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

Stage B4. Assignment: adjudication, and transmission after revision.

Model: build and check legs as arm.json sets (your platform file lists them); the blind sample's dispatch outside your session (FLEET.md). Legs: adjudication (one or more, bounded by packets of contested items), then build, dispatch and check for the blind sample.

Steps: B4a-01, B4a-02 ... one per adjudication packet; B4b the blind sample's build (it ends at --cut packets-out), the dispatch, B4c its check, the placements final (assign_final.jsonl), the handback and the interview. Each leg is named by its step (STANDING §2).

Safe stops: after each adjudication packet is ruled and logged, and after each section of the transmission check. 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 B4 --pause (STANDING §2).

Startup packet: _head.md, this file, FLEET.md, the last five diary entries, validate.py --open B4; in the check leg, the harness's one line: the readers left their files in _fleet_raw/B4/.

Reads allowed: assign_raw.jsonl, b3_classes.jsonl (the contested inventory, §1), multiplacement_log.jsonl (every why_not_one), fidelity-B3.md, position-B3.md, notes-B3.md (the disagreement inventory, the gloss-only placements), stage B3's gold keys (the tags in batches-B3.jsonl), author_items.jsonl, gold_items.jsonl, scheme_final.* with families.jsonl, confusions.md and pilot_report.md (stage B2's predictions), stability-B2.md, notes-B2.md (the agreement definition); the contested items through mem.py theme and codes --sub, quotes through cite.py; _fleet_raw/B4/ in the check leg. Forbidden: the canvas as an undeclared ground (rule 7); your notes as a ground for a ruling; a reader's identity while you rule; any edit to the scheme; to a blind reader, everything §10 withholds.

Outputs: disagreements.jsonl; assign_final.jsonl; transmission-B4.md; coder_conflicts.jsonl; adjudication_report.md; RETURN-B4.md; reassign.jsonl; transmission-after-revision-B4.md; notes-B4.md; a manifest entry.

1. What comes here

Only the contested items, and the placements made from a gloss. Three placements decide which items are contested (J, 2026-10-12): the two readers' pass-1 lines, replicates 1 and 2, and your B1 grouping, the item's third-order themes among the members of scheme/scheme_current.jsonl (sha256 8b67360b8d0dfefc), joined by sub and second-order label through package/store/themes.jsonl, a set with no order. A reader's list is its third_order in order. Gold copies and the duplicate's second copy are left out. tools/assemble.py B3 writes each item's class into b3_classes.jsonl; that file is the contested inventory, built by the tool, not by hand.

The placement stands, with no adjudication:

1. The two readers return the same list, order included, and its set equals the B1 grouping. The readers' list stands, multi or not.
2. The two readers return the same list, and its set is smaller than the B1 grouping and inside it. The B1 grouping stands: the readers' list first, then B1's other themes in ascending id order. Its line in assign_final.jsonl carries "source": "b1_grouping_kept", so that stage C1 can tell it from placements made under the frozen scheme.

The item comes here:

3. The readers differ, including the same themes in another order (J: "just add the case to adjudication"); and an item with one replicate missing, which cannot be agreed (missing_replicate in b3_classes.jsonl).
4. The readers agree on a set that is not inside the B1 grouping.
5. The item has no B1 grouping (58 at B1).

One exception stands whatever the class: a placement into a family neither reader opened (families_opened without the chosen family; unopened_family in b3_classes.jsonl) is read by you, since two readers can agree from a gloss; where you move one, it is logged as a ruling like any other. An item two readers holding the same written scheme placed differently is a boundary the scheme did not decide, with the alternatives attached.

2. The gold placements come out first

Drop the seeded gold placements before you adjudicate anything, using stage B3's keys. A seeded placement is measurement, not corpus evidence; left in, it enters assign_final.jsonl and every denominator below, and breaks rule 4 quietly.

What is dropped is the gold placement, not the second-order theme. A theme carrying gold placements is still a theme of the analysed set and still owes assign_final.jsonl its line, taken from the live replicates. Drop by (item, agent, replicate) on the gold-keyed rows, never by key. After the drop the line count still reconciles to counts.py's assignment_items_expected over the analysed set; if it does not, you dropped themes.

Fidelity on the gold placements is stage B3's figure, in fidelity-B3.md. Read it, do not recompute it, and say in notes-B4.md how many placements the drop removed.

3. Transmission, decomposed

Compute it before anything here touches the file. Drop the gold placements from the rates as §2 drops them from assign_final.jsonl, but keep them for the decomposition, which is what your own answers are for. Your answer exists for every gold item and every item stage B2 piloted, in gold_items.jsonl and author_items.jsonl; FLEET.md's four rows are computed there:

The third row is never revised away. The fourth row's items never reach the adjudication, because both readers agreed: they are in assign_raw.jsonl and nowhere else, and author-error below catches only the items that split. Split the fourth row both ways, you were wrong or the scheme taught something you did not intend, and say in notes-B4.md which way each went. Outside the gold and pilot sets your own answer does not exist, so the split rate there cannot be decomposed: say so, and do not extrapolate from the gold set, which was drawn to be hard.

transmission-B4.md gives the rows with their items, by family. Name it in handback-B4.md; stage C1 reads it among its sources.

4. The conflicts pass

coder_conflicts.jsonl is produced here, once, and nowhere else: two coders describing the same practice in incompatible terms, seen only by reading across communities; the second-order themes whose theme_note and concepts pull apart under the scheme. The pass runs over theme_note alone, with no concepts beneath it. Run it before you rule: a conflict between coders explains some of the disagreements between readers, and you want to know which first. Stages C1 and C2 read this file.

5. Declared before the first item, in notes-B4.md

The ruling policy. At R = 2 every split is one reading against another and there is no majority to lean on:

• how many items you expect to rule on, from stage B3's inventory, and how many packets that is at five hundred per leg (v42; one hundred before): a large inventory means more legs, never a rule applied in bulk. A scheme of hundreds splits more than a scheme of forty, mostly between siblings in one family, where the discriminator is the sentence to quote; gloss-only placements are ruled first; and where many splits fall on one sibling pair, the boundary is ruled once, in one paragraph quoting the discriminator, applied to every item on the pair and logged as one ruling with its item list, because a boundary ruled item by item is ruled inconsistently;
• what a reader's confidence and why count for, and whether a clear reading against an arguable one is a prior or nothing;
• whether an adjudicated placement may name a third-order theme neither reader chose: it may, logged loudly, because it means both readers missed something the file did say;
• what you do when the honest answer is that the scheme does not decide. That must be available as an outcome, recorded as unresolved and carried, not resolved by preference.

6. Blindness, and its limits

You see the item, the scheme, and the two placements, not who made them. Agent id, batch, replicate, position and timestamps are stripped from the packet before you look (no v38 tool yet for the stripping; done by hand and logged); the funnel is a property of the fleet, not evidence about this item. Both readings are shown, with their why and confidence: two readings, neither a prior over the other. You wrote the scheme, and no procedure removes your pull toward your own reading; record named_neither every time your placement is one no reader chose.

7. Adjudication, by packet

One packet of at most five hundred contested items per leg (v42), by family, gloss-only placements first. For each item you rule: the taxonomy from stage B2, the placement that stands, and one clause of why from the item's material through mem.py theme and codes, the deciding sentence quoted through cite.py. A ruling never cites a room note. The classification matters more than the placement:

• scheme-ambiguous: the belonging test or the discriminator does not decide. Name the field and quote the sentence. These aggregate into stage C1's revision list.
• item-is-a-bag: the coder's second-order theme holds two practices, so no single placement is right. Not a scheme defect.
• reader-error: the file does decide and a reader missed it. Suspect this last, and never assign it when two independent readers made the same miss: that is a scheme defect that looks like reader error.
• author-error: you were wrong and the readers were right. It exists wherever your own answer is on record, the gold items and anything you placed at stage B2; an adjudication that never returns one is measuring obedience. Carry the count to adjudication_report.md.
• genuine-joint: the material satisfies two gates and one placement cannot carry it, rule 0's exception on the record. A split here is readers applying rule 0 differently; the count measures how well the rule travelled, a fact about the instructions, not the readers.

Where the two readers agree and you would not have, you may overrule with the same line and the taxonomy value author-disagrees, counted separately; you do not re-read the agreed items in search of these.

{"sub": "...", "theme_id": "...", "readers": [["T3-007"], ["T3-011"]], "ruling": ["T3-007"],
 "unresolved": false, "taxonomy": "scheme-ambiguous", "field": "must_not_show",
 "theme": "T3-007", "quote": "the sentence that failed to decide",
 "boundary_ruling": "T3-007|T3-011 or null",
 "named_neither": false, "confidence": "clear|arguable", "why": "...", "leg": "...",
 "support_by_theme": {"T3-007": ["c0013"]}, "also_present": {}}

Every ruling names the concepts behind each theme it places (support_by_theme, rule 0 as amended, J 2026-10-11); a bag keeps the themes it does not place under also_present, each with its concepts. assign_final.jsonl carries both from the ruling or from the readers' lines.

ruling is empty and unresolved true where the scheme does not decide; a boundary ruled once carries the pair in boundary_ruling on every item it decided, and the ruling itself is one paragraph in adjudication_report.md with its item list. assign_final.jsonl is every item of the analysed set with its standing placement.

8. Three things not to do

Do not let a clear reading decide silently. A reader's confidence may weigh on the placement; the disagreement still gets its row, its taxonomy, and its quoted sentence.

Do not revise the scheme here. Stage C1 does that with the whole inventory in front of it; fixing a sentence mid-adjudication adjudicates the items before and after it against different files. Log the defect and leave the file alone.

Do not let the unresolved count reach zero by effort. An adjudication that resolves everything has resolved some things by preference.

9. adjudication_report.md

• Items adjudicated, against stage B3's disagreement inventory.
• The taxonomy split, the most important result: mostly item-is-a-bag is a good scheme meeting messy material; mostly scheme-ambiguous is a scheme that did not survive being handed over.
• Third-order themes by ambiguity, ranked. Stage C1 opens with this list.
• Fields by ambiguity: must_show, must_not_show, decide_when_both or the discriminators. Four numbers.
• The named_neither rate, and your reading of it.
• The unresolved count, named, with the author-error and author-disagrees counts.
• The boundary rulings, one paragraph each with its item list.
• Agreement before and after, with the note that the figure after is not a reliability figure; it is agreement following an adjudication you performed.
• Whether stage B2's predictions held, now that every theme in the analysed set has answered.

10. The blind sample: build leg

This measures transmission after revision (FLEET.md), never called reliability: whether the line you settled can be reached again by readers who were not here. Fleet agreement is already computed, corpus-wide, at stage B3, before anyone decided; this adds fresh readers against the record every later stage rests on. It is small by design: the fleet already read everything twice.

Declared before you draw, in notes-B4.md:

• The sample: five hundred items, two hundred and fifty drawn uniformly at random from the analysed set and two hundred and fifty from the adjudicated items, by a recorded seed (J, 2026-10-12: "250 each for adjudication and random"). The two strata are reported separately and never pooled: the first says whether the scheme decides the ordinary item, the second whether your rulings can be reached by a stranger. Within the uniform stratum, the items kept under §1's rule 2 are reported as a row of their own: they were neither adjudicated nor placed under the frozen scheme.
• Two fresh readers per item, as arm.json sets this stage's readers, with no memory of the run, one dispatch. Each reading is compared with the settled line on its own: one thousand comparisons, a rate over readings; the readers' agreement with each other is reported beside it.
• The agreement definition, stage B2's and B3's from notes-B2.md, unchanged, or the figures cannot be set beside each other: sets by Jaccard where rule 0 allowed more than one placement, alpha computed on it; whether ["T3-004"] against ["T3-004","T3-011"] is agreement, disagreement or partial; whether chance correction is over themes, over sets, or over presence/absence per theme.
• Exemplars: none unless stage B3 served them, and then the same arm, declared on both sides of the comparison (FLEET.md). This measures whether the written scheme decides.

RETURN-B4.md as at B3: the return line, rule 0 in the contract's words, the end marker and the confirm line. packets.py B4 builds the packets under the 50 KB ceiling: the scheme as frozen (the index page, the family pages as the reader opens them), the sampled second-order themes with their theme_note and concepts, and no ruling in sight. Nothing else. Not assign_final.jsonl, not disagreements.jsonl, not gold_items.jsonl, not a hint that these items were decided before, and never your room notes, the head or findings.jsonl. If a reader needs your notes to decide, the scheme did not decide, and that is the finding. Append the manifest entry on stage A1's schema (readers' model, seed, decoding, reader count) before the first packet leaves. Stop here. Close with done.sh B4 --cut packets-out. Your session ends; nothing more of this stage runs on your model until the check leg. The dispatch runs without you, on the platform's reader route. The next launch of stage B4 is the check leg, and its packet tells you one thing: the readers left their files in _fleet_raw/B4/.

11. Dispatch leg

Not yours. The dispatch sends each packet to two fresh readers on the routine reader model, never the same reader for both readings of an item, collects the returns into _fleet_raw/B4/, validates them against RETURN-B4.md, reissues what fails the schema, and closes. Nothing in it reads your files.

12. Check leg: the comparison and the figures

You are relaunched. Assemble the returns keyed on packet heading and cell number. The comparison, per reading, the fresh placement against the settled one, into reassign.jsonl:

• agrees: the scheme decides, and the same way twice;
• disagrees, and the item was a disagreement here: expected. Compare the two rates; a boundary that split the fleet and then split a fresh reader shows consistency, not a new problem;
• disagrees, and the item was unanimous in the fleet: either the fleet shared an assumption the scheme does not license, or the adjudication moved something it should not have. Name these individually, and give each a paragraph;
• disagrees, and the settled placement was unresolved: no information. Exclude these from the denominator rather than counting them as a failure, and say how many.

transmission-after-revision-B4.md: the headline figures side by side, fleet agreement from stage B3 and fresh-against-settled agreement per stratum, with alpha per stratum on its two hundred and fifty items and the stability floor from stability-B2.md beside them, the uniform stratum's rule-2 items in a row of their own, the definition stated once above them; per family, not per theme: two hundred and fifty items a stratum over nearly two hundred themes leave most themes with one or two items, or none, in the sample. The per-theme figure stage C1 needs is stage B3's agreement, over every item and every theme; point to it, do not recompute it. Then the unanimous-but-disagreed items, listed; and the gap between the uniform stratum's figure and stage B2's gold figure: the pilot measured the scheme on deliberately hard items, and where the gold number is not the lower of the two, the hard set was not hard. Wherever a figure from this section appears, say what it licenses, in FLEET.md's words: readers share an architecture with you, so the figure bounds whether the written instrument decides, never whether it is right.

Moments to log

The gold drop with its count; the transmission rows; the conflicts pass; each declaration of §5; each packet ruled with its taxonomy counts; each boundary ruled once, with its item list; each overrule; the blind sample's draw with its seed, its build, its dispatch and its result; every unanimous item a fresh reader placed elsewhere.

Final message

The claim (which family the scheme decides worst and why), the item that best supports it, the item that most threatens it, and the condition that separates them; then the counts: contested items inherited and adjudicated; rulings by taxonomy; author-error and author-disagrees; the three third-order themes with the most scheme-ambiguous rows; the field that failed most; the named_neither and unresolved counts; boundaries ruled once, and the items they decided; whether the pilot's predictions held on the full corpus; the four transmission rows; transmission after revision per stratum with its definition, alpha against the stability floor, the three worst families, and the unanimous items a fresh reader placed elsewhere; the single boundary you most expect stage C1 to have to move; the licences exercised.

Close

done.sh B4 at every leg; paste validate.py B4's report into notes-B4.md; handback-B4.md; the head; the closing diary line.

Interview point. B4c writes out/interview/B4c.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 B4 --cut interview: the placements final after the fleet placed every theme and the adjudication ruled: where the readers and you disagreed, what the pilot and the fleet taught you. 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).