npbuilds Statement Audit — Director
The last line of defense before a statement is handed to a solver. This director realizes six-eyes' Phase 6 (Audit): it runs the candidate statement through the standard failure-mode checks and ends with the six-axis grader whose sub-threshold result triggers the re-state loop. Audit is adversarial by design — its job is to break the statement, not to bless it.
Routing Table
| Child skill | Checks for | Trigger if it fails |
|---|---|---|
xy-detector |
Asking about an attempted solution rather than the real goal | Return the inferred underlying goal; re-state |
answerability-tester |
Could any evidence in principle resolve this? | Unanswerable → reframe or narrow |
falsifiability-checker |
What observation would prove it wrong? (Popper) | Unfalsifiable → tighten the claim |
scope-bounder |
Are time / population / context boundaries explicit? | Unbounded → add scope |
statement-grader |
Six axes: specificity, falsifiability, scope, audience-fit, answerability, root-vs-symptom | Any axis < 3/5 → re-state loop |
Routing Logic
- Run the checks in order —
xy-detectorfirst (an XY pattern invalidates everything downstream), then answerability, falsifiability, scope. - End with
statement-grader. Any axis below 3/5 returns control tosix-eyes, which re-enters Phase 4 (Reframe) with the failure diagnosis as input. Cap at 3 iterations. - Surface the per-axis scores so the convergence trajectory is auditable, never loop silently.
Scope Boundaries
- In scope: validating logic, answerability, falsifiability, scope, and overall quality.
- Out of scope: generating fixes (→
problem-reframing), shaping the output (→statement-compression). Audit diagnoses failure; the orchestrator routes the fix. - Escalate to
six-eyeson multi-axis failure across iterations — that signals a deep-lane upgrade, not another cosmetic re-state.