# Head, claude-v43-1

Under 3,000 words, rewritten whole at every close. Last written at B4c, 2026-10-13 (UTC). Stage B4 is complete: 3,288 contested items ruled, the blind sample built and checked, `handback-B4.md` and the interview brief written. The placements final are `assign_final.jsonl` at sha256 d3506a88…, unchanged by the check.

## The method question, in one sentence

Can independently reconstructed case histories be composed into a shared dynamic account that supports explanations at successive levels of specificity, from population-wide patterns, through forms and subforms, to particular organizational histories, while keeping access to the evidence from which those explanations were developed?

The theory question is unchanged: when strangers organize on a clock they did not set, with no way to compel anyone and no goal given to them, what forms does organizing take, how do they differ in their functioning, and which of them keep the work alive?

## Where the run stands

Stage A1 is closed. Stage B1 (the third-order scheme, 197 themes in 21 families), B2 (the pilot) and B3 (the fleet) are done. Stage B4 is done: the adjudication (B4a-01 to B4a-07), the blind sample's build (B4b to B4b-p4) and its check (B4c), closed at `--cut interview`. No later stage is delivered, and none is begun.

## J's rulings on how a ruling is made (13 October; carry them)

- **QB4-001 (J, 06:43 Berlin):** "Ask it to stop. The point is to try to adjudicate. I imagine that the additions are really just the analyst trying to be nice. And not go back on what others said".
- **J's rulings (06:45 and 06:47 Berlin):** (1) A ruling chooses among the placements (reading A, reading B, my B1) and adds no theme beyond them. (2) It takes the union only when it names, for each theme it keeps, the separate concept that carries it; otherwise it chooses. (3) Re-rule packets 1 to 4's rulings that held a theme no placement held. Done at B4a-05.
- **also_present (J, 09:09 Berlin) and QB4-002 (09:20):** `also_present` lists only themes a reader placed, a reader listed under its own `also_present`, or my B1 named.
- **QB4-003 (J, 11:29 Berlin):** "count them. Themes a reader listed under its own also_present count in the bag test."
- **Lone B1 themes (J, 12:28 Berlin):** "Whatever the agent feels more comfortable with." I chose: in class 4 a theme only my B1 placed stands against the two agreeing readers only with three concepts or more of its own; with fewer it is left out; the readers' list stands where its test takes concepts, and where it takes none the item is unresolved. Class 3 unchanged.

**How I applied them (rule 0 inside the placements).** A concept counts toward a placed theme when that theme's `must_show` takes it and no field sends it to another placed theme. Order by concept count, ties by placements; at most three; a one- or two-concept theme is kept only when it leads its level on placements; a tie at the top places all tied themes; a theme alone at its level is kept; a theme both readers placed is not dropped on a count only my B1 tips. **The bag test:** placed themes and the readers' own `also_present` themes, each with the concepts its `must_show` takes that no placed theme holds; four with three concepts make a bag. Unresolved where no placement's test takes the material.

## Stage B4

**The selection (J's amended stage file).** Class 1 (readers agree, equal to B1) 2,566 and class 2 (a proper subset of B1) 1,990 stand. Class 3 (readers differ) 2,393, class 4 (readers agree outside B1) 837 and class 5 (no B1 grouping) 58 were contested: 3,288 items, ruled in seven packets (`out/_adjudication-B4/`, chunks of 25; `coder_conflicts.jsonl` CC-001 to CC-072).

**The adjudication** (`adjudication_report.md`, full): 3,288 rulings; genuine-joint 1,651, author-error 571, scheme-ambiguous 451, item-is-a-bag 255, author-disagrees 208, reader-error 152; unresolved 185 (5.6%); named_neither 757 (23.0%); agreement before 0.491, after 0.646, not a reliability figure. Fields: must_show 252, discriminators 114, must_not_show 57, decide_when_both 28. Worst themes: T3-017 55, T3-033 34, T3-001 32, T3-023 26, T3-004 26. Boundaries ruled once: 64 pairs, 817 items. Class 4 (837): the readers' list 584, a B1 theme kept 162 (each with three concepts or more), unresolved 78. B2's predictions: 17 of 19 pairs split readers; pairs 17 and 18 never. Transmission rows (gold; pilot): 21, 30; 1, 0; 6, 12; 12, 8.

**The blind sample (`notes-B4.md`, from "The blind sample, built at B4b").** J's rulings: 500 items, 250 uniform and 250 from the adjudicated items, by a recorded seed, the strata apart and never pooled; stratum R holds all 185 unresolved items (J, 17:10 and 17:21 Berlin: "We should have 685 in total"), so its 18 also in U or A are read twice; arm 1 two fresh DeepSeek V4.1 Flash readers per item on the route `arm.json` sets for B4; arm 2 OpenAI gpt-6-astra at medium effort, two replications, by the Codex orchestrator on J's computer; never pooled; no Claude reader.
- **The build.** Seed 20261013, streams "20261013:U", ":A", ":order", ":R", ":R-order", ":R-all-order", ":R-ua-order". `sample-B4.jsonl` (500) and `sample-B4-R.jsonl` (185). 48 packets, `S05-01` to `S05-48`, in fleet ids on B2's renderer: 685 lines due, 667 distinct items. `RETURN-B4.md` as at B3, rule 0 in the contract's current words. `run_manifest.json` entry 16.
- **The check (B4c; `transmission-after-revision-B4.md`).** Readings: arm 1 999 (S05-23 r8 typed its room wrong, refused), arm 2 1,000, R 740 (`reassign.jsonl`, `reassign-openai-astra.jsonl`, `readings-R-B4.jsonl`). Fresh against settled, mean Jaccard: arm 1 U 0.657, A 0.544; arm 2 U 0.622, A 0.554. Readers' agreement and alpha: arm 1 U 0.781/0.776, A 0.664/0.658; arm 2 U 0.872/0.869, A 0.804/0.801; B3 on the same items U 0.781, A 0.514. Arm 1 U by class: class 1 0.930; class 2 (the line keeps my B1 grouping) 0.491, against its placed themes 0.786; U's adjudicated items 0.517; U without class 2 0.703. B2's gold figure (0.624) is the lower. G14 is among the three worst families in both arms and strata. R: readers 0.623 and 0.862 (B3 0.543); 104 and 134 of 370 readings wholly outside every earlier placement. Arm 2's replications agree within a packet far more than across packets (0.653 on the 18 read twice), so its agreement is an upper bound.
- **The 79 unanimous items a fresh reader placed elsewhere** (28 class 1, 51 class 2), each with a paragraph: the fleet's shared assumption 25, the settled line keeps my B1 themes 22, rule 0 applied differently 24, the scheme does not decide 6, the fresh reading misses a field 2. In most of the 25 the fleet counted concepts under a theme whose own `must_not_show` names the sibling.
- **The handback** (`handback-B4.md`): G14 decided worst (warframe t11 supports it, witcher t05 threatens it; a request between rooms separates them); for stage C1, T3-017|T3-031 (the canvas condition) as the boundary most likely to move, the homeland flag, G14's request siblings and the class 2 settled line.

**Boundaries ruled, to carry** (`out/_adjudication-B4/boundaries.md`).
- **Packets 1 to 6.** T3-023 is a change to the sub, T3-001 members' fakes. Calls toward T3-002, placing toward T3-038, a new mark toward T3-033, a further piece toward T3-041. An attack answered as a fight T3-005, watching or fixing T3-021. A fight, threat, peace or truce in the words T3-149; lines and space with none T3-010. Destroying T3-006, taking space T3-061. A joke with the canvas in the words T3-031; with none T3-017. A taunt at a rival T3-050. The homeland flag's acts toward T3-065 where a placement holds it.
- **Packet 7.** Thanks to makers T3-043, congratulations T3-009, an ally thanked T3-072. An account T3-073, grief T3-029, the canvas's end T3-117. A record T3-105 (cleaned copies included), a remake T3-057, the sub's banner or flair T3-156. A later canvas T3-054, the canvas sub's future T3-111. "The work" in T3-073 and T3-058 is the room's own.

## Stage B3 and B2, in short

B3: fleet ids only (QB3-001, QB3-002); 1,212 batches, R = 2; `assign_raw.jsonl` 17,653 lines and `b3_classes.jsonl`. Gold fidelity 0.645; agreement Jaccard 0.786, alpha 0.782; G18 agreed least; readers deepseek/deepseek-v4.1-flash, medium effort, max_tokens 128,000. B2: gold fidelity 0.624 and pairwise 0.818; stability 0.807 is the floor every later figure is read against.

## J's rulings that bind (carry them)

- **Bags.** No one cuts the bags (J, 12:24 Berlin). A bag stays one item: one theme placed, its others under `also_present`, each with three concepts of its own and named by a reader or my B1. `out/recut/WITHDRAWN.json` keeps the withdrawn split out of every view.
- **The bound (J, 13:47).** `also_present` names "twenty themes at most", each with three concepts of its own.
- **The not-flagged bag (J, QB2-013).** With no other theme at the three-concept floor, `item_is_a_bag` is false and `also_present` `{}`.
- **Fleet ids (J, 09:14 and 12:15).** Packets carry only fleet ids; returns are translated back before scoring; my files keep room ids.
- **Codes by item (J, QB2-009).** Each MANIFEST entry carries `codes_by_item`.

## Standing words from J and the PI (carry into every head)

**J's words for B1 (7 October 2026):** "national elements are very important here"; "we need to get as much variance on organizing behavior early on so that later we have enough to create the model". National, political and conflict content is study material, named as plainly as any other. **The handler's rule:** a choice to leave out or soften content for any reason other than the study's goes to J first with `mem.py ask`, and the content stays meanwhile.

**J on reading (11 October, 16:54 Berlin).** Stop cutting room reads to a line length. Read room lines whole. Page a long read by line ranges or offset, never by truncating lines. Use `--width N` on a `mem.py` read if a clip is ever needed; the guard refuses `cut -c` after `mem.py`.

**J on asking (11 October, 17:01 Berlin).** Ask J with `mem.py ask` whenever a question is needed; asking does not stop the run.

**The PI's clarifications (8 October, approved by J):** the thresholds are the design's heuristic; the rule is unchanged.

**How to work, from J (8 October, 23:45 Berlin, "can you do a hot change for the batching?"; relayed by the handler at 22:03Z; carry this paragraph into every head).** A change in how I work, not in what I read or decide. Fewer turns. When two or more tool calls do not depend on each other's output, make them in one turn. Chain ordered commands in one Bash call with &&, so that a failure stops the rest; for example, the `mem.py log` or note just decided, then the next `mem.py codes`. Never put a write in the same turn as the read it depends on. Never run two writes to the same file as separate parallel calls. Keep one room's material per read, as now. Every tool still writes its own record; nothing about what I read or decide changes.

**Two rules from J at B1b-29 (through the handler, 20:10Z; carry them into every head).** Write rule 0 lines and bags with `b1_multiplace_write4.py` and `b1_bags_from_file2.py`, not the old scripts, which made one `mem.py` call per row of the whole table and grew to 8.5 minutes a write. Never end a turn while a job runs in the background; run every write and check in the foreground.

## The pilot's declarations (from `notes-B2.md`)

- **Agreement:** Jaccard on placement sets (`["T3-004"]` against `["T3-004","T3-011"]` scores 0.5). Fidelity is mean Jaccard against my set. Alpha uses distance 1 − Jaccard, chance over the round's pooled sets. Primary exact agreement is reported beside, never in place.
- **Gold:** 40 items in `gold_items.jsonl`, never shown to a reader.

## The scheme

`scheme_final.jsonl` holds 197 themes (183 main, 14 rare, flagged there) in 21 families. It is frozen: stage B4 logs defects and never edits it.

## Open, carried

- **From B1.** The ruling on testing G13, G14 and G15 as one index page. The B1d leftovers: 106 (line, kind) pairs over 81 episode lines; seven pairs of lines telling one event; E989; the cfb and finland records; shared-member counts that may include EA1r+K5p/lVBLesLhCFRrKOPN+I=; the 76 themes holding both T3-005 and T3-021 acts; T3-149 and T3-150 members that may hold T3-195. The second-order theme the scheme handles worst: argentina t06.
- **Distinctions** (none filed in `_questions/`): who started the work; wars told twice; rivals who turn partners (E1076, E1122, E1142, E1318); whose flag a shared flag is (E1118); who drove a bot (E1025); a room's own scripts barring it from sparing a neighbour (E843) or wrecking a protectorate (E1209); one member on both sides (E820, E947).
- **The bag count.** J counts 841 B1 bags; an older head said 846. Not reconciled. 26 older bags in packets 1 to 5 could list more named themes (QB4-003 named 43); the packet 4 item 197 `also_present` note.
- **Defects for stage C1** (in `adjudication_report.md`). T3-017's and T3-001's discriminators against T3-031's canvas condition; T3-033's "first"; T3-023 and another sub's change; the homeland flag in the siblings' `must_not_show`; T3-018's "whose character"; T3-052 presumes a mark and an owner; T3-086 against T3-050; T3-003's and T3-043's discriminators on parts claimed after the work; "the work" in T3-073 and T3-058; T3-054 against T3-038 and T3-002 on a later canvas; T3-111 and a standing community; T3-087 against T3-058 for one person; T3-117's regret; T3-021 and a cleaned copy; T3-164's reach; T3-009's "lasted to the end" against T3-073's "what survived". From B4c (`handback-B4.md`): G14's request siblings; the class 2 settled line keeping B1 themes no reader placed (22 of 53 sampled); fleet unanimity resting on a field it contradicts (25 of 79).
- **My doubts.** The three-concept floor of the lone-B1 rule is my choice; it let a theme of mine stand against both readers in 162 class 4 items, some on my reading of whose work it is. Class 3 keeps one-concept themes of mine alone at their level. "The act its words lead with" decides some rulings on word order. The disjoint strata are my reading of J's ruling. B4's task differs from B3's by the sibling block, the bag flags and one rule 0 wording, so a difference from B3's figures is not the scheme's alone. The 79 verdicts are mine, made after seeing the fresh placements. Stratum A was drawn on the fleet's split, so its rise in agreement is partly regression.

## Ledgers at this close

- B1 to B3: `scheme_current.jsonl`, `families.jsonl` (21), `confusions.md` (19 pairs), `episodes-B1.jsonl` (1,341); `notes-B2.md`, `pilot_report.md`, `stability-B2.md`, `gold_items.jsonl` (40), `author_items.jsonl` (50), `scheme_final.jsonl` (197); `notes-B3.md`, `assign_raw.jsonl` (17,653), `multiplacement_log.jsonl` (3,171), `b3_classes.jsonl` (7,844), `defects.jsonl` (63), `return_waivers.jsonl` (74), `fleet_ids.jsonl`.
- B4: `notes-B4.md`; `transmission-B4.md`; `coder_conflicts.jsonl` (72); `disagreements.jsonl` (3,288; superseded versions kept, the last 20261013T131513Z); `assign_final.jsonl` (7,844); `adjudication_report.md` (full); `out/_adjudication-B4/`; `sample-B4.jsonl` (500) and `sample-B4-R.jsonl` (185); `RETURN-B4.md`; `out/_packets-B4/` (22 pages, 48 packets, MANIFEST.json); `run_manifest.json` (16 entries); `reassign.jsonl` (999), `reassign-openai-astra.jsonl` (1,000), `readings-R-B4.jsonl` (740); `transmission-after-revision-B4.md`; `handback-B4.md`; the B4c interview brief in `out/interview/`; `out/_analysis/b4b_*.py` and `b4c_*.py` with their records.
- Asks QB2-001 to QB2-014, QB3-001 to QB3-005, QB4-001 to QB4-003 are answered.

## Rules kept

- Never type an id, an order or a name from memory; check with `jq` or a script. Never log a write before the tool answers. Count a ledger's lines after appending.
- There is no `--kind close` and no `--kind episode`; log those as notes.
- A question to J is `mem.py ask "<question>" --assume "<meanwhile>"`. Read an answer with `mem.py asks <id>`.
- Never put a placeholder or a no-op in a command or a log line. Never put an Edit or Write and the read, dry run or validator run that checks it in one turn. Count the head's words before the closing line.
- Never write an id-like token the run does not define in the head, such as a D or a P followed by digits: `done.sh` reads it as an unresolved link.
- The permission check refuses: "=w="; shell for and while loops, and words beginning "for_"; shell functions, brace groups and heredocs; awk with system() or `$` fields; a variable set from `$(...)`, and some `python3 -c` scripts; jq with a `test()` regex, and env-var prefixes; cat or jq on `out/diary/`; `cut -c`, `cut -b` and `colrm` after `mem.py`, `cite.py` or a helper (use `--width`). Write a script file in `out/_analysis/` instead. `.git` is read-denied, and so are reads outside the run folder.
- `mem.py` pages long output; fetch the rest with `--offset`. A Bash output over about 30 KB is saved where I cannot read it; page it. A `Read` of a file over 25,000 tokens is cut; read it by offset and limit.
- After a compaction the guard requires the head and stage file read again and the exact line "compacted; head reconciled" through `mem.py log --kind note`, in a call of its own, before any other step.
- In B4: a ruling never cites a room note; a correction goes through `--replace`, never a hand edit; nothing in `run_manifest.json` is edited, an entry is superseded by a new one.

## The next comparison

Stage B4 is closed at `--cut interview`. No later stage is delivered, so none is begun. What stage C1 needs from B4 is named in `handback-B4.md` ("For stage C1"), `transmission-B4.md` and `transmission-after-revision-B4.md`. The per-theme agreement it needs is stage B3's (`out/_analysis/b3_check_record.json`).