The added file is a staging copy of historical agenda-panel tallies into an intake batch; each decision carries an exact provenance pair (legacy commit ab69854... plus a file-and-index locator, e.g. proposals/agenda/2026-08-31-agenda-33432478115/scores.yaml#tallies[3]) and duplicates the original tally verbatim in a `legacy.record` block, so nothing is silently rewritten. Fields that were never recorded in the source runs — runId, model, generatedAt, inputHash, reason — are explicitly null rather than reconstructed, which is the honest handling §3.8 asks for; no locator is invented and no score is changed or re-certified. The tooling found no citation identifiers in the added lines, so §3 real-citation concerns do not arise. The docs/DATA_MODEL.md paragraph accurately describes the staging step, states the existing scores.yaml files remain authoritative, and does not present anything as ratified or human-reviewed; no check, gate, or review budget is altered, and AGENTS.md is untouched. My one reservation is the repeated `details.seatStatus` anchors ('scored 78' / GPT-5.1 'scored 76') applied identically across three different runs while many per-item entries in the 2026-08-31 run record GPT-5.1 as 'failed' — a summary field that sits awkwardly with the item-level data — but this is run-level bench metadata with no published consequence and no locator claim, so I treat it as ordinary imperfection rather than fabricated provenance. Nothing here reaches readers, no assessment is mutated, and the OMITTED FILES list is empty.

The change adds an append-only historical copy while leaving the existing `scores.yaml` files present and authoritative, so it does not silently alter prior judgments. Each decision preserves an exact legacy commit and file locator, along with the recorded tally and named panel seats. Missing reasoning, run IDs, models, generation times, and input receipts are represented as null rather than reconstructed or falsely attributed. No citations, confidential source material, reader-facing assessments, constitutional text, or validation checks are added or weakened.

The pull request adds an immutable intake batch under `proposals/intake/` and updates `docs/DATA_MODEL.md` to stage 78 historical agenda scores ahead of an intake migration. No normative rules or files such as `AGENTS.md` are modified. The records accurately document provenance and commit refs while honestly noting where historical reasoning or receipts were not retained, adhering to epistemic tracking requirements without altering published canonical assessments.

The diff only adds a descriptive paragraph to docs/DATA_MODEL.md explaining the staging intent and a single immutable intake YAML that copies 78 historical agenda tallies. Each record keeps the original candidate text, highs/advances/concerns/seats exactly as tallied, points at the authoritative source via legacy.commit ab698546200770ac95612af49f5056d30fd6f21b plus the scores.yaml#tallies[N] locator, and explicitly nulls runId/model/generatedAt/inputHash/candidateHash and other fields the PR states were never retained. No AGENTS.md edit, no citation identifiers, no public assessment or reader change, no reconstructed reasoning presented as fresh, and no check-weakening; the originals remain authoritative until a later consumer cutover.

This is a staging migration, not a publication: it adds a new immutable intake file and a documentation note, and explicitly leaves the authoritative scores.yaml files and all readers/workers untouched until a separate consumer PR, so no published judgment is silently changed (§3.14, §3.15). Provenance is handled honestly: every record preserves the exact legacy commit and tally locator (§3.8), and fields that were not retained — runId, model, inputHash, generatedAt, reviewer reasoning — are set to null and declared unknown rather than fabricated, which is exactly the right way to migrate lossy history. No citations or identifiers are added (citation verification found none), no AGENTS.md or gate/review-limit change appears in the diff, and the batch is presented as unratified history awaiting review, not as a fresh judgment. The visible records are internally consistent between the top-level scores and their legacy copies. I cannot re-verify the claimed identity match against commit ab698546… myself, but the diff supplies a verifiable locator and the PR runs CI, so there is no positive evidence of fabrication or check-weakening here.