git:20260901.829eb8b to git:20260904.5a122b3

8 added, 26 removed. Audit A to A.

---
name: principle-blind-the-investigator
- description: "Apply when dispatching research, scouts, or verification helpers. Hand the investigator the question, never the wanted answer — an investigator that knows the conclusion anchors to it and verifies nothing."
+ description: 'Keeps desired outcomes out of research prompts. Apply when dispatching research, scouting, or verification.'
user-invocable: false
---
# Blind the Investigator
- Hand an investigator the question, never the wanted answer. A researcher
- that knows what is being built returns opinions instead of facts, and a
- verification helper that knows your conclusion will anchor to it and
- verify nothing.
-
- **Why:** Knowing the desired conclusion is the corruption — the
- investigator rationalizes toward it and the answer arrives pre-agreed.
- What the blinding withholds is the framing, not the subject: an
- investigator still needs the question, the vocabulary, and the ground to
- walk.
+ Give investigators the question, vocabulary, and evidence source; withhold the wanted answer and framing.
- **Pattern:**
- - Research consumes neutral questions, never the task framing. A missing
- piece of context surfaces as an open question, not a guess at intent.
- - The isolation extends downward: a scout's prompt carries only verbatim
- question text and stated context — no framing added, no goal named, no
- speculation about intent.
- - Verification helpers get neutral, falsifiable claims with file:line —
- never your verdict, severity, or reasoning. A rule-violation claim
- still carries the rule, because the rule is what makes it falsifiable.
- - One skeptic per claim, always fresh.
- - Treat any leakage as a critical defect: stop and report.
- - The review-gate instance of this — fresh context, no shared history —
- and the one-claim-one-fresh-judge rule are owned by
- `skills/principle-generator-evaluator/SKILL.md`; this is the same move
- applied upstream of review.
+ - Research from neutral questions, never the task framing; surface missing context as an open question, not guessed intent.
+ - Give scouts only verbatim question text and stated context: add no framing, goal, or intent speculation.
+ - Give verification helpers neutral, falsifiable claims with file:line; omit your verdict, severity, and reasoning, but include any rule that makes a violation falsifiable.
+ - Assign one fresh skeptic per claim.
+ - Treat leakage as a critical defect: stop and report.
+ - Apply `principle-generator-evaluator` for review-gate fresh context, no shared history, and one-claim-one-fresh-judge separation.