Whole-Run Investigation

A whole-run investigation (WRI) reads one campaign, or several as equals: each one's log, the check program staged beside it and the harness states it kept, against the question the operator asked. It ends in one vetted synthesis that says whether each reading may be believed, what limited it, and which owner each next change belongs to. It reads and proposes. It starts no solve, no session and no round, it appends to no log, and the host's verifier stays the one owner of every verdict.

The catalogue has six deterministic rows and thirty-one lanes. from-v1.md maps the earlier catalogue to this one. Lane 31 reads a promotion and the climb after it, with every campaign under review side by side. AGENTS.md sections 1 to 8 own the campaign's rules.

These narrower questions need no lanes. Start there when one is the whole question. A wri.ts command's --log accepts a path or a campaign's whole name, as bun run runs does, and start and climb take several, separated by commas. accept.ts and settle.ts below write to the campaign and belong to their own authorised procedures; they are not commands to run inside a read-only investigation.

question	command
Is a run alive, and what did it last do?	bun run runs pulse
Is there enough supply for a comparison, and what can it show?	bun run runs view <name>: its set-up line and the table's move
Does the checker stand on fresh controls?	bun tools/c/accept.ts <the campaign's flags>; stored answers are reported separately
What does a repaired checker move, and does the repair stand?	bun tools/c/repaired.ts --log <log> --bundle <repaired pack>
A report names tasks set aside on rejections: what does it read once they are labelled?	bun tools/d2/settle.ts <the campaign's flags> --session <id>, then label the rejections it draws
Which sessions ran, from which parent, with which reading?	bun tools/wri/wri.ts read --log <log>, row F and the CHAIN line
Did a promotion climb, beside the campaigns that ran the same pack, and did one campaign's /tmp reach another's solves?	bun tools/wri/wri.ts climb --log <log>,<log>,<log>: one row a campaign, then each solve that named what another campaign's session left

The procedure

Six steps. Keep the investigation's files in one directory, <dir>, under the ignored .ana/wri/<date>-<campaign>/. Do not start a step until the check named for the one before holds.

1. Start. wri.ts start --log <log>[,<log>...] --out <dir> folds each log into the six rows and the triggers, writes them to <dir>/overview.md, and writes <dir>/shared-instructions.md with each campaign's row filled in and two sections left blank. Where one question spans several campaigns (other domains, other hosts, other commits of one pack), name them all: they are read as equals. Every lane reads every campaign, opens where a trigger fired on any of them, and triages each finding against the others on its Seen in: line; lane 31 is also given their climbs side by side. The logs are read again when the lanes are built. With bun --env-file=.env.runs tools/wri/wri.ts ... --log <name>, a campaign on another machine is fetched automatically into data/runs/<host>/<name>/. Explicit log paths select an offline snapshot. runs fetch <name> remains available to refresh a copy explicitly. Lanes read the copy with this checkout's tree, so check out the commit the receipt names where the host line says the code differs. lanes rereads the paths stored by start; it does not refresh a live remote campaign behind the investigation's back. Start another investigation to take a new snapshot. Check: read the host-code digest line. Use the matching campaign tree it names, where one exists; otherwise say beside each source claim that it concerns another tree. The digest is not a commit: the launch receipt supplies the commit separately. Settle rows A to F from the log.
2. Edit the shared instructions. Write the two blank sections: the orientation, which is the operator's question, and the moved variable with its prior state. Check: no ## section is blank, and the hard rules are still there.
3. Build the lanes. wri.ts lanes --out <dir> writes one prompt per lane under <dir>/lanes/ and <dir>/launch.json, one session for each lane it built. With no --lanes it opens every lane a trigger started and the four standing lanes. By default it builds no isolated lane: it names the ones whose trigger fired for the primary to assess. A prompt built before is kept and excluded from any newly written launch.json. Check: exit 0, and count the sessions that this call built. If it built none, launch nothing: an earlier launch.json remains on disk. The count is the number of sessions to authorise, not a dollar estimate.
4. Launch. launch.json is a codex-gpt-swarm manifest: gpt-sessions.ts --manifest <dir>/launch.json --output-dir <new dir> ... on the model the operator chose. A Claude session is handed its row's task, one line, and nothing else; start them all in one message. Check: read the model each transcript recorded before using its report.
5. Collect. wri.ts collect --out <dir> holds every expected report to its shape and writes <dir>/reports.md. Read lane 4's report, then decide lane 5; read lanes 11, 15 and 18, then decide lane 19, applying the WINS limitation below. Build lane 5 with wri.ts lanes --out <dir> --lanes 5, or lane 19 with --lanes 19, then launch the one lane it says it built, alone, and collect again. Check: collect exits 0. It exits 1 while a report is missing or invalid, or no report is expected. Review the evidence too: a valid shape does not establish a sound finding.
6. Synthesise. wri.ts synthesis --out <dir> renders <dir>/synthesis-prompt.md; one session writes <dir>/synthesis.md, and the primary vets it ("Vetting" below). Check: every consequential claim is marked as checked against the log or as resting on one report.

Commands

Each runs as bun tools/wri/wri.ts <verb> from the checkout. The usage block at the head of tools/wri/wri.ts is the source.

step	command	writes
read	read --log <log>	nothing: the rows, the triggers and the host line, printed
compare climbs	climb --log <log>,<log>[,<log>...]	nothing: the campaign table and shared temporary paths, printed
1	start --log <log>[,<log>...] --out <dir>	overview.md, shared-instructions.md, investigation.json
3	lanes --out <dir> [--lanes 1-4,7]	lanes/NN-prompt.md, launch.json
5	collect --out <dir>	reports.md
6	synthesis --out <dir>	synthesis-prompt.md
lanes 1, 5	grade --log <log> [--bundle <pack>] --task <taskId> --answer <file>	nothing: one verdict line. --bundle names a stored pack; a campaign that drafted its exam names none

One investigation reads one question, over one campaign or several. Two campaigns of one pack, one under a checker and one under its repair, are read apart, and lane 6 of the later one joins them.

Sizing

One session reads one lane, on every campaign under review: more campaigns make each session longer, and add none. selected opens all four standing lanes even before a comparison reports, alongside any triggered open lanes. Count the manifest rather than assuming a size. When the budget is smaller than the build, read the groups in this order, because each rests on the one before:

1. The instrument (1 to 6). A comparison counts verdicts. Until the refusals read sound, a win may be the check refusing the parent's right answer. Lane 4 first.
2. The comparison (20 to 25). Whether the pairs are pairs, how the scored set was chosen and what the test could show at its size. Lanes 21 and 22 first.
3. The harness (13 to 19). What changed and whether it could know the exam. Lane 18 first.
4. The tasks (7 to 12), the solve (26 to 28), the review loop (29 to 30).
5. The climb (31). Once a promotion is on record, open it before a person is told the campaign improved a harness: it reads what the promotion rested on and what followed, and it rests on lanes 13, 19, 22 and 25 where they ran.

--lanes takes whatever is asked for. A lane named without its trigger opens and is told so by its own overview. An isolated lane is refused without its trigger. The tool permits named isolated lanes in one selection, so the primary must build and launch each alone.

Report checks

reportOf accepts finding, no finding and could not read. It checks the lane heading, equal counts of lines beginning Observation:, Cause:, Owner:, Reverses if: and Decision:, a written finding for a finding verdict, no finding fields for a no finding verdict, and an owner from the catalogue. It does not check the cited evidence or enforce exactly one owner.

• reporting gives a no finding verdict its own ## Checked section, which says what the lane checked and found sound, with ## Findings left empty. The collector does not read that section, so read it before accepting the verdict: an empty one leaves the lane unsubstantiated.
• WINS lists lane 19's solve ids and starts nothing by itself (selected). Lane 19 runs alone: it opens only where it is named and WIN AT A SMALL WALL SHARE or ANSWER FROM A PROGRAM fired.

Hard rules

In every lane prompt

The preset carries these in their own section, so an edited copy keeps them unless the primary deletes them on purpose.

• Quote nothing the walls keep apart. No task's text, answer or hidden value, no check's failure text, no prompt, no passage of a trace, no reviewer's prose. A lane session reads all of them in the log; its report is read by people and sessions that may not. Event kinds, solve ids, digests, check ids, counts, paths and line numbers only.
• Never write under the campaign's folder, and never append to a log.
• Start no solve, session, round or check run. wri.ts grade is the one exception, where a lane names it.
• Open the report early and append findings as they hold; write a verdict only once its evidence is read.
• Stay independent. No lane opens another's prompt or report.
• Read source in the checkout the campaign's row names, and say when it is not the tree that ran.

For the primary

• Settle rows A to F yourself. They are the identities every finding binds to, and no session repeats them.
• Isolated lanes run alone. Lanes 5 and 19 freeze a result before they read anything that could lead it, so neither is given the orientation, the moved variable or another lane's trigger.
• No report reaches a Builder, an Examiner or a Referee. Lane 19 reads traces of scored tasks, and several lanes read hidden values: a finding handed to a session that writes the harness or the tasks is the leak lane 18 looks for.
• A pack/ finding is a decision, never a patch. No seat repairs a campaign's check program. The finding goes to the checker's owner, and a repair is another campaign (accept.ts --freeze, repaired.ts, carry.ts).
• Verify every session's model from its transcript before using its report.

Editing the shared instructions

<dir>/shared-instructions.md is the text every open lane carries, whole, and the synthesis carries it too. Isolated lanes receive only its campaign and hard-rule sections, the deterministic rows, their own triggers, the catalogue's stance and owner table, their lane and the reporting rules. HTML comments are dropped when prompts are built, a ## section left blank refuses lanes, and deleting a heading drops its section on purpose.

section	written by	for
Orientation	the primary	the operator's question and where its evidence is, never the answer expected
The moved variable and prior state	the primary	what changed since the campaign before, the earlier findings this one checks and what would falsify each; "none" when nothing moved
Campaigns under review	start	each campaign, its log, its size, the checkout and its host line
Hard rules	the preset	the rules above

A lane told what to find finds it. An earlier diagnosis goes under the prior state as a hypothesis, never under the orientation as a premise.

Vetting

The synthesis template owns what the note owes: the four links for each reading, the binding constraint, one findings table, conflicts, a build plan, operator decisions and predictions. The primary then treats the note as research, in two acts kept apart:

• Triage sorts each finding by kind and owner (host defect, pack, operator decision, prediction, already changed at the head) and checks nothing.
• Vetting checks a finding against the logs, the source at the head and, where it names them, the kept harnesses and the check program, then rules it confirmed (recounted or read here), plausible (resting on one lane's report) or refuted.

Vet every finding, and above all every one that would change code or a decision. Write the result as <dir>/vetting.md, one row a finding with its verdict and what was checked, and never report a finding that was only triaged as vetted. On 22 October eight of thirteen findings were first reported as adjudicated after triage alone.

• Start from the operator's decision. Usually one of: believe this step, run another round, repair the checker, run the Examiner again. Order the synthesis by that.
• Read the instrument before the tally. A rejection counts only once it is ruled a true fail. In a simulated campaign a too-strict check gave 6 wins to 0 and a step; the same campaign under the repaired check gave 1 win and no-step. A comparison now sets a task aside on a rejection ruled a check defect or unclear, or ruled by nobody, and its report holds the tally were those rejections to stand (counted). Only a person's labels settle it (tools/d2/settle.ts), and the settled report says what was labelled, disputed and restored. An investigation proposes that reading; it never makes it.
• Read not-scored and an empty no-step as no test. A comparison that tallied no task says nothing of its harness. Row D's count of held-out tasks, and lane 8's of those waiting for a label, say whether the next one can do better.
• Keep the observation apart from the cause. "Forty of fifty-five refusals name one check" is an observation; "the check is too strict" and "the solver cannot do it" are two causes, and lane 5 is what separates them.
• Recount. Every "N of M" the synthesis leans on is recounted from the log's events.
• Say whose reading a count is. A ruling is the Referee's. Its agreement with a person is unmeasured until hand labels exist.
• Confirm against current source. Before a patch row becomes work, check that the defect still stands. Follow AGENTS.md section 11 and stack-pr for its branch and PR.

Where the detail lives

file	owns
references/lanes.md	the stance, the owner table, rows A to F and lanes 1 to 31
references/from-v1.md	what became of v1's nine rows and thirty-eight lanes, and what that says about v2's direction
references/shared-instructions.template.md	the preset every open lane carries
references/synthesis.template.md	the synthesis session's prompt
tools/wri/read.ts	the fold of a log into rows and triggers
tools/wri/catalogue.ts, tools/wri/investigation.ts, tools/wri/wri.ts	the catalogue's parser, the prompts and the report check, and the verbs