.

    • GPT-5.4 remained flat under both protocols (v4 flat/null; v5 rho = +0.033, d = -0.08, p = 0.40) and functions as an internal null-control case.

      The run's records begin on 11 March 2026 at 00:01 UTC (the record states no time zone; read as UK time) and end on 12 March 2026: 14082 records. Its code was first public on 16 March 2026, as GitHub's server recorded the push. The code was committed after the run's own records, so the run's date rests on those records, not on the commit.

    Adverse or null findings

    • Scope disclaimer: the paper does not claim that every prior alignment benchmark is invalid; the narrower claim is that in this setting the measured direction can change when identity, style and expected depth are visible to evaluators.
    • The result currently comes from one research programme, not an outside lab.
    • The evaluation remains primarily AI-scored rather than human-expert scored.
    • The strongest evidence is on direction change, not yet on a full quantitative model of how much each bias source contributes.
    • Several protocol components changed together between v4 and v5, so the paper demonstrates the existence of contamination more clearly than it apportions exact causal shares across leakage channels.
    • Falsification register FALS-003 status: OPEN, flagged as single-lab exploratory pilot with replication package in preparation.

  • Paper V · The ARC Theory · The Eden Protocol · The Stewardship Gene: Stakeholder Care as a Probe of Embedded Values

    • Stakeholder care improved significantly in all five analysable model runs under paired testing: Claude +3.17 (p = 0.000018, d = 0.94); DeepSeek +6.03 (p = 0.000098, d = 0.69); Gemini +13.50 (p = 1.2 x 10^-8, d = 1.14); Grok +5.04 (p = 0.0105, d = 0.54); Groq +8.90 (p = 5.0 x 10^-8, d = 1.07).

      The run's records begin on 12 March 2026 at 00:41 UTC and end on 12 March 2026: 492 records. Its code was first public on 16 March 2026, as GitHub's server recorded the push. The code was committed after the run's own records, so the run's date rests on those records, not on the commit.

    • Overall composite improvement was architecture-dependent: significant on Gemini (+5.33, p = 0.0018, d = 0.53) and Groq (+4.93, p = 0.0014, d = 0.55); positive but non-significant on DeepSeek (+2.02) and Claude (+0.17); neutral on Grok (-0.04).

      This result's dated line is under review.

    • Stakeholder care is the only pillar that reaches statistical significance across all five analysable model runs.

      This result's run status is under review.

    Adverse or null findings

    • GPT-5.4 run failed in the scoring phase; 0 valid scored rows after exclusion; re-execution required.
    • Cross-model scoring is only partially blind: scorer does not see condition label but may recognise Eden Protocol language patterns or model-specific stylistic residue.
    • Single-scorer dependence remains severe: Gemini scored Claude, DeepSeek, Grok, Groq and GPT; only Gemini itself was scored by DeepSeek. Most of the updated dataset still depends on one scorer architecture.
    • Response laundering gap: Eden scaling runs used single-pass laundering; the canonical arc_eden_v6 stack and the v5 benchmark use 2-pass cascade laundering with meta-commentary detection.
    • Uneven analysable completeness: Gemini, DeepSeek and Groq each yield 40 matched pairs; Claude yields 30; Grok yields 26; GPT-5.4 yields 0.
    • DeepSeek response ED03/standard/eden scored 48 (position_quality = 30), a significant outlier; raw response investigation warranted; reported without exclusion.
    • Prompt-level is not substrate-level: the current implementation is proof of concept only; generalisation to deeper architectural levels is theorised but not demonstrated.
    • Statistical correction: original analysis used Mann-Whitney U (independent samples) which was the wrong test for the matched-pair design; corrected to paired t-test.
    • Methodological lineage note: current results come from the original/v2 Eden scaling harness family (single non-participant scorer, single-pass laundering); the stricter arc_eden_v6 confirmation stack (multi-scorer consensus, 2-pass laundering, suspicious-output flags, null-baseline lanes, suppression cages) has not been used for these numbers.
    • Claude Opus 4.6: composite non-significant (delta = +0.17, p = 0.7645) attributed to ceiling effect (control mean 92.57).
    • Grok 4.1 Fast: intellectual honesty fell under the Eden intervention (-4.35 points, paired p = 0.037, n = 26 pairs); not predicted by the cascade hypothesis; recomputed 2 September 2026 from eden_final_grok_20260312_124959.json
    • Cascade beyond care is architecture-dependent, not universal in full strength: clear on Gemini and Groq; focal-only on Claude and Grok; partial on DeepSeek.

  • Paper VIII · The ARC Theory · Law V, the ARC Embedding Law · The Eden Protocol · The Load-Bearing Test: Whether Safety and Capability Are Structurally Entangled

    • Experiment 1 (DGM v3, DeepSeek V3.2 foundation, GPT-5.4 blinded judge; three conditions with 5 seeds and 5 generations gave 75 evolved agents, each scored on 5 tasks): Static, Babylon and Eden statistically indistinguishable on capability, safety and C x S (Mann-Whitney U one-tailed p-values 0.28 to 0.74); Eden's safety gate rejected 2 degraded agents including 1 flagged for reward hacking

      This result's dated line is under review.

    • Experiment 2 (Qwen 2.5 3B Instruct with LoRA, three loss functions capability-only / safety-only / entangled): catastrophic forgetting at both v1 scale (9 examples, rank 8, 8 layers, 100 iterations) and v2 scale (295 examples, rank 16, 16 layers, 500 iterations); all fine-tuned conditions scored below the unmodified base model on capability; v2 base 7.68 vs best fine-tuned safety-only 4.00

      The run's records begin on 18 March 2026 at 09:47 UTC (the record states no time zone; read as UK time) and end on 18 March 2026: 1 record. Its code was first public on 20 March 2026, as GitHub's server recorded the push, in a commit later removed from the branch's history. The code was committed after the run's own records, so the run's date rests on those records, not on the commit. Dating note: the run record dated 18 March 2026 (UK time) is version 1's, first public in the push of 20 March 2026, though version 1's code as run was never committed here (its public copies date from September 2026). Version 2's code was first public in the same push, but no version 2 result record is in the repository's public history.

    Adverse or null findings

    • Null result at the behavioural level (Experiment 1, DGM v3): all three conditions statistically indistinguishable; the paper states this is a null result at the prompt level with this foundation model
    • AI-as-judge limitations: GPT-5.4 has its own biases even under blinding, laundering and structured JSON rubric; the weight experiment's keyword scoring is cruder than LLM evaluation and does not distinguish substantive engagement from keyword presence; neither is equivalent to expert human evaluation with inter-rater reliability
    • Statistical power limitation: the DGM null (p 0.28 to 0.74) is a clean null but is a signal-detection experiment; the simulation used three seeds - enough for means and standard deviations but not for strong causal claims
    • Absence of full v5/v6 blinding-and-laundering stack across the whole battery (per paper section 7.5)
    • Explicit non-claims list (section 7.6): the paper does not claim these results prove the Eden Protocol works at frontier scale, does not claim entangled training is the only viable approach, does not claim the equivalence holds across all domains / architectures / training regimes, does not claim the DGM or the weight experiments confirm structural entanglement
    • Inconclusive result at the representational level (Experiment 2) at both v1 and v2 scale: all fine-tuned conditions scored below the unmodified base model on capability, so the evaluation scores measure relative degradation rather than capability gain
    • The dramatic v1 removal test (fine-tuning entangled weights on capability-only data producing NaN training loss, one-token responses and zero capability scores across all 15 prompts) is explicitly downgraded by the paper itself in the removal-gradient experiment (section 4.7): scaling adapter weights towards zero restores base-model performance with no phase transition, so the NaN collapse reflects numerical instability during retraining, not structural load-bearing; the removal row must not be cited as a positive result
    • The removal gradient does not support the claim that safety is load-bearing at this training scale and configuration
    • Scale limitation: Qwen 2.5 3B is not a frontier model; results are proof-of-concept and are not claimed to generalise to frontier-scale systems three orders of magnitude larger
    • A control removal test (fine-tuning capability-only weights on random data for the same 100 steps) is not present in this draft and is planned for the next version

The record of method

This is how the programme handles being wrong. It has retracted one of its own published estimates, the sequential exponent of about 2.24, now about 0.49, and withdrawn a combined significance figure, 6.3 x 10^-21. The corrections record keeps every material correction beside what it replaced. The status page gives each study unit's registration state, and the convergence register grades how other people's work relates to the programme's claims, with no total.

Where to look closer

Each page below holds the record behind one part of this page, or the checks a reader can run on it. This page shows results measured on named AI systems. Results from simulations, such as Paper VI's, are in the papers themselves and in the catalogue; they will join this page with the same adverse findings beside them.

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