aithinklab/runesmith-evidence

Runesmith evidence

The measurements behind the paper Beyond the Model (AI ThinkLab, 6 October 2026; not peer reviewed; DOI 10.5281/zenodo.23196763; website aithinklab.com), published with the Runesmith 1.0 release. Runesmith itself, its model card and the paper are in the model repository AIThinkLab/runesmith; the code is at https://github.com/Veristria/runesmith. The paper is also at https://aithinklab.com/paper/beyond-the-model.pdf, and its technical report (the full method and the complete negative record) at https://aithinklab.com/paper/beyond-the-model-technical-report.pdf.

This dataset is small: 170 claim rows, six short tables, the outcome tables, reports and recomputation scripts of the sealed studies SR7, U1, U2, U2b, LOC1 and SI, and the seal records, including LOC1's amendments, result wording and OpenTimestamps proofs. It lets a reader check the numbers the paper reports against the study results, recompute the exact tests of SR7, U1 and LOC1 from the published per-session tables, check LOC1's seals against the Bitcoin blockchain, and see the weakness of each claim next to the claim. It is not the raw record.

Provenance

• Claims index (claims.jsonl). The technical report cites rows of a claims index by identifier (M-0118 and so on). The index was made from the research record (the report's source document) by two AI extractors working separately: GPT Astra 6 (864 entries) and a Claude Opus model (604 entries). The two lists were reconciled into 1,309 rows, 159 of them found by both. This file holds the 170 rows the report cites, with ranges such as M-0225 to M-0231 expanded.
• Study results (sealed_arms.csv, sealed_tests.csv). Copied from the RESULT.json files of two sealed studies (SR7-FRONTIER and SR6-W, analysed on 2026-09-25) and, for SR7's per-session split and SR6-W's false promotions, from the technical report's extraction packs, which quote the study records. Nothing was recomputed.
• Memory-retrieval cohorts (m1_cohorts.csv). Copied from the technical report's table and text for the M1 study (section 5.2 of the report) and from the technical report's extraction pack, which quote the research record. The raw per-block receipts are not in this dataset.
• Runesmith Motion (motion_metrics.csv, j11_stretches.csv, j11_events.csv). Counts from the project's own files and ledger at three dated reads (2026-10-01T16:52:16Z, 2026-10-04T20:36Z and 2026-10-05T01:05Z), as the technical report's section 5.5.2 gives them, and the entries of the project log for 2026-10-04 and 2026-10-05, put in plain words. Reads are never mixed: every row carries its own read time.
• SR7's sessions (sr7_sessions.csv, sr7_recompute.py). The 324 session outcomes of SR7 (task, repository, arm, round, strict success, final status, cycle seconds), taken from the study's session receipts; the script recomputes the exact sign-flip test from this table alone and reproduces p = 443/524288.
• The October studies (after the record's cut, 2026-10-05; the technical report's sections 5.2.6 and 4.1, and section 7 of the paper). U1 ran the SR7 generation C7, its predecessor B and the shipped step g0 on fresh tasks from two public repositories, next to the same model with a simple fixed script in place of Runesmith's repair organ (u1_sessions.csv, u1_REPORT.md, u1_RESULT.json, u1_FIRST_EDIT_PROXY.json, and the recomputation by a second script written inside the project, u1_RECOMPUTE.json, with that script, u1_recompute.py, which also runs from the published table alone with --csv). U1's sealed protocol and its four amendments are included with their seal files (u1_PROTOCOL.md, u1_SEAL.txt, u1_AMENDMENT_1.md to u1_AMENDMENT_4.md, u1_SEAL_AMENDMENT_1.txt to u1_SEAL_AMENDMENT_4.txt), copied byte for byte with their Windows line endings so that each file's SHA-256 equals the digest in its seal file. U2 repeated U1's comparison on the same 17 tasks with a small and a large model on free routes (u2_sessions.csv, u2_REPORT.md, u2_RESULT.json, u2_RECOMPUTE.json; its sealed protocol u2_PROTOCOL.md with u2_SEAL.txt). U2b repeated the comparison on the 21 tasks U1 never started, three rounds, three model sizes, on free routes (u2b_sessions.csv, u2b_REPORT.md, u2b_RESULT.json, u2b_RECOMPUTE.json). SI estimated starvation integrity on 104 designed opportunities with no model available (si_opportunities.csv, si_REPORT.md). The files are the studies' own outputs, copied unchanged.
• LOC1, the sealed test of C7 against the shipped step g0 (2026-10-06; the paper's section 4). On 56 fresh tasks in three large public repositories, with one free repair model, LOC1 compared g0, C7 and HEUR, a hand-written trace-aware rule written by the study's designer. The outcome table is loc1_sessions.csv (one row per session: model, route, task, repository, arm, round, strict success, status, calls, cycle seconds, and analysed); the study's own report and result are loc1_REPORT.md and loc1_RESULT.json, and a second script's recomputation is loc1_RECOMPUTE.json. python loc1_recompute.py, run in a folder holding loc1_sessions.csv and loc1_RESULT.json, recomputes the three sealed tests from the table alone with exact arithmetic and reproduces all three p-values: C7 against g0 p = 5835/2097152 (PASS at its Bonferroni level of 0.0083), C7 against HEUR mirrored p = 101/131072 (HEUR_BETTER) and the replication on a second model, one task, p = 1 (NO_DIFFERENCE_DETECTABLE). The design was sealed before any task or session (protocol sealed 2026-10-06T10:41:04Z; first session 12:05:45Z; analysis 16:36:53Z), the wording of the paper's result section was sealed before any outcome was read (11:34:30Z, with a factual addendum at 12:07:51Z), and the fingerprints of the sealed files were submitted to OpenTimestamps, which anchored them in Bitcoin block 970170 (mined at 12:12:09Z) and block 970174 (12:48:21Z). The sealed files are included byte for byte: loc1_SEAL.txt, loc1_AMENDMENT_1.md to loc1_AMENDMENT_3.md with loc1_SEAL_AMENDMENT_1.txt to loc1_SEAL_AMENDMENT_3.txt, the stage-B protocol loc1_PROTOCOL.json with loc1_SEAL_STAGE_B.txt, LOC1_RESULTS_BLIND_TEXT.md and LOC1_RESULTS_BLIND_ADDENDUM_1.md with their .seal.txt files, and two preregistration manifests, PREREGISTRATION_MANIFEST_2026-10-06.txt and PREREGISTRATION_MANIFEST_2026-10-06b.txt, each with its OpenTimestamps proof (.ots, binary). To check a proof, put a manifest and its .ots side by side and run ots verify PREREGISTRATION_MANIFEST_2026-10-06.txt.ots with the OpenTimestamps client (it needs access to a Bitcoin node or block explorer); the second manifest was written after Amendment 2 was final, and the first lists an earlier draft of that amendment (the final one is bound in loc1_PROTOCOL.json, which the second manifest covers). A manifest line that begins RUNESMITH-LOC1-2026-10-06/ names the file here prefixed loc1_; the manifest's lines for a different study (NATURAL) are not shipped. The sealed PROTOCOL.md itself is not shipped, because it names the laboratory's internal systems and local paths; loc1_PROTOCOL.public.md is its text with each redaction marked and loc1_REDACTIONS.md lists them, and the sealed bytes (their SHA-256 is in the first manifest, which the first proof anchors) are available to reviewers on request.
• Seals (seals.csv). For SR5, SR5b, SR6-W, SR7, U1, U2, U2b, LOC1 and SI: the protocol's SHA-256, its seal time, the first evaluation start and the analysis time, so a reader can check that each seal precedes the outcomes it governs. The seals were recorded by the laboratory's own machinery; the U1, U2, U2b, LOC1 and SI protocols were also pushed to a private GitHub repository of the laboratory before their first session, and LOC1's were also anchored publicly (see the LOC1 entry above).

Files

File	Rows	What it holds
claims.jsonl	170	One row per claim the technical report cites: id (the report's M- number), claim, numbers (effect, n, test, p, interval), conditions, not_claimed, section, evidence_grade, evidence_grade_second, found_by, scrubbed.
sealed_arms.csv	4	Per-arm counts of the two sealed studies: sessions, strict successes, organ errors, false promotions, seconds per strict success, session statuses.
sealed_tests.csv	2	One row per sealed test: verdict, test, p-values, the better/worse/tied counts, the declared alpha, the family-wise threshold.
m1_cohorts.csv	5	The three M1 confirmation cohorts of the memory-retrieval study, the post-hoc conventional-index control and the same-code negative control.
motion_metrics.csv	82	Runesmith Motion's build metrics at three dated reads: plan items, applied drafts and their authors, model calls, tokens, cost records, check approvals, file sizes.
j11_stretches.csv	2	The two stretches in which the operator only read the project, measured against the lead investigator's definition of unattended.
j11_events.csv	13	The log entries of 2026-10-04 and 2026-10-05 around those stretches.
sr7_sessions.csv	324	SR7's sessions: task, repository, arm, round, strict success, final status, cycle seconds.
u1_sessions.csv	180	U1's session records: task, repository, arm, round, strict success, status, calls, cycle seconds, route, phase, and analysed (1 for the 170 sessions of the 17 tasks in the analysis).
si_opportunities.csv	104	SI's opportunities: template, condition, variant, tasks and how each ended, model exposure, and the outcome of every integrity rule.
u2b_sessions.csv	456	U2b's sessions: three model sizes, three rounds per task and arm, on the 21 tasks U1 never started.
u2_sessions.csv	68	U2's sessions: the same fields as U1's, for the small and the large model (17 tasks, two arms, one round each).
loc1_sessions.csv	542	LOC1's sessions: three arms (g0, C7, HEUR) with gpt-oss-20b on 59 tasks planned, 56 analysed, three rounds each, and two arms with gemini-3.5-flash on one analysed task; analysed is 1 for the 510 sessions in the analysis.
loc1_REPORT.md, loc1_RESULT.json, loc1_RECOMPUTE.json		LOC1's report, result and the recomputation by a second script.
loc1_POSTHOC_VIEW_CROSSTAB.json		Post-hoc, descriptive: strict success by arm and by whether the faulty file was in the model's edit view; 123 of 124 repairs had it in view.
loc1_recompute.py		Recomputes LOC1's three sealed tests from loc1_sessions.csv and compares them with loc1_RESULT.json.
loc1_PROTOCOL.public.md, loc1_REDACTIONS.md		LOC1's sealed protocol with marked redactions, and the list of the redactions.
loc1_SEAL*.txt, loc1_AMENDMENT_*.md, loc1_PROTOCOL.json		LOC1's seal files, amendments and stage-B protocol, byte for byte.
LOC1_RESULTS_BLIND_*.md, LOC1_RESULTS_BLIND_*.seal.txt		The sealed, results-blind wording of the paper's LOC1 section, with its addendum and seal files, byte for byte.
PREREGISTRATION_MANIFEST_2026-10-06.txt, PREREGISTRATION_MANIFEST_2026-10-06b.txt and their .ots proofs		SHA-256 of the sealed files, and their OpenTimestamps proofs (anchored in Bitcoin blocks 970170 and 970174).
seals.csv	9	Seal records of the sealed studies (see Provenance).

Fields that need explaining

• claims.jsonl: section is the numbering of the research record, not of the paper. evidence_grade is the index's own scale, 1 (strongest) to 6: 1 sealed, preregistered and independently recomputed; 2 sealed null, or closed before confirmation; 3 unit-tested engineering claim; 4 observed or exploratory; 5 descriptive; 6 hypothesis, untested. It is not the technical report's A-to-D scale. It is the extractor's judgement. found_by is astra, opus or both; the text is the first extractor's where both found the row, with empty conditions, not_claimed and numbers fields filled from the second. evidence_grade_second is set only where the second extractor graded the row differently (6 rows). scrubbed marks the rows whose text was edited for release (see below).
• sealed_arms.csv: strict_successes is judged on a session's final state, so a session can end budget_exhausted and still be strictly correct. For SR7 the four-way split is given: B is 34 public-pass and strict, 2 public-pass and not strict (the false promotions), 1 budget-exhausted and strict, 125 budget-exhausted and not strict; C7 is 55, 2, 2 and 103. strict_rate is derived here from the counts. In j11_stretches.csv, duration_minutes is derived here from the log times.
• sealed_tests.csv: arm A is the arm the hypothesis favours (C7 in SR7, SUIT in SR6-W). The units of SR7 are 54 task components with three rounds each, so its 162 sessions are not 162 independent tasks.
• m1_cohorts.csv: ratios are as the record reports them, with the direction in ratio_definition. Blank cells mean the quantity does not apply to that row.
• motion_metrics.csv: value is numeric; timestamps are in value_text. Authors are given by model name.

What this dataset does not show

• It is not an uplift measurement. No study compares a model working alone with the same model inside Runesmith. SR7 compares two generations of Runesmith's repair step with each other; SR6-W compares the shipped repair step with a fixed scaffold and the scaffold did better in direction (5/68 against 12/68, verdict NULL). U1 compared C7 with that scaffold driving the same model, inside the same kernel, and found no detectable difference (per task C7 better on 5, worse on 6, tied on 6; 20/51 against 26/51 sessions; 80% power only for an advantage of about 36 points).
• SR7 is one result. It holds inside its envelope: synthetic single-line regressions in two repositories, a paid author model, one cheap repair model, a comparison with the predecessor B, no ablation of the five changes. A small replication attempt on public repositories (U1: C7 20/51 against B 15/51) went in SR7's direction and was not significant. Six of the seven sealed tests of the series did not show their effect, nor did U1's two tests. LOC1 then compared C7 with the shipped step g0 on 56 fresh tasks in three large public repositories (42 of 168 sessions against 17; exact one-sided p = 5835/2097152 = 0.00278, PASS at its Bonferroni level of 0.0083), and SR7's p-value survives Bonferroni correction across all 23 sealed or registered comparisons the program has run. But in the same study a hand-written trace-aware rule, HEUR, repaired 65 of 168, more than C7 (mirrored p = 101/131072 = 0.00077, verdict HEUR_BETTER); the replication on a second model completed one task and could detect no difference; and LOC1 is one repair model, synthetic single-line regressions, three repositories, a rule written after C7 existed, and a change to the study's concurrency cap during the run that the paper reports. It licenses "C7 repaired more than the shipped step g0 within LOC1's scope" and "a sealed two-step record"; it does not license any claim that C7 beats a hand-written rule, a scaffold or a model working alone.
• The claims are extractions, not primary sources. The rows are sentences written by two AI extractors, with their own judgement of grade. The index has known faults: 182 of the 1,309 rows failed an automated check (most of them stale line numbers, which are not carried here), and the reconciliation method is not documented.M-0860 extracts an archived protocol file rather than the record's text (it gives SR7's task pool as 4 and 50 tasks from two repositories; the record says only "54 fresh tasks"). Some index rows have no backing text in the record; the paper uses none of those.
• No independent replication. The recomputations are second scripts written inside the project from the same raw files. All studies were run by AI ThinkLab with AI collaborators.
• No raw receipts. Apart from the outcome tables of SR7, U1, U2, U2b, LOC1 and SI, there are no per-session, per-task or per-block records, no task lists, no prompts and no model outputs here. The index's receipts (file paths and digests) and source.lines fields were dropped; the claim text is as extracted, so it can name the public repositories the paper names (the private ones are lettered, see Scrubbing) and can quote short digest prefixes or line references.
• Runesmith Motion's numbers move, which is why they are given at dated reads. "Done" means that the approved checks passed and the draft was applied, not that the feature works. "Free" is a route designation, not metered cost: the recorded cost estimate is zero or absent on every call, and the operator's own model usage is not metered anywhere.
• The unattended stretches are observational. The operator in these logs is an AI (Claude Opus 5.5) acting as the owner; the log is its own. The stretch of 2026-10-05 is two consecutive milestones, not sustained unattended operation, and it was logged after the paper's count of 2026-10-05T01:05Z, so it is not among the figures of the technical report's section 5.5.2.

Scrubbing

The text of the claim rows was left as extracted, with these exceptions: in 4 rows (M-0228, M-0890, M-1193, M-1210) the name of an internal system (the one that routes model calls, or the owner's application suite) was replaced by the technical report's own wording ("the route layer", "the owner's application suite") or the clause naming it was removed; in 7 rows (M-0120, M-0129, M-0145, M-0543, M-0551, M-0552, M-0561) a local folder path quoted in the record's section title was removed; in 6 rows (M-0118, M-0338, M-0859, M-0860, M-1143, M-1197) the names of the laboratory's private repositories (the task pools of SR7 and SR6) were replaced by "repository A" to "repository D". scrubbed is true on all of them; that is all. The index's receipts and source.lines fields are not carried. The files were scanned for key shapes, e-mail addresses, local paths and the names of internal systems before release.

How to cite

The paper:

@misc{horpestad2026beyond,
  author = {Horpestad, Lars O.},
  title  = {Beyond the Model: Runesmith, an Open Runtime That Improves Software and Itself, and the Instrument--Substrate Hypothesis},
  year   = {2026},
  month  = oct,
  note   = {AI ThinkLab. Not peer reviewed.}
}

This dataset:

@misc{aithinklab2026runesmithevidence,
  author       = {{AI ThinkLab}},
  title        = {Runesmith evidence: the measurements behind Beyond the Model},
  year         = {2026},
  month        = oct,
  howpublished = {Hugging Face dataset, AIThinkLab/runesmith-evidence}
}

License

The data in this repository is released under CC BY 4.0. The Runesmith code and model card are under Apache-2.0 in the repositories linked above.