# Evidence package — the grounded-reference programme (A13 to A18)

*2026-09-12. How to find, check and regenerate every figure in the two
manuscripts, *History-grounded reference and transfer* and *Authority
and reference repair*. The package is the repository itself at a named
commit; this note is its index. Hashes of every listed artefact are in
`grounded-reference-evidence-manifest.md`, generated from the committed
objects by `tools/evidence_manifest.py`.*

## What a reader can check, and how

Every claim in the manuscripts is one of three kinds, and each kind
points at a different artefact:

1. **A preregistered verdict** (A14's formation and transfer penalty;
   A16's dimension-specific reading; A17's inconclusive verdicts; A18's
   authority, correctness and restoration verdicts). Check it by finding
   the study's frozen text in the preregistration, the rule it names,
   and the analysis output the rule was computed on; then regenerate the
   output with the command below and compare.
2. **A descriptive figure** (A17's event-note cells; the qualification
   set's production counts; usage). Check it in the same analysis
   outputs, in the cells marked as descriptive; no verdict is attached.
3. **An accounting statement** (calls, attempts, retries, pauses,
   deviations, audits). Check it in the driver logs, the seed logs, the
   preregistration's dated amendments, and the audit output.

## The record, study by study

| study | frozen text (commit) | amendments | records | analysis output | tool and command |
|---|---|---|---|---|---|
| A13 | preregistration entry A13 (aac5d43) | ruling f091285 | `results/dialects/history/a13-s10…12-deepseek` | per-seed summaries in the result note | `tools/summarize_history.py --run <dir>` |
| A14 | A14 (78fa0bb) | A14.1 b13489a; A14.2 in the preregistration | `a14-s13…28-deepseek` (`.failed1` attempts kept) | `a14-population.json`, `.all16.json`, `.prereg.json` | `python tools/summarize_population.py --seeds 13-28 --pattern "results/dialects/history/a14-s{seed}-deepseek" --expected-model deepseek-v4-flash` |
| A15 | A15 3388de1; A15.1 36c1b43 | A15.2 220ed3e; A15.3 ecb3895; A15.4 in the preregistration | `a15-s29…59`, `a15pre-s45…47` | `a15-structure.json` | `python tools/summarize_structure.py --seeds 48-59 --pattern "results/dialects/history/a15-s{seed}-deepseek-flash" --expected-model deepseek-flash` |
| A16 | A16 27eaecd | A16.1 a5a3542; A16.2 6ef7f67; A16.3 50be940; audit note in the result note | `a16-s60…71-deepseek-flash` | `a16-crossed.json` | `python tools/summarize_structure.py --seeds 60-71 --pattern "results/dialects/history/a16-s{seed}-deepseek-flash" --expected-model deepseek-flash` |
| A17.0 qualification set | A17.0 30f3b1c; A17.0c–e in the preregistration; A17.0f 2cbd00e; A17.0g 289ae53; A17.0g.1 f93e461 | bookkeeping entries dated in the preregistration; the carrier and multiplicity rulings | `a17pre-s72…107` (aborted, failed and mislaunch copies kept and labelled) | `a17pre-decision-*.json` (archived); `a17pre-qualification-audit.json` (written rule) | `python tools/audit_qualification.py --out results/dialects/history/a17pre-qualification-audit.json` |
| A17 | A17 2b5f99e (implementation 86d8e06); frozen text `a17-freeze-proposal.md` 39c9908 | A17.1 e20d5bc; A17.2 49b44e6; A17.3 78537f1 | `a17-s108…119-codex-gpt-6-astra`; `a17-driver.log` | `a17-repair.json`; `a17-r0-crossed.json` | `python tools/summarize_repair.py --runs results/dialects/history --seeds 108-119 --model codex-gpt-6-astra --out results/dialects/history/a17-repair.json` |
| A18 | A18 1090174 (implementation eb0da6a); frozen text `a18-design-proposal.md` v2.1 | none; `tools/power_a18.py` for the planning scenarios | `a18-s108…119-codex-gpt-6-astra`; `a18-driver.log` | `a18-authority.json` | `python tools/summarize_a18.py --runs results/dialects/history --seeds 108,110-119 --model codex-gpt-6-astra --out results/dialects/history/a18-authority.json` |

Every regeneration is deterministic: the bootstrap uses generator seed
20260909 with 10,000 resamples, and the tools read only the archived
records. The tests that pin the analysis rules (`tests/test_population.py`,
`tests/test_history.py`, `tests/test_repair.py`,
`tests/test_repair_analysis.py`, `tests/test_authority.py`) run with
`python -m pytest tests/<file> -q`; the scripted-LLM tests need the
the simulator simulator and skip without it.

## What is in a run directory

`config.json` (the frozen configuration, the resolved model identity,
the accepted identifiers, the relations for crossed designs);
`calls.jsonl` (every model call: tag, request, response, resolved
model, usage, wall time, attempts); `episodes.jsonl` and `acts.jsonl`
(formation); `notebooks.json` and `records.json` (the checkpoint);
`qualification.jsonl`, `competence.jsonl`, `gate_pre.json`, `gate.json`
(the gates); `probes.jsonl` (R0 exchanges with the frozen classification
per note); `probes_<condition>.jsonl` and `done_<condition>.json` (A17
and A18 passes); `done.json`. Attempts that did not complete are kept
beside the records with `.failed1`, `.failed2`, `.aborted` or
`.mislaunch-appended` suffixes and are never read into a decision.

## The rulings

Every panel ruling that changed a claim or a design is quoted or
paraphrased in the preregistration at its date, and the manuscript
corrections that followed each review are listed in the two review
files `review-bbcfe62.md` and `review-bbcfe62-summary.md` and in the
commit messages 48b36ec, 3225c71, a2b49d7, 781c086.

## Known limits of the record

The DeepSeek pathway archived the resolved model per call and detected
the provider's model change; the Codex pathway reports no answering
model per call, so gpt-6-astra runs carry the requested identifier
only. Usage on subscriptions is tokens, not a bill. A17 ran two draws a
probe where its frozen text said three (A17.2). The "shared acts"
relation of A16 is planned, not recorded. The qualification decisions
were made with a predicate that omitted the written production
threshold and with a stricter invalid-act cutoff than written; the
audit tool reproduces both the archived and the written-rule readings.
The three 2026 arXiv papers and the four classical citations in the
related-work sections were verified against their records on 12
September 2026; the recorded prior-work search of the same day, its
queries, hits, verdicts and API refusals, is in
`grounded-reference-search-protocol.md`. The collation published on 12
September carried two paragraphs that its committed source pieces did
not, because the builder read uncommitted working-tree edits; the
published text was preserved, the sources were brought up to it, the
collation was rebuilt from a clean checkout and shown to reproduce the
published one apart from its commit stamp, and the manifest was
regenerated with the superseded one retained. The defect, its cause,
the check and the two paragraphs, which await their own review, are in
`grounded-reference-reconciliation-2026-09-12.md`.