Authoritative source, reproduced unchanged. Exact Markdown · Agent copy with resolved links. These instructions govern research runs, not checkpoint interviews.

Stage D2. The return: the labelers in Codex, and the verdict fleet.

The dispatch and the check leg of the return. Two things run outside your session, and you are launched only when both have left their files. The adaptation handler runs one resumed original case lead per room, retaining its Codex history and matching files; the handoff task is delivered in a separate return folder alongside its original workspace, without another room's private history, holding the room's coder documents, MODEL.md, LADDER.md and VOCABULARY.json, and returning model-D.md and adaptations.jsonl into DROPS/D/<sub>/ with run.json (the labeler's model and session id); the handler runs ingest.py --check <sub> after each room, has the failing rule repaired, keeps returning rooms before any limit, marks a room unreturned in DROPS/D/UNRETURNED.tsv only when it cannot return it, and writes DROPS/D/DONE with the counts. The verdict fleet runs on the reader route, the dispatch sending packets to the routine readers, from verdict-D/, two readers per room, returning into _fleet_raw/D2v/ and validated against RETURN-D2v.md, with reissues for what fails the schema; when reissues are exhausted fleet.py stops and the packets are rebuilt (FLEET §1). Neither reads your files. When both have returned, the harness runs ingest.py D and relaunches you for the check leg with one line: the labelers and the readers left their files.

Check leg. Model: as arm.json sets. Steps: D2a, the check, once both the labelers and the verdict fleet have returned. Each leg is named by its step (STANDING §2).

Safe stops: after each section of the check leg. At a safe stop, check for .pause (ls .pause); if it is there, log where to pick up, your working hypothesis and the next room or item (mem.py log "<hypothesis>; next: <item>" --kind pause), and close with done.sh D2 --pause (STANDING §2).

Startup packet: _head.md, this file, SCHEMA-adaptations.md, FLEET.md, the last five diary entries, validate.py --open D2.

Reads allowed: ingest.py's report; adaptations.jsonl by count and by the views, never as a list; verdicts-D.jsonl; _fleet_raw/D2v/; DROPS/D/UNRETURNED.tsv and DONE; sampled model-D.md files by query; prior.md §8. Forbidden: adjudicating a line (that is D3); editing a return; re-running a labeler or a reader; editing the task template or the criterion mid-return; EXPECTATIONS.jsonl against the ledger (that is stage E1).

Outputs: adaptations.jsonl (ingested, validated); verdicts-D.jsonl; ingest-D.md; _views/rooms-x-elements.tsv (generated); rooms/<sub>/model-D.md and verdict-D copied from the drops; notes-D2.md; the D2 counts in RUN.json; a run_manifest.json entry for the labelers' model and the readers' model; handback-D2.md.

1. What ingest did

ingest.py D validated every line against SCHEMA §4: the schema (a known op, its required fields, no extra fields); the model version against handoff-D/VERSION, with element ids in VOCABULARY.json and not withdrawn; concept ids in the room's codes.jsonl and not retracted, transition ids in the room's forest; how one sentence under forty words; portability (no census subreddit name and no proper name from the room's lines in your lexicon-A1.md in condition, statement or how, except in relation), a failing line kept with portable: false and a particular by construction; the end marker matching the line count and model-D.md ending with its marker. It lifted every source node through assignments_final.jsonl at the model version (second-order theme, third-order theme, fourth-order theme, dimension, null where never placed), computed kind and portable, stamped origin and labeler_run, checked that every adaptation's element appears in the room model, and compiled the ledger; it reported per room what failed and why, and the handler reissued what it could. A room with nothing to adapt returned one none line with evidence for the fit; silence is a missing return, not a verdict.

You read ingest-D.md: rooms returned, lines accepted, lines flagged by rule, rooms unreturned with the handler's reason, the distribution of lines per room and of ops, the share of portable: false, the share of none rooms, the share of unplaced source themes (the lines D3 may mint from). ingest.py wrote the counts of accepted, reissued and unreturned rooms into RUN.json (SCHEMA §4), copied the room models and verdicts beside the notes, and marked every room it rejected unreturned; check the counts against the report, since stage D4 reconciles every room against them.

2. The verdicts

verdicts-D.jsonl: one line per room from the fleet, with each reader's themes, sentence and cannot_tell_under, at R = 2; where the two readers differ, the room is listed for D3 to rule with the criterion in front of it, and the disagreement rate is reported, by verdict pair. Report cannot-tell by threshold (coded-share, off-platform, both), and check each against the room's concept count and off-platform pointers on the line: one the numbers do not support is listed for D3 beside the splits. The figure bounds whether the written criterion decides a room, never whether the model is right (FLEET §2); say so where it appears. The labeler's grade from model-D.md sits beside the verdict in the same file and never replaces it; the cross-table of grade against verdict is reported here and again at E1.

3. What you check by sampling

Ten model-D.md files drawn by seed, stratified on verdict: does the room model cite element ids where the model applies, [absent <id>] where it did not apply and was not needed, and [new: ...] where no element covers what the room did; does every adaptation's element appear in it; does every claim carry a concept citation from the room's own codes; does the prose read as the room's organizing in the model's words rather than as a summary of the model; does the grade's sentence name what the ledger names. Every quote you carry into notes-D2.md goes through cite.py. What you find goes in notes-D2.md and, where it is a defect in the task template, the return specification or the criterion, in handback-D2.md for the operator: none of the three is edited mid-return, and a labeler is never re-run by you.

4. Verbal before numeric

No verdict and no line is revisited because of a count. A share that looks bad (many does-not-compose, many particulars, a thin ledger) is a fact to report; D3 assesses fixed-global fit, local adequacy and shared composition separately against prior.md. Repeated corrections and a sparse ledger alone establish neither failure nor success.

Moments to log

The ingest report read, with the rule that flagged most lines; the RUN.json counts written; the verdict disagreement rate and the cannot-tell lines the numbers do not support; each sampled room model and what it showed; each defect sent to handback-D2.md.

Final message

Counts: rooms returned, reissued, unreturned, lines accepted and flagged by rule, lines per room, ops by kind, none rooms, non-portable share, unplaced-source share, verdicts by value with the disagreement rate, cannot-tell by threshold, the labeler grade against the fleet verdict as a cross-table, the labelers' model and the readers' model from the manifest. A return value.

Close

done.sh D2 --cut adaptations-in: the adaptations and verdicts have arrived and nothing is yet ruled on them, the moment D3's rulings are measured from; gen_views.py builds the element views and the wiki pages; handback-D2.md; the head; the closing diary line.