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

The adaptation schema

Everything downstream of the return (adjudication, the taxonomy, the branch descriptions, the reconstruction test, the element views) works only if an adaptation is a typed operation on a named element of the global model, with evidence, and not a paragraph. This document fixes the types. It is deliberately small: five element kinds, seven operations, one ledger line, one adjudication line, and the files derived from them.

1. Elements of the global model

The model the analyst builds at stages C2 to C5 is stored in model/elements.jsonl, one line per element, and model/arrows.jsonl. Five kinds:

Element line:

{"id": "P07", "kind": "process", "statement": "...", "tier": "T4-003", "dimension": "D2",
 "evidence": {"findings": ["F-031"], "codes": ["asoiaf:c0042", "denmark:c0117"]},
 "since": "C2-pass-3", "status": "standing"}

tier links the element to the fourth-order theme it belongs to; dimension to the aggregate dimension. status is standing, bounded, withdrawn or candidate (a proposed element accepted at D3 and not yet instantiated in a second room); a bounded element carries the condition ids that bound it. Every change to an element is a line in model/history.jsonl with the move (one of Grodal, Anteby and Holm's eight) and the leg. Ids are never reused; a withdrawn id stays withdrawn.

The id vocabulary the labelers receive at the return (D1 to D3) is the projection of elements.jsonl and arrows.jsonl at the model version hash written into the handoff: for an element id, kind, statement, tier and dimension; for an arrow id, kind, from, to and mechanism (and its statement where it has one); and status: tentative on an entry stage C4 took from the field. Nothing else.

2. Operations

Seven operations. The labeler chooses the operation; the kind (specification, elaboration, correction) follows from it and is not chosen. The aggregator may reclassify at D3, and the reclassification is logged.

The operations are this instrument's operationalization of Fisher and Aguinis (2017): instantiate and absent are vertical contrasting (the theory applied at another level of detail); add-condition and restrict are new construct specification of a condition; add-step is construct splitting; link is structuring specific and sequence relations; contradict and propose are horizontal contrasting. Busse, Kach and Wagner (2017) motivate exploring boundaries through construct refinement, moderators and mediators. Our two correction codes distinguish discrepancies, not mandatory outcomes: adjudication may bound, refine, extend or replace a formulation.

Source and target nodes . Every operation names a target in the model (an element id, or a dimension id for link) and, where it adds anything, a source in the room's own ladder: a second-order theme id of the room (denmark:t07) or a code id. add-condition and restrict give the condition as a source theme id; add-step gives the finer process as a source theme id; link connects a source theme to a target element or dimension; instantiate and absent name the target and cite codes. Free text lives only in how. Two labelers can no longer describe one adaptation in two vocabularies, because they point at nodes.

At ingest each source node is lifted through assignments_final.jsonl as it stands at the model version the handoff carries (stage C1's placements, re-decided by every logged recut since): second-order theme → third-order theme → fourth-order theme → dimension. The lifted path is stored on the line ("lift": ["denmark:t07", "T3-014", "T4-003", "D2"], with null at a level the theme was never placed at). An unplaced second-order theme lifts to nothing above itself; those are the lines from which D3 mints new third-order themes when several rooms' unplaced themes point at the same target.

An eighth form, propose, is not an operation on an element: it names something the room needed that no element covers. It carries a statement and evidence and is unattached until D3, where it may become a shared candidate element or a locally validated particular. Unshared does not mean false.

Relational adaptations use the same operations with a relation field naming the other room ("relation": "placenl"), so "defence delegated to a named neighbour" is instantiate on the defence process with a relation. Naming another room in relation does not make the line a particular; naming the room itself in condition does (§4).

3. The ledger line

adaptations.jsonl, one line per adaptation, written by ingest.py from the labelers' returns, never by hand:

{"id": "denmark-007", "sub": "denmark", "op": "add-condition", "element": "P07",
 "source": "denmark:t07", "lift": ["denmark:t07", "T3-014", "T4-003", "D2"],
 "condition": "the artwork is under attack from two sides at once",
 "statement": null, "how": "explains why denmark recruited outside placers on night two and not before",
 "evidence": {"codes": ["denmark:c0117", "denmark:c0118"], "transitions": ["denmark:t-C5"]},
 "relation": null, "portable": true, "kind": "elaboration",
 "origin": "labeler", "model_version": "sha256:…", "labeler_run": "codex-2026-11-…"}

Required for every line: sub, op, element (or propose with statement), at least one code id in evidence.codes, how as one sentence, model_version. source (a theme or code id of the room) is required for every op except absent, and for contradict it may be a code id rather than a theme; for instantiate it is the theme that shows the element operating, so that instantiations carry a key and enter the distance as specifications; condition and statement are the source theme's label copied for reading, not free text, except that contradict and propose may state in words what no theme names; from, to for link are source and target. kind, portable and lift are computed at ingest.

A room with nothing to adapt returns one line: {"sub": ..., "op": "none", "how": "the global model explains this room as stated", "evidence": {"codes": [...]}}, with evidence for the fit. Silence is a missing return, not a verdict.

4. Validation at ingest

ingest.py D accepts a room's return only if every line passes:

1. Schema: the fields in §3, the op in §2, no extra fields.
2. Element ids exist in the model at model_version, and are not withdrawn. A return against a stale version is rejected whole and reissued with the current packet.
3. Code ids exist in the room's codes.jsonl in the store, and are not retracted. A transition id exists in the room's forest.
4. how is one sentence (one terminal stop; under 40 words).
5. Portability: condition and statement contain no subreddit name from the census and no proper name from the room's lines in the analyst's lexicon-A1.md, except in relation. A line that fails is kept with portable: false and is a particular by construction (contract rule 12).
6. The return carries the end marker from RETURN-D.md and the room's line count matches its header.

Rejected returns go back through the handler with the failing rule named. Completion comes before any limit: the handler returns every room it can, and marks a room unreturned only when it cannot. Counts of accepted, reissued and unreturned rooms go into RUN.json at the close of D2.

5. The room model

rooms/<sub>/model-D.md, written by the labeler beside the ledger, is case.md rewritten in the model's vocabulary. Its sections are the aggregate dimensions, in the model's order. Inside a section the labeler writes prose, and every claim carries either an element id in brackets ([P07]) or a [new: …] marker for what no element covers, and a code citation ([c0117]) from the room's own codes. The file ends with a grade the labeler writes, composes, composes-with-residue or does-not-compose, and one sentence of why. The grade is the labeler's; the verdict, with its fourth value cannot-tell, is the fleet's (FLEET.md), and the two are reported side by side. The lead's grade concerns the adapted local explanation; the fleet verdict concerns the fixed global model. The return also requires local_status: validated | unresolved | insufficient-evidence and a ## Local exceptions section linking discrepancies to evidence and adaptations. Local validation is distinct from portability.

ingest.py checks the room model for the same id and code validity as the ledger, and that every adaptation line's element appears in the room model at least once.

6. Adjudication

adjudications.jsonl, one line per adaptation, written by the analyst at D3 through mem.py adjudicate:

{"adaptation": "denmark-007", "decision": "accept", "kind": "elaboration",
 "reason": "the condition is stated without the room, and asoiaf-011 meets it",
 "pair": "asoiaf-011", "change": "h-0412", "leg": "D3-P07"}

decision is accept, modify, reject, particular or pending. modify carries the rewritten condition or statement in a revised field, which passes the same portability check as the labeler's line, and the original stays in the ledger. A supported portable line with no shared warrant is particular. A line with insufficient evidence remains pending and unresolved at D3 close; pending lines are counted as such. pair names the second room whose accepted line meets the same candidate correction. Recurrence prompts a reasoned shared-warrant assessment, not automatic model revision; change names the logged revision or explicit scope decision when one is made. kind here is the analyst's, and differs from the ledger's when the analyst reclassifies.

D3 reads by element: mem.py adaptations --element P07 returns every accepted-or-pending line touching P07 across rooms, sorted by op. One leg per element group, one hundred lines per packet at most. After the elements, the propose lines: each becomes a candidate element (an elements.jsonl line with since: D3 and status candidate) or a particular.

7. What is built from the ledger

All generated by gen_views.py at leg close; none hand-edited.

• rooms-x-elements.tsv: rows rooms, columns elements, cell the op or blank. The view D3 reads before any element.
• rooms-x-adaptations.tsv: one matrix per ladder level and per kind, plus the combined one. A key at level L is op:target:<node at L> from the line's lift, for accepted and modified lines at the model version frozen at D3's close (every line is re-lifted to that version before stage D4; rooms whose elements changed at D3 were re-checked). Particulars and pending lines are excluded. The distance, written once so nobody re-derives it:
  • For rooms i and j, let T be the set of targets touched by either room's lines. For each target t in T, let L be the deepest ladder level at which some line of i and some line of j on t have the same lifted node (3, 4, dimension, or the bare target), and let w(L) be 4, 2, 1, 0.5; if only one room touches t, the term is 0.
  • Let g be the operation agreement over the two lines that met at L: 1 for the same op, 0.5 for ops of the same kind, 0 for different kinds, and -1 for instantiate against absent or add-condition against contradict on that target. Where several pairs of lines meet at L, take the pair with the largest g.
  • Similarity S(i, j) = sum over t in T of w(L_t) times g_t, divided by 4 times |T|, so S lies in [-1, 1]. Dividing by |T| removes the mass effect: a room with more lines gains no closeness from having more. Distance d = (1 - S) / 2: 0 for identical fully comparable adaptation profiles, 0.5 for rooms that share no target (unrelated), 1 for rooms whose lines oppose each other everywhere. Rooms whose line sets are both empty (none) have d = 0 with each other and 0.5 with every adapted room. Explicit observed-absent or not-needed lines on the same target contribute the deepest-level same-op weight without requiring a source lift. Unobserved and not-testable lines remain in local accounts but are excluded from shared distances. Missing lift entries never become shared None keys; partially comparable non-absence profiles can retain nonzero distance even when their raw sets match. tools/distance.py computes it.
  • Per-kind matrices use only lines of that kind; the tree-likeness statistic (Gray, Bryant and Greenhill 2010) is computed on the combined level-3 matrix. The weights and grades are declared in prior.md before stage D4 opens and are not changed after.
• exception-report-D.md: per room, particulars against accepted shared lines; per element, the count of rooms adapting it by op; the sentence the stage file requires when all of a room's adaptations are particulars.
• forms/<id>.md at stage D4: a branch as the set of shared keys, its members, and the analyst's description of what the branch does differently.

8. Reconstruction

rebuild.py <sub> uses the effective adjudicated ledger. It includes accepted, modified and supported particular lines in local reconstruction. It lists pending lines as unresolved, never asserted support. The latest accepted room return replaces that room's earlier ledger even if its new ids list is empty. Original ledgers and model versions remain available.

Recheck ids include a digest of the line content and model version: R-<sub>-<digest>-NNN. An unchanged return is idempotent; a changed return cannot inherit an earlier line's adjudication. Prior recheck rows remain in the ledger and the latest accepted manifest selects current ids. Superseded compile files and local account copies remain under return-history/.

adaptation-returns.jsonl records each accepted return with sub, phase (initial or recheck), status: accepted, model_version, ids, local_status and local_grade. An old local return carries forward to a changed model only when D3 explicitly marks the room unaffected with a reason; all affected cases return to their original leads. The skeleton reports the current model hash, return version and local version status.

An absent line produces no only for observed-absent. not-needed, unobserved and not-testable retain their own meanings. Legacy unspecified absence is conservatively unobserved. Unknown states never enter shared adaptation distances as negative evidence.

At E1 a fresh reader receives the skeleton and the coder's blind case.md with the room's second-order themes, never model-D.md. Score all warranted claims, record evidence limits, and report reconstruction fidelity by branch and by historical global verdict. This is an in-sample preservation and composition check, not independent truth estimation.

9. The handoff folder (stage D1)

out/handoff-D/ written by handoff.py D:

• RETURN-D.md: the return specification, verbatim in every task, the schema of §3 and §5, the end marker, the reissue rule.
• VOCABULARY.json: the id vocabulary at the model version hash.
• MODEL.md: the global model as prose with ids, the account from stage C5 cut to the ceiling the labeler's context allows, plus the wiki pages for the elements the room's second-order themes were placed under.
• LADDER.md: the coding table, quote to aggregate dimension, with representative codes for every tier the model uses, projected for the dimensions and elements in MODEL.md .
• tasks/<sub>/TASK.md: the labeler's instruction: read the coder documents, rewrite case.md in the model's vocabulary, write the ledger; the room's case.md, codes.jsonl, themes.jsonl, brief.md, forest.md; the placements of the room's second-order themes; the two outputs and where they go.
• HANDLER.md: the adaptation handler's brief: dispatch order (by size ascending, so failures surface early), how to run a labeler in Codex, the return path DROPS/D/<sub>/, the reissue rule, and what never to do (no access to out/ beyond the folder, no edits to the vocabulary, no merging of rooms).

10. Decisions taken, and their consequences

1. Evidence states are distinct. A room need not use every global element. Only evidenced absence or lack of need supports a negative assertion. Silence and untestability are unresolved evidence, never a contradiction. An evidenced absence challenges a universal claim only after its scope and evidence are examined.
2. Completion over reissue. No reissue limit as a rule; the handler's job is to get every room returned and to check that the returns work, and ingest.py reports what fails and why. A room is unreturned only if the handler says so, and the count is reported.
3. Proposed elements enter the model in the same run. At D3 a propose line that the analyst accepts becomes an element and the model is rewritten with it. Then a re-check: a few of the rooms whose evidence produced the change are sent back through the labeler against the revised model, in a small second handoff (handoff-D/recheck/), and their new returns say whether the rewrite fixed what the evidence pointed at. The re-check rooms and their before-and-after ledgers are reported at 18.
4. The labeler's packet carries the whole ladder. Beside the wiki pages of the elements its themes were placed under, the labeler receives the coding table with representative codes for every tier the model uses: quote, first-order concept, second-order theme, third-order theme, fourth-order theme, aggregate dimension, projected for the dimensions and elements in MODEL.md. A labeler that sees only its own codes cannot understand an aggregate dimension, so it cannot say what the model fails to explain. handoff.py builds this projection from _views/data-structure.tsv.