Ask the right question

The two workflow families tell you what to PRODUCE. This skill tells you what to ASK, of whom, in what order — the reasoning that happens between an operator message and the first phase artifact. Working here fails in two ways: asking nothing (you invent facts and build on them) or asking the operator everything (their attention is the scarcest budget in the system). Both are the same mistake — a question routed to the wrong authority.

1. The three authorities, in cost order

Every question has a cheapest competent authority. Ascend only when the cheaper one genuinely cannot answer.

Authority	Answers questions like	Cost
THE RECORD — log.jsonl, objects/, typed verdicts, blobs, instruments (verify_root, tools/root_sweep.py, tools/docs_verify.py, the gate)	what happened; which value was rejected; do two instruments agree; did anything move	a command
THE FRAMEWORK — CLAUDE.md, docs/map/ (Traps first), docs/ERRATA.md, the tranche ledgers (REQUEST/GOAL/PARKED), prior DELIVERY reconciliations	what is frozen; what has gone wrong here before; what did the operator already rule; where does X live	a file read
THE OPERATOR	genuine forks the record and framework underdetermine; frozen-surface approval; taste on user-facing shape	the scarcest budget there is

Two rules that are always in force:

• Cite the instrument with the number. Two instruments can both be right and disagree: the root census is 45 by direct manifest load over git ls-files and 42 rows by root_sweep.py, which scans experiments/ only. A number without its instrument is not a fact yet (docs/ERRATA.md E5, E8; DR-INV-frozen-surfaces, "The root sweep").
• Model prose is never evidence. Yours included. If your answer to a record-question does not end in a command output or a file path, you have not asked the record yet.

2. Reading the operator

The operator writes tersely and expects the repo's context to carry the rest. Each row is a real committed exchange — look it up when unsure.

The operator says	It means	Not	Committed evidence
"Do it" / "go ahead" after you stated a plan	approval of EXACTLY that plan	a new vague instruction; license to widen	FIX.md, experiments/2026-08-03-fix-attached-evidence-integrity ("operator approval on the stated plan: 'Do it.'")
"…as you go" ("fix documentation as you go")	a STANDING grant — but you must bound it in the tranche ledger and park the rest	permission to survey and rewrite everything	that tranche's GOAL.md, "Documentation (operator-granted scope addition)"
"are you using X?" about your process	a prompt: check honestly, answer plainly, and apply X now — the honest "no" is expected and useful	an accusation to deflect, or a yes/no to answer without checking	REPRO.md map-finding section of the same tranche (the answer "not until now" surfaced a missing seam)
"start an errata please" (artifact named, shape unstated)	derive the shape from repo conventions (honest ledger, append-only, evidence pointers) and build it; say what you assumed	a reason to ask what an errata is	docs/ERRATA.md header block
an apparent typo or wrong name ("Claude.me")	resolve by repo context and proceed; note the resolution in one clause	a blocker; a thing to ask about	CLAUDE.md is the only near-match in the tree
a new instruction mid-tranche	APPEND to REQUEST.md verbatim, reconcile through dr-spec-change	something to absorb silently into the current step	dr-change-orchestrator, "The ledger rule"

The general rule behind every row: the operator's words are AUTHORITY, so treat them the way the harness treats a record — quote them verbatim into the ledger, then interpret them in writing where the interpretation can be reviewed (REQUEST.md quotes; SPEC.md interprets). If you find yourself paraphrasing the operator from memory, you have already lost the thread.

3. Ask the record first

Before any theory about a failure or a surprise, in order:

1. Which instrument produced this? Name it. A verdict without an instrument is hearsay (section 1).
2. What does the typed artifact say verbatim? run-status.json state and stop_reason; the violation's check and detail; the blob under blobs/ with the rejected value. Both cycle-0 deaths in this repo's history were misattributed by readers who theorized before opening the diagnostic blob (CLAUDE.md, hard-won invariants; DR-SUB-capabilities Traps, jolt run-b4d6dfda).
3. Do two instruments agree? rc codes, verify_root, the audit JSON, the report layer. Agreement is signal; disagreement is a finding about an INSTRUMENT, not yet about the system (ERRATA E8: a fix's predicted verdict flip appeared in verify_root but not in the sweep, because the sweep also binds the stored terminal summary — frozen evidence).
4. Has this gone wrong here before? The covering map document's Traps section, then docs/ERRATA.md. A recurrence is the cheapest diagnosis available.
5. What would falsify my current reading? If nothing could, it is not a reading, it is a mood.

4. Derive before asking

(Absorbed from dr-decide-or-ask, kw8imd lineage, commit 86f1248e — credited, superseded here.) Before composing ANY question to the operator, derive their answer from what they have already said: this tranche's verbatim words, CLAUDE.md's standing rules, prior rulings in DELIVERY reconciliations and RESULTS conventions. This operator's recorded values: smallest correct change; fix readers, never invalidate committed records; no frozen surface without explicit approval; honesty over polish — negative results recorded as negative; a reachable defect is fixed or parked, never shipped silently.

The dominance test: would every reasonable operator holding those values choose the same option? Then the fork is false — decide, act, and record one line: "Decided without asking (dominant under your recorded values): — override any time."

What earns a question: options that survive the dominance test AND change real stakes — frozen-surface or irreversible action, >2x effort divergence between defensible readings, a conflict WITHIN the operator's own words, or taste on user-facing shape. Batch every such fork into ONE question set per tranche; lead with your recommendation and its one-sentence reason; state each option's consequence in the operator's terms (risk, cost to them, honesty of the record), including the do-nothing consequence when it leaves something reachable-broken. When the earning reason is frozen-surface contact, the question MUST embed tools/blast_radius.py's BLAST_RADIUS_RESULT_V1 result — this is section 1's own "cite the instrument with the number" rule, applied to this specific instrument (Rung G6, docs/map/INV-frozen-surfaces.md). A STOP that describes contact without pasting the gate's own computed list is words given over an inferred surface, not a disclosed one (docs/ERRATA_EXECUTOR.md, the 2026-08-09 entry; design premise: experiments/2026-08-10-change-blast-radius-analysis/REQUEST.md).

5. Frame forks falsifiably

When you are genuinely uncertain between two readings, do not pick one and build — and do not ask yet either. Write BOTH readings as alternatives the record can decide between, each with the evidence that would prove it, the way the attached-evidence tranche's GOAL.md framed W ("the writer breached; the finding is correct; the root stays invalid") against R ("the reader over-demands; the finding is spurious") before any code was read. Then collect the deciding evidence. The fork that survives IS your diagnosis; the fork that died is your "Ruled out" section. A goal written this way is decidable under either outcome, which is what makes it a goal rather than a hope (experiments/2026-08-03-fix-attached-evidence-integrity/GOAL.md, DIAGNOSIS.md).

6. The wrong-question table

Questions this repo has already paid for asking — or paid for not asking. Each cites the committed scar.

Wrong question	Right question	Scar
"Which schema rule failed?" (theorizing at cycle-0)	"What does the diagnostic blob say verbatim?"	CLAUDE.md hard-won invariants: both cycle-0 deaths misattributed on first reading
"Is the count 42 or 45?"	"Which instrument, over which scope, produced each count?"	ERRATA E3/E5: both were true; the census check went stale asserting one of them
"Why is this root invalid?" (accepting the finding's prose)	"Does the artifact the finding names actually exist in the state?"	ERRATA E6: the detail text named as missing an artifact present at seq 4
"Which artifact mentions the source record?" (shape as identity)	"Which artifact carries the discriminator only the trusted writer can stamp?"	DR-SUB-verification Traps: a mention-shaped predicate selected ordinary citations
"Can I reuse this passing check?"	"Can this check FAIL, and on whose machine?"	ERRATA E7: four map checks passed only on the machine whose gitignored roots they opened
"Should I ask the operator which option they prefer?"	"Do the operator's recorded values already decide this?"	Section 4; a false fork spends the budget the whole system is designed to conserve

Exit criterion. Before you send any question upward or act on any assumption: every question you almost asked is either answered-with-a-command (record), answered-with-a-citation (framework), decided-and-recorded (dominant), or sitting in ONE batched question set with a recommendation. Zero questions whose answer a reader of CLAUDE.md, the map's Traps, and this tranche's ledger could have given without you.