Assumption Auditor

Expose the premises that could silently break the work. This skill is for auditing assumptions, not for doing the full plan, research, implementation, or adversarial QA pass.

Workflow

1. State the claim or action being audited.

   • Name the answer, plan, artifact, decision, code change, or workflow that depends on assumptions.
   • If the target is unclear, audit the user's latest request and the intended next action.

2. Separate assumptions by type.

   • Facts: external, local, temporal, technical, legal, financial, medical, or source-derived.
   • Interpretations: what the user probably meant, which audience matters, which quality bar applies.
   • Environment: tools, paths, permissions, dependencies, credentials, services, model behavior, runtime state.
   • Design choices: defaults, priorities, tradeoffs, non-goals, compatibility expectations.
   • Success criteria: what counts as done, good, current, safe, or acceptable.

3. Rate each assumption.

   • Confidence: high / medium / low.
   • Blast radius: high / medium / low if the assumption is false.
   • Drift risk: stable / version-sensitive / time-sensitive.
   • Verification cost: cheap / moderate / expensive / impossible in this turn.

4. Decide the next treatment.

   • Verify now when confidence is low, blast radius is high, and verification is cheap.
   • Ask the user when the missing premise is a preference or decision boundary that cannot be inferred safely.
   • Proceed with a recorded assumption when the default is low-risk and reversible.
   • Route to $uncertainty-router when the main problem is choosing the cheapest verification path.
   • Route to $codex-adversarial-qa when the artifact needs hostile scenario testing, not just premise inspection.

5. Keep the audit compact.

   • Lead with the assumptions that could change the answer or work.
   • Do not list obvious background truths unless they affect risk.
   • Do not block execution on low-risk assumptions; record them and continue when the broader task calls for action.

Output Shape

Use this structure unless the user requested another format:

• Audited target
• Critical assumptions
• Moderate assumptions
• Safe defaults / recorded assumptions
• Verification plan
• User question, only if needed

For tabular work, use columns:

Assumption | Type | Confidence | Blast radius | Drift risk | Verification | Treatment

Stop Conditions

• Stop when the remaining uncertainty no longer changes the next action.
• Do not turn an assumption audit into a full implementation plan.
• Do not present unverified assumptions as confirmed facts.
• When using memory-derived facts without current verification, mark them as memory-derived and note stale risk when relevant.