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

Stage D5. The branch accounts, at every level.

Model: as arm.json sets for this stage (your platform file lists the model and effort). Legs: one per branch, then one for the waterfall and survival. Fleet: none.

Steps: D5a-01, D5a-02 ... one per branch; D5b the waterfall and survival. Each leg is named by its step (STANDING §2).

Safe stops: after each branch's account. 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 D5 --pause (STANDING §2).

Startup packet: _head.md, this file, the last five diary entries, validate.py --open D5; the branch's forms/<id>.md; the hierarchy shelf. First leg only: branches.jsonl, crosslinks.jsonl, tree-D4.md, chunk-share-D4.md.

Reads allowed: the branch's members by query (rooms/<sub>/model-D.md, case.md, concepts through cite.py, mem.py note, adaptations --element, contrast); verdicts-D.jsonl and exception-report-D.md; account.md; model/arrows.jsonl; propositions.jsonl; rebuild.py <sub>; _views/; the organizing_end_hour and aftermath share per room. Forbidden: rooms outside the branch except as the contrast the account names; the canvas as an undeclared ground (rule 7), projected on afterwards; recutting a branch or changing the model; the chunk-share tree as a ground.

Outputs: forms/<id>-account.md per branch; branches-D5.jsonl; waterfall-D5.md; survival-D5.md; lines in findings.jsonl and defects.jsonl; notes-D5.md.

The branches are the forms organizing took under your own return; read their members as rooms, in any order, through any file. Two things are fixed: a claim about absence carries §1's arithmetic, and the branches were cut from adaptations your own model provoked and cannot vouch for it. If stage D4 drew no tree, or the population is one form with variation, write this stage anyway, about what the branches do track, and say in every account that they are groupings. The unadapted group gets an account of its own; a branch mostly cannot-tell is said to be so before it is described.

1. Before you call anything absent

Rule 6. Declare here, before any comparison, the minimum count of within-case datings an arrow stands on, from your own subs counts. For every arrow you compare across branches, compute the expectation: its corpus-wide subs count over the analysed set from counts.py, times the branch's n; smallest and largest branch first. Below one the arrow is not-testable-here, and no boundary condition and no propositions_fail may be written from it. Arrow presence inside a branch is arithmetic about your own model, cut on the adaptations it provoked: present here, absent there describes a form and never bounds a proposition, unless you can say what distinguishes the difference from a restatement of the cut.

2. Per branch, the account

forms/<id>-account.md, in this order:

How these rooms organized, as a narrative: a few hundred words a stranger to the model could follow, with participants' words through cite.py, written before the comparisons.
What the branch does: the elements it shares with the global model unchanged, as the model's prose says them.
What it does differently: the shared keys read as prose, each with the rooms and the how sentences that carry it, and what a room of this branch does that the global model would not predict. Say what the branch's claim rests on (contract §1): the shared keys are a pattern across rooms and are validated as one, by the ledger's counts and D4's alternatives; what the rooms did with them is a claim on their words and is validated by cite.py. One kind of evidence does not stand in for the other. Then stage D4's name read against the members: the strongest argument that it names the keys rather than what the rooms did, and what a reader who suspected so would look for; a name you cannot defend from the members is a renaming proposal in notes-D5.md.
The concepts it employs: representative first-order concepts from the members, through cite.py, at least one per shared key.
Composition here. Keep three versioned judgments separate. verdicts-D.jsonl diagnoses the original unadapted global model and remains historical. adaptation-returns.jsonl, the current model hash and D3's unaffected-case reasons establish each adapted-local status. Name all unresolved or insufficient-evidence local accounts and their limits. Then state whether the shared branch explanation is supported, unsupported or unresolved and why. An original global failure is not a current local or shared failure.
How the model runs here: which elements do the work, which sit idle (absent), which fit badly, what these rooms do that it has no place for (propose and particular lines). Where the room models and your reading disagree, say which you believe.
The one relationship that runs differently here: one arrow, both sides quoted.
Fourth-order themes over- and under-represented against the population base rate (no v38 view yet; by script, logged), in the rooms' own words; never a property of the rooms (contract §5).
Arrows present, absent or not-testable-here with the expectation on each; propositions that hold, fail against a real expectation, or are not testable.
How representative members do the work: two or three rooms, by a rule you state, each told in a paragraph in the branch's terms with citations, one the member that fits worst, one read at length with why you chose it and what is unrepresentative about it.
The nearest branch and what separates them, with the room that could have gone either way; every cross-link touching this branch, named.
The form test (contract §1, Doty and Glick): the constructs the branch binds, the relations among them. Map them to the four organizing problems where justified, allowing an empty mapping and a supported theoretical remainder. Then, for what varies among members, whether the form accommodates it, a subform is warranted (a candidate for stage D4, written to notes-D5.md, never cut here), or it is a particular of one history. A branch for which you cannot write this is reported as a grouping, not a form.
What would make you doubt this is a form at all.

branches-D5.jsonl, one line per branch:

{"branch": "F03", "name": "...", "n": 23,
 "verdicts": {"composes": 14, "composes-with-residue": 5, "does-not-compose": 2, "cannot-tell": 2},
 "global_model_version": "<original handoff VERSION>", "local_model_version": "<current hash>",
 "local_statuses": {"room": "validated|unresolved|insufficient-evidence"}, "unresolved_members": [...],
 "shared_status": "supported|unsupported|unresolved", "shared_warrant": "evidence and scope",
 "construct_bindings": ["P07", "T4-005"], "binding_account": "relations and conditions",
 "binds": [...], "leaves_free": [...], "over": ["T4-005"], "under": [...],
 "arrows_present": [{"arrow": "A03", "expected": 2.1}], "arrows_absent": [...], "arrows_not_testable": [...],
 "propositions_hold": ["N12"], "propositions_fail": [...], "propositions_not_testable": [...],
 "survival": {"median_organizing_end_hour": 0, "aftermath_share": 0.0,
              "work_remained": {"kept": 0, "lost": 0, "silent": 0}},
 "canvas": {"median_area_px": 0, "rooms_with_artwork": 0, "read": "one clause"},
 "unmodelled": "...", "form_or_grouping": "form", "confidence": "..."}

3. The waterfall

waterfall-D5.md: one reader's walk from the account, through one branch, through one subform, to one room's model, to its adaptations, to the quotes; and the walk back up. Say at each step what was added and whether it contradicted the level above (contract §1, the check in both directions), and read the walk against prior.md's expectation of what the lower levels describe. Choose the branch and the room by a rule stated before the walk.

4. Survival, per branch

survival-D5.md: organizing_end_hour, the aftermath share and work_remained (how many members' records say the work was kept, lost, or say nothing: silence counts, contract §1) per branch, from the record and never from the canvas, with the branch's ending described in its own terms: what stopped the coordinating, and what the rooms did after. The canvas fields are projected on afterwards and read against this, with the direction of the bias named (rule 7): what the rooms of this form left on the canvas, against the narrative in §2 and the account's paragraph on what the model implies a room keeps. Where the §2 narrative dates a change in what the members did, say which bracket of prior.md §4 it falls in and whether the members' placements by hour (canvas) move at that hour: a timing the canvas corroborates is said to be, a mechanism never is, and a timing it contradicts is named as a question for the account. The contract's §5 sentence on what the canvas measures stands beside it; nothing here is a score.

5. What you must refuse to do

Do not characterise rooms stage D4 declared cousins as siblings. Do not treat a network stage D4 could not draw as a tree of forms. Do not let branch size become importance. Do not recut anything or change the model here: a fourth-order theme that lumps two things a branch separates, a renaming, a subform, goes to notes-D5.md for stage D6, which may act on it as an amendment, and to defects.jsonl for stage E1; a claim a branch bounds is a findings.jsonl line under rule 14. A closed branch's account is not edited; a description later found wrong is superseded by a new file with what showed it (rule 18).

Moments to log

Each branch opened; the expectations computed; the representative members chosen with the rule; the form test's result; the survival figures; the waterfall's steps; what the members showed that the room notes did not predict.

Final message

The claim (what the forms are and how they differ in their functioning), the branch that best supports it, the branch that most threatens it, the condition; then: branches accounted; branches that failed the form test; names that survived the members' reading; the branch whose reading disagreed most with the room models; where the uncomposed rooms sit; arrow absences real and not-testable-here, by count; propositions that failed against a real expectation, and the findings.jsonl lines they bounded; the waterfall's contradictions if any; survival by branch; the most consequential difference between two forms; the licences exercised.

Close

done.sh D5 at every branch; handback-D5.md; the head; the closing diary line.

Negative shared result

If D4 finds no supported shared branch, keep branches-D5.jsonl empty and explain the result in notes-D5.md. The waterfall goes from the global model directly to a selected local account and its evidence. Report remaining local evidence and unresolved cases. Do not invent a branch. binds and leaves_free contain only optional four-problem mappings; construct_bindings must resolve to empirical construct ids, and a form requires their binding_account.