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

Stage C5. The account.

Model: as arm.json sets for this stage (your platform file lists the model and effort). Legs: build (one leg per pass of the account, closed at the pass's end or earlier when you choose; every move logged with mem.py log --kind move, its Grodal, Anteby and Holm move named, and snapped with python3 tools/snap.py C5 move-<move>-<n> when it is made); dispatch (three readers on the platform's reviewer model, outside your session, sent by fleet.py); check (you are relaunched, told the readers left their files in _fleet_raw/C5/, and revise, one leg per pass). Fleet: three readers, once. This is the last stage before the model and the tiers are fixed. After it nothing below the model moves except as an affected-case-validated amendment at D3 (rule 11); D6 records separate unvalidated proposals, and stages D1 to E1 measure against it.

Steps: C5a-01 ... one per pass of the account; C5b the readers' build (it ends at --cut packets-out), the dispatch, C5c-01, C5c-02 ... the readers' check, one leg per pass over the returns; C5d the Gioia table, the dynamic model and the figure, the handback and the interview. Each leg is named by its step (STANDING §2).

Safe stops: after each logged move, after each section of the account, and after the data structure and the figure. 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 C5 --pause (STANDING §2).

Startup packet: _head.md, this file, the last five diary entries, validate.py --open C5; the last story, model/story-pN.md of the pass C2 kept. In a check leg the harness adds one line: the readers left their files in _fleet_raw/C5/.

Reads allowed: everything of yours: model/ as it stands after C3 and C4 (mem.py diff-model <snap>, with the name of C2's last close snap in out/_snap/, shows what they changed; its arrows carry the mechanisms C3 supplied and the from-field hypotheses C4 entered, which the propositions run on; read manifest-C4.txt first, then manifest-C3.txt), model.md, structure.md, dimensions.jsonl, fourth/fourth.jsonl, concept.md, literature-C3.md, mapping-C3.jsonl, field-C4.md, findings.jsonl, within_case.md, story_rooms-pN.jsonl; your notes through mem.py note; counts.py; rooms by query (mem.py room, forest, where, codes, pairs, rooms --with, contrast); quotes through cite.py; _fleet_raw/C5/ in check legs. Forbidden: new elements (the model is fixed at the close of C5 except through validated amendments at D3); clustering anything (stage D4); the canvas as an undeclared ground (rule 7); your notes to any reader.

Outputs: account.md, the final version of the model's explanation; sketches.md; account_passes.jsonl; prose-test-C5.md; propositions.jsonl; author_claims.jsonl; RETURN-C5.md and three packets; readers-C5/ (three returns) and notes-C5.md answering them; data_structure-C5.md; figure-C5.md (the model figure as text and as the generated drawing); implications-C5.md; findings.jsonl lines; tier_history.jsonl and model/history.jsonl lines for any recut; 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); a run_manifest.json entry; handback-C5.md; the model frozen (done.sh records the hash of the model and its tiers at C5d's close, out/_closes/model-C5.sha256).

Write sections 1 to 5 before 6, and 6 before the packets go out; the figure and the data structure are re-emitted at C5d from the model that survives the readers. A cut can still move here; test every one you suspect while the account is open. If a proposition will not state because a fourth-order theme is wrong, move it now, logged and written as stage C2 says (scheme_v2.jsonl and mem.py reassign for a third-order recut), re-placing what the amendment touches yourself and redrawing whatever depended on it.

1. Write the model as prose, before you draw anything

account.md is the model in continuous prose, one to two pages: how organizing came about in these rooms and how the parts worked together, so that a reader who was not here can follow it once through and say what it claims. Tell it both ways: from the population down to a room, and from three rooms up to the population. Every claim goes through an element id and a quote or a transition; the concept sits in its place; the loops and the irreversibles across hour 73 are named; where the model should break is said in its own terms. It describes the universe, not the mode: read it against a room you know is odd and make it say what happened there as one way of running the same structure; every element should have a range. The last story of stage C2 is its first draft, and the rooms that story could not tell (story_rooms-pN.jsonl) are its first test. Name every community the account cannot reach, one clause each; the return (D1 to D3) returns them anyway (rule 12) and stage D5 reads them inside their branches.

Three things this catches. A model you cannot write as prose is a taxonomy with arrows added; if every paragraph is and also, say so. An arrow that will not survive being stated as when X, Y follows, because Z is not grounded, whatever arrows.jsonl says. A probabilistic claim needs evidence for its tendency and scope. If the relevant condition is unknown, state that uncertainty; do not invent a boundary or turn uncertain support into a deterministic mechanism (rule 16).

2. The prose test on the names

Every fourth-order theme and dimension is a construct declared at stage C2 (borrowed, adapted, coined). In prose-test-C5.md, write one sentence per construct in which the name does theoretical work, because the room had X, Y could happen: the sentence a reader would say about a room using the name and nothing else. A name that only labels the sentence's subject and does no work in it is a description one tier too high; recut it under rule 11, logged, and keep the grounded phrase as the gloss. A name a reader will mistake for its mistaken_for construct without being told needs its difference stated in the account. A third-order theme that turns out to be named for a construct is renamed in the material's words, logged.

3. Draw it more than one way

sketches.md: at least three arrangements of the same model, each with a paragraph on what it showed and what it hid; by time on the canvas clock, by tier (conditions above, mechanisms in the middle, what follows below), by trajectory (one room's path, then a second whose path differs). If two drawings put the same element in different places, that element is doing two jobs and is a candidate for a recut. Working sketches; section 10 specifies the figure.

4. Simplify, and say what it cost

Cut the model to what a reader can hold at the top; the count is yours, set by what keeps composition and explanatory power alive (contract §2), with why in notes-C5.md. For every element cut: what it was doing and what the model can no longer account for without it. For every element kept over an alternative: which alternative and why. A cut element keeps its id in model/ with its status changed and a history.jsonl line with the move, logged and snapped when it is made.

5. Recut when the account says to

Rule 11 governs, and a recut here is the ordinary outcome, each a tier_history.jsonl line with its move, logged and snapped when it is made. Keep the construct declaration current with every recut, and re-examine every findings.jsonl line it touches in the same act (rule 14). Before pass 1, write in notes-C5.md what would make you prefer a later account to this one; a criterion written after the passes exist is not a criterion.

A counterexample requires explanation. Record the evidence and compare local adaptation, construct or mechanism refinement, a supported scope condition, and withdrawal. Do not invent a condition to preserve a claim. An isolated case can validate a local exception without automatically revising the shared model. Preserve unresolved evidence and record the decision in findings.jsonl and model history, including any weakened or withdrawn claim.

6. The propositions

A proposition here can carry a number: a claim of the form authority is claimed and no dated consequence of it appears anywhere is answerable against the analysed set, whichever way the count comes out.

Write the propositions the model implies into propositions.jsonl, one line each, numbered: the figure tags its arrows with them and stage D5 says which hold in each branch.

{"id": "N12", "statement": "...", "rests_on": ["A03", "A07"], "theorised": false,
 "finding": "F-031", "carried_by": "words|pattern|both", "rival": "one clause"}

For each, in the line or in its paragraph in account.md:

• The claim, in one sentence, in the model's own terms.
• The mechanism, carried up from the arrow that grounds it and marked from-literature where stage C3 supplied it. A proposition without a mechanism is a pattern. Where the proposition sits at the dimension tier and the arrow that grounds it sits lower, say so, and say whether every arrow beneath it shares that mechanism; if not, narrow the proposition to the subset that does, or drop it.
• The evidence in this corpus, with its count, its gate (rule 5), and the contract's §5 warning where the count is an absence; rule 6 governs every absence: say how many communities could have carried the thing before you say it is missing from any of them.
• Whether the arrows beneath it are observed or theorised (theorised). A proposition resting wholly on reading-alone arrows is a conjecture drawn from the data structure, and legitimate. Mark it.
• What carries it (carried_by, contract §1): words, a claim about what rooms did or meant, validated by exemplars that pass its gate and by the return; pattern, a claim about a distribution, an order on the clock or a difference between groups of rooms, validated by counts with their denominators and by the documented events of the 73 hours; both only with one named primary in the paragraph. The evidence you name is of the kind you declare; a pattern claim offered quotes as its proof, or a words claim offered a count, is the defect stage E1 looks for.
• The strongest rival (rival): the other reading of the same evidence that a careful stranger would offer, and what in the record decides between them; where nothing in the record does, the claim is thin (rule 8) and says so.
• What would falsify it in a population that is not this one.
• Where it should not hold: the boundary, in the constructs the model is built from. Corpus size, subreddit topic and which rooms the frame holds are limits on the sample, not boundary conditions; a boundary that names a surface feature of the setting is not one.

Do not write more than the model earns: four defensible propositions are better than a dozen padded ones. A thin claim is made and labelled thin (rule 8), with what it would take to firm it up; the evidence is weak, but the difference was large and the record suggests why is preferred to silence.

A proposition that says a problem was solved is a claim, not a report. One about task division, task allocation, reward provision or information provision is stated from your structure and mapping-C3.jsonl; the return (D1 to D3) reads it room by room. Write it so that it can fail.

7. The passes

Every pass, in the build leg and in the check leg, is one leg and one line in account_passes.jsonl:

{"pass": 2, "t": "<ISO>", "leg": "C5-build-2", "what_changed": "one clause",
 "tier_events": ["<ids from tier_history.jsonl and model/history.jsonl>"],
 "sketches_tried": 3, "elements_cut": ["..."], "criterion_says": "prefer|reject|no change"}

Stop when a pass changes no element and no cut. No cap; from pass 4 on say in notes-C5.md what the criterion says about each. When the passes stop, build the packets and close.

8. Three readers

Build leg, last pass. Append this stage's entry to run_manifest.json (the readers' model, seed, decoding, count). Then, in this order:

1. author_claims.jsonl first: five to nine lines, each one claim you believe account.md makes, in a form a reader could agree or disagree with: {"id": "c3", "claim": "...", "where": "paragraph 4"}.
2. Three rooms named in notes-C5.md: one the account reaches easily, one the last story could not tell, one from the side of the onset contrast the account has visited least. Draw them by query, never by opening notes.
3. RETURN-C5.md, the one return specification (FLEET.md): the return line below, the end marker, and the line telling the reader to confirm it reached it. Three packets (packets.py files --stage C5 --name C5-reader-k <files>, the three rooms dumped first from the verbs into files), each with account.md, figure-C5.md, propositions.jsonl, the three rooms through the verbs with room served --no-canvas, forest and cite.py (never your notes), and nothing else of yours, under the 50 KB ceiling. The instruction, verbatim: Read the account as a stranger. Say where it carries what it means and where it does not, both ways: a sentence you cannot follow, a claim you cannot find the evidence for, a room you can name that the account does not fit, jargon a reader would not accept. Do not propose another model. A disagreement is never a verdict. Then: For each of the three rooms you were given, say whether the account describes what happened there and, where it does not, what the room did instead, one paragraph per room.

The return line: {"reader": "r1", "served": "<sha256 of account.md>", "claims": ["..."], "relations": ["A03 acts on P04"], "could_not_follow": ["quoted"], "unsupported": "...", "rooms": [{"sub": "...", "fits": true, "instead": "..."}]}, five to nine claims in the reader's own words. Stop here. Close with done.sh C5 --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 reviewer route, each reader in a session of its own. The next launch of stage C5 is the check leg, and its packet tells you one thing: the readers left their files in _fleet_raw/C5/.

Dispatch leg. Not yours. The dispatch sends each packet to a fresh reader on the platform's reviewer model, collects the returns into _fleet_raw/C5/, validates them against RETURN-C5.md, reissues what fails the schema, and closes. Nothing in it reads your files.

Check leg. You are relaunched. Copy the validated returns into readers-C5/. Score fidelity both ways in notes-C5.md: your claims each reader recovered, and claims they recovered that you did not write. A claim all three found that you never intended is the account teaching something unintended. Every divergence is a defect in the account, not the reader; author-error is a permitted verdict. Answer every item under rule 13: an error is fixed; a room the account does not describe is a bound written at section 5 or a residue named for the return (D1 to D3); a disagreement about how to tell it is printed beside the account; every change is logged and is a pass (section 7). Nothing a reader says changes an arrow's grounding. The readers are not judges: they tell you whether the account carries what you meant and whether it describes their three rooms, and nothing else. They run once; no second round.

9. The data structure, final: data_structure-C5.md

Nothing is moving now. The display over every concept is generated: gen_views.py writes _views/data-structure.md and .tsv at every close, one row per first-order concept up to its dimension, never edited (rule 15). What you write is data_structure-C5.md, the table this method is named for, from the tiers as they finally stand, all five of them: full over the tiers, representative in its quotes. Six columns, one row per third-order theme at minimum, more where a theme needs them, so that every third-order theme, fourth-order theme and dimension appears:

The leftmost column is material through cite.py, never a label you wrote. Where a quote will not carry a row on its own, use the coder's second-order theme label and say so in the cell. Every row's quote carries its metadata, in the row or in a footnote keyed to it: community, author handle as the record carries it, the hour since the canvas opened, and the concept id. Rows show the range rather than the centre; name in notes-C5.md the rule that chose them, and the paper's figure is a sample of these rows chosen by that rule. Carry the residue: below the table, the third-order themes no fourth-order theme picked up and the fourth-order themes no dimension gathers, each with the account from structure.md and one of its quotes. Keep it to what one person can read; structure.md carries the rest.

10. The figure

One figure, taken in at one pass, in figure-C5.md, drawn once with the theorised arrows and once without; gen_views.py draws both from model/ and propositions.jsonl into _views/figure-model.md (Mermaid) at every close, and the text beside them is yours.

• The things at the top and the arrows among them, the size section 4 reduced the model to (as many as a reader can take in at one pass; say why the count), every element and arrow by id.
• Every arrow is tagged with the id of the proposition that carries it. An arrow no proposition carries is not drawn.
• Conditions on one side, what follows on the other, and things named for what they do in the middle, the mediators. If the figure has no middle column, say so: a two-column lattice with tagged arrows is a taxonomy, and a reader should be told which they are looking at. Show loops as loops, only where model/loops.jsonl records the full dated cycle at its stated within-room or inter-room level. Do not turn cross-case similarity into a process loop.
• The imposed terminus belongs in the picture: placement stopped at hour 73 for everyone.

The text specifies the figure precisely enough that someone else could draw it: every box, every arrow, what each arrow is grounded in, the instance shown for it (the quote or the forest transition from its support or support_transition, so that the arrow is a thing that happened and not a line), its proposition id, and the community count on it. An arrow with no instance is not drawn (contract §2). Then, in one paragraph, what the figure leaves out: the arrows no proposition carried, the third-order themes the model does not use, the register the topic-based coder tier scattered.

11. Transferability, and where the model should fail

Say what carries. Establish from the evidence the relevant conditions: voluntary participation, existing authority, resource access, legal or informal commitments, the external clock and whether failure dissolves a group. Do not assume all rooms share these conditions. Name settings where the conditions plausibly hold and where they do not, what the model predicts, and what this record cannot check. State the frozen frame and A1 exclusions from their files: 705 rooms in this snapshot, with 706 a target census, not complete observational coverage. The artwork-based frame cannot warrant claims about all communities that organized.

Name the conditions under which the model should hold and should break, in the things the model is built from. Corpus size, subreddit topic and which rooms the frame holds are limits on the sample and do not count here.

12. Implications

implications-C5.md: what the model implies for the theory question, in the model's own terms; what a counterfactual at a leaf would borrow from a neighbouring leaf, with the mechanism explicit (contract §1, prediction secondary); what the return (D1 to D3) should find if the model composes, written as expectations the handoff can carry (which elements should instantiate in which kinds of room, which should be absent), so that the return (D1 to D3) is a replication in Eisenhardt's sense and not a re-reading. Stage D1 writes EXPECTATIONS.jsonl from this file; stage E1 scores the ledger against it. It also says, in one paragraph, what the model implies for which forms keep the work alive, answered from the record (contract §1): organizing_end_hour, the aftermath share and work_remained in organizing_end-A1.jsonl, where silence leaves the outcome unknown and the last coordinating turn is a last observation, not a demonstrated end. The canvas is not evidence here.

13. What you may not do, and what the canvas is here

You may not cluster anything; that is stage D4. A room the account cannot reach is named, not fitted with a new element. If the account claims which of the contract's §1 four problems your structure reaches richly and which thinly, that claim is stage C3's mapping restated, and the return (D1 to D3) reads it beside the ledger.

The canvas is what was made (rule 7). The account may say, in one paragraph, how the rooms whose record says their work was kept (work_remained, stage A1) organized differently from those whose record says it was lost or falls silent, in the model's terms; you held the canvas throughout, so this is a comparison and never a prediction, and stage D5 reads it against the canvas by branch. The paragraph says whose pixels it is about: much of a room's artwork was placed by people who never wrote in the room (stage A1 reported the share placed by those who wrote in it), so a claim about what the room's members kept and a claim about what survived on the room's patch are two claims, and the paragraph makes one of them by name or is not written. It may not be built from the canvas: no element is cut, kept or named because of what a room drew, and no proposition takes the canvas as its evidence.

Moments to log

The criterion; the account's drafts, one entry per pass; each sketch and what it hid; each element cut and what it cost; each name that failed the prose test; each recut with its move; each finding bounded or changed; each proposition written, thin ones marked; the three rooms named for the readers; each reader item and its answer; the rule that chose the table's rows; each expectation written for the return.

Final message

The claim, the room that best supports it, the room that most threatens it, and the condition that separates them. Then: passes run and what each changed; arrangements drawn and which the account rests on; what you cut and what the model can no longer account for; tier events by tier, and whether any third-order recut moved second-order themes; names recut by the prose test; the readers' fidelity both ways, their items by kind and what changed, and what they said of their three rooms; the criterion and what it says about the pass you kept; propositions, how many thin, how many wholly on theorised arrows, how many from-literature; things and arrows in the figure, arrows shown by a quote and by a transition, arrows dropped; rows in the data structure and the rule that chose them; the expectations written for the return; what the account showed that the room notes did not predict; the sentence in account.md you are least sure a reader will take as meant; the licences exercised (METHOD.md). A return value.

Close

done.sh C5 at the end of every pass, or earlier when you choose, every recut, element cut or dropped hypothesis logged and snapped as it was made; paste the close-check report into notes-C5.md; the harness hashes model/ and the account; handback-C5.md; the head; the closing diary line.

Interview point. C5d writes out/interview/C5d.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 C5 --cut interview: the model finished: the data structure (the Gioia table) written, the dynamic model explained, the figure plotted. 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).