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

Stage D3. The return: adjudication by element.

Model: as arm.json sets for this stage (your platform file lists the model and effort). Legs: one per element group (a fourth-order theme's elements, or one element with many lines), packets of at most one hundred lines; then one leg for the unattached propose lines; then one for the verdicts and the uncomposed rooms; then the re-check handoff; then the close. A leg may close earlier when you choose. Every model revision is a move, logged with mem.py log --kind move (the Grodal, Anteby and Holm move named) and snapped with python3 tools/snap.py D3 move-<move>-<n> when it is made. Fleet: the labelers again, for the re-check rooms only, through the handler; the verdict fleet again for rooms whose verdicts rested on a rewritten element. Both run outside your session.

Steps: D3a-01, D3a-02 ... one per element group, then as further D3a legs the unattached propose lines and the verdicts with the uncomposed rooms; D3b the re-check, if you build one; D3c the freeze and the re-read, the handback and the interview. Each leg is named by its step (STANDING §2).

Safe stops: after each element group is ruled and logged. 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 D3 --pause (STANDING §2).

Startup packet, every leg: _head.md, this file, SCHEMA-adaptations.md, the last five diary entries, validate.py --open D3; the rooms-x-elements view's summary for the group.

Reads allowed: mem.py adaptations --element <id>; the room's model and coder documents by query when a line needs them (mem.py room, forest, theme, codes, cite.py); verdicts-D.jsonl; findings.jsonl by --touching; prior.md; the last story_rooms-pN.jsonl. Forbidden: the raw ledger as a list; loosening an element until a room fits (rule 12); promoting a case-only element to the global model without a shared warrant; EXPECTATIONS.jsonl (stage E1 scores it); the canvas as an undeclared ground (rule 7).

Outputs: adjudications.jsonl; model/ revisions with history.jsonl lines; tier_history.jsonl for any tier move; when the third-order tier is recut, scheme_v2.jsonl and assignments_final.jsonl rewritten (mem.py mint, then mem.py reassign, as stage C2 says); handoff-D/recheck/, adaptations-recheck.jsonl; uncomposed-D.md; findings.jsonl lines; verdicts-D.jsonl rulings; exception-report-D.md; notes-D3.md; the model frozen as one version, hashed; handback-D3.md.

1. By element, never as a list

Read _views/rooms-x-elements.tsv first: where the mass is. Then one element group per leg: mem.py adaptations --element P07 returns every accepted-or-pending line touching P07, sorted by op. For each line: accept, modify (with revised, which passes the same portability check as the labeler's line; the original stays in the ledger), reject, or particular, with one reason from the line's evidence and the room's material, into adjudications.jsonl through mem.py adjudicate (SCHEMA §6), with pair where a second room's accepted line meets the same correction and kind where you reclassify. A portable line with no shared warrant is a locally supported particular if its evidence suffices, otherwise pending and unresolved. A line naming the room itself in condition is a particular; one naming another room in relation is not. Quotes go through cite.py.

2. Corrections met twice

Two accepted corrections with the same target and the same lifted node at the third or fourth level are the trigger: you re-examine the element in the same leg, and your written reason is the warrant (rule 12), never a count. A revision is a history.jsonl line with the move (one of Grodal, Anteby and Holm's eight), the adaptation ids that caused it, a move line in the diary and a snap, made when you revise; substantive alignment warrants a cluster branch or a change of global scope, construct or mechanism. The count does not choose the revision; retain the formulation with a written reason if the alleged alignment does not hold; every finding touching the element is re-examined under rule 14 in the same act, as a new findings.jsonl line carrying the id it supersedes. A tier move is a tier_history.jsonl line under rule 11's amendment regime. absent is ordinary and never a correction unless the model stated the element as universal, which you decide here and log. You do not loosen an element until a room fits.

3. The unattached lines

propose lines touch no element. Read them by lifted node: several rooms' unplaced second-order themes pointing at the same target with no third-order theme holding them are minted as one (mem.py mint, a tier_history line with move splitting or merging, logged and snapped), the new theme written to scheme_v2.jsonl and its members moved with mem.py reassign, and the lines are re-lifted (python3 tools/ingest.py D --relift). A proposal you accept becomes an element with status candidate (since: D3), and the model is rewritten with it; one you reject is a particular with a reason. An element supported only in one room remains in that local model and adaptation; it does not enter the global model solely to fit the case.

4. The verdicts, and the rooms the model cannot compose

Rule the split verdicts with the criterion in front of you, and the cannot-tell lines stage D2 listed as unsupported by the room's numbers: a ruling is a verdicts-D.jsonl line with ruled_by: analyst, changed_from (the readers' verdicts) and the reason; the readers' lines stay. Rule 13 governs; the readers are not judges.

Then uncomposed-D.md: every does-not-compose room read whole through its forest, in its own words (rule 12): what it did, what the model would need in order to say it, whether the model should have that, and the condition that separates it from the rooms that compose. Name the kind, yours and not the reader's: outside-the-model, the room organized and the model cannot say how, the counterexample that bounds the model; out-of-scope, the room did not organize, or did something the model is not about (received the event as news, talked about another room's work, played), which bounds the model's scope and not its claims; too-thin, the room is above prior.md's floor and its record still does not show how it organized, which bounds the record, read with its off-platform pointers beside it. The out-of-scope rooms are also read together, since what they did instead may be a kind of its own.

Local validation precedes a shared claim. Resolve each supported exception in its local model and ledger, including unshared particulars. Shared conditions require substantive comparison of the rooms and a written warrant. Repeated corrections can define a cluster branch; they do not force the same revision everywhere. If no defensible formulation survives, replace or withdraw the claim under rule 14. Record unresolved local evidence explicitly. An exception cannot be printed beside an unchanged universal claim as a substitute for a scope decision. Preserve every pre-adaptation verdict alongside local and cluster judgments. The list is carried to D4, D5 and E1. This is an in-sample return run as a census, in the spirit of Eisenhardt's replication logic: a model that composes everything by being loose is weaker than one that composes most rooms and can say which it cannot and why.

The field's entries. out/sealed/FIELD.jsonl names the tentative entries stage C4 took from the field. Rule them as any element, by the rooms' lines: each becomes standing, bounded or withdrawn under rule 11, with its history.jsonl line, and none stays tentative past the freeze. Report them apart in notes-D3.md under a heading The field's entries, one line per id: how many rooms instantiated, restricted, contradicted and left it absent, and the ruling.

5. The re-check

For every rewritten element, all rooms whose adapted account relies on it go back to their resumed return lead against the revised model. Include positive and contrasting cases and record how affected rooms were identified. Write recheck-scope.json with affected_rooms (a list), unaffected_rooms (a room-to-reason mapping) and reason for the scope. These sets must partition the analysed corpus. If the scope is uncertain, recheck all rooms. Every affected room needs a revised local account with its local status; unresolved cases stay explicit. Do not freeze a newer model hash than the one returned to the affected cases. A three-room sample may pilot the revision; it does not certify untouched affected accounts. Preserve both versions. First write MODEL-D3.md, the revised model as prose with ids (the re-check never reads the prose written before the revision); then handoff.py D --recheck <sub>,<sub> builds handoff-D/recheck/ with the revised vocabulary, VERSION and model, and the verdict packets for the same rooms in verdict-D/recheck/. Close the leg with done.sh D3 --cut packets-out. The handler runs the labelers under the same brief into DROPS/D/recheck/, and dispatch.sh D3 sends the verdict packets to the verdict fleet (_fleet_raw/D3v/); when both have returned, the harness runs ingest.py D --recheck, which compiles adaptations-recheck.jsonl (ids R-<sub>-<content-and-version-digest>-NNN) and verdicts-recheck-D.jsonl, and opens D3b. You read the before-and-after ledgers of those rooms and say in notes-D3.md whether the rewrite fixed what the evidence pointed at; E1 reports them. Both verdict lines stand, with the version on each.

6. The close

Pending lines retain their unresolved status and are counted separately from supported particulars. A locally validated account cannot contain pending claims. Adjudicate every original and recheck line; a completed recheck never silently accepts its lines. exception-report-D.md: per room, particulars against accepted shared lines; per element, rooms adapting it by op; the composition criteria in prior.md. Report fixed-global verdicts, local validation status, particulars and corrections met twice separately. Repeated corrections are candidates for clusters, never a failure threshold. Count distinct shared changes as well as ledger lines; assess cluster adequacy at D4 to D5. Where all adaptations remain unrelated particulars, say: a model whose adaptations are all particulars has not composed. The model at the close is frozen as one version: done.sh records its hash at D3c's close (out/_closes/model-D3.sha256), and the hash goes into RUN.json; python3 tools/ingest.py D --relift then re-lifts every ledger line to it (the check at D3c and at D4 refuses a ledger that is not re-lifted); every ledger line is re-lifted to it before stage D4, and the account, the figure and the data structure are re-read against every amendment (rule 11). Write the re-read into notes-D3.md under a heading Re-read against the amendments, one line per amendment: its id, what in the account, the figure or the data structure it changed, or "nothing".

The counts, last. After every ruling is written and not before, validate.py D3 computes and you read: the four verdicts' counts over the analysed set, and does-not-compose by kind; cannot-tell by threshold; the rooms stage C2's last story could not tell, how many of them compose now, how many do not, and whether any room that story told does not compose here, the measure of how well the loop prepared for this test; the composition rate of the rooms read against a story at C2 beside the rate of the rooms never read against one, from story_rooms-pN.jsonl and verdicts-D.jsonl, the nearest the run comes to measuring what the stories were worth. Report them in notes-D3.md. No ruling is revisited because of a count: a bad number is a fact to report, not a thing to fix.

Moments to log

Each element group opened and closed with its counts by decision; each revision with its adaptation ids and move; each minting and candidate element; each verdict ruled with changed_from; each uncomposed room with its kind and condition; each finding bounded, and each condition left as one room's description; the re-check rooms named; the composition criteria; the counts read.

Final message

The claim (what the return taught the model), the room that best supports it, the room that most threatens it, the condition that separates them; then: lines by decision, revisions and their causes, elements bounded, minted or made candidate, re-check rooms and what the rewrite fixed, the diagnostics against prior.md, verdicts ruled, the four verdicts and the kinds of does-not-compose, the bounds with their second room and the conditions still one room's description, residue kinds named, how the last story's misses fared and the two composition rates, what the return showed that the room notes did not predict; the licences exercised (METHOD.md). A return value.

Close

done.sh D3 at every leg, every revision logged and snapped as it was made; the close-check report into notes-D3.md; handback-D3.md; the head; the closing diary line.

Interview point. D3c writes out/interview/D3c.md (under 800 words: what you did, what you found, what surprised you, what you doubt, and the three questions you would most like to be asked) and closes with done.sh D3 --cut interview: the model returned to every room and adjudicated: where it composed, where it did not, what the corrections changed. The principal investigator may talk with you here, in a fork of this leg, run in a copy of your folder as it stood at the close (not read-only; nothing done there reaches the run); the run goes on from the cut either way. done.sh files the brief under the leg, the stage, the step and the minute, and refuses the cut without the brief (and, at a stage's end, without the handback).