Dr Ask The Right Question

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(Absorbed from dr-decide-or-ask, kw8imd lineage, commit 86f1248e — credited, superseded here.)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): <choice> — 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.

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.