validate-reasoning

name	validate-reasoning
description	Validate reasoning quality after generating a response, code solution, or analysis. Use when: the user invokes /validate-reasoning, asks to check reasoning, wants to verify claims are grounded, or suspects hallucination. Targets: hallucination, unfaithful chain-of-thought, overconfidence, self-contradiction.
user-invocable	true

Validate Reasoning — Anti-Hallucination & Logic Audit

You are performing a structured reasoning audit on your most recent response (or on the content the user specifies). This is a POST-GENERATION check. Go through each step methodically.

Step 1: Grounding Audit

Review every factual claim, code reference, and assertion in the response. Classify each as:

• [GROUNDED] — Directly supported by files read, context provided, or tool output in this conversation
• [INFERRED] — Logically derived from grounded facts (state the inference chain)
• [TRAINING] — From general knowledge / training data (not verified against this project)
• [UNVERIFIED] — Cannot determine the source; may be fabricated

List the top claims with their classifications. Flag any [UNVERIFIED] items prominently.

Step 2: Logical Consistency Check

Scan the response for:

• Contradictions (X is stated in one place, not-X in another)
• Unsupported logical jumps (conclusion does not follow from premises)
• Circular reasoning (A because B, B because A)
• Missing premises (conclusion requires an unstated assumption)

If contradictions are found, state them explicitly.

Step 3: Confidence Calibration

For each major claim or recommendation:

• Is the confidence level appropriate given the evidence?
• Are there hedges where there should be certainty (or vice versa)?
• Would it be more honest to say "I don't know" or "I'm uncertain"?

Flag any overconfident claims (high certainty + weak evidence).

Step 4: Code/File Verification

If the response references specific files, functions, variables, or APIs:

• Use Read/Grep to verify they actually exist
• Confirm the behavior described matches the actual code
• Flag any references that cannot be verified

Step 5: Summary Verdict

Provide a brief summary:

• Grounding score: X of Y claims are grounded
• Consistency: Pass / Issues found
• Confidence calibration: Appropriate / Overconfident on [items]
• Action items: What should be corrected or re-examined

Be direct and honest. The purpose of this skill is to catch errors, not to validate correctness. Err on the side of flagging concerns.