Adversarial Reasoning

When NOT to Use

• Trivial fact recall
• Simple arithmetic (unless ambiguity exists)
• Purely stylistic writing tasks
• Low-stakes conversational questions

For those, answer directly with a quick assumption spot-check.

Procedure

Stage 1: Problem Framing

State precisely:

• What question is actually being asked
• What the key unknowns are
• What constraints must be preserved
• What ambiguities could materially change the answer

If ambiguity exists, either branch explicitly or state the dominant interpretation and why.

Stage 2: Candidate Space

List 2–4 plausible candidates, interpretations, or solution paths. Only include candidates that are actually plausible — not combinatorial noise.

Stage 3: Assumption Extraction

For each candidate, identify:

• Explicit assumptions
• Hidden assumptions
• Definitional dependencies
• Required environmental conditions

Stage 4: Negative Analysis (Internal)

For each weak or invalid candidate, provide:

• Failure label(s) from the typed list below
• Concise explanation of why it fails
• Whether the failure is fatal or repairable
• A counterexample or contradiction when available

Allowed failure labels: false_premise · unsupported_assumption · invalid_inference · missing_constraint · boundary_condition_failure · contradiction_with_known_facts · internal_inconsistency · category_error · equivocation · overgeneralization · underspecified_claim · solves_adjacent_problem · locally_valid_globally_invalid · circular_reasoning · non_falsifiable · conclusion_too_strong

If none apply cleanly, say so explicitly and describe the uncertainty.

Stage 5: Positive Construction

Build the strongest surviving candidate:

• Restate allowed assumptions
• Show the valid inference chain
• Identify the exact structural difference from failed candidates
• State the answer with calibrated confidence

The answer must be contrastive — it must explain why it is right relative to the nearby wrong answers.

Stage 6: Cross-Model Adversarial Critique

This is the key step. Do NOT self-critique here. Shell out to independent models.

Format a payload containing:

1. The original question
2. The positive candidate from Stage 5
3. The assumptions it relies on
4. The rejected alternatives and why they were rejected
5. The inference chain

Then call the adversarial critique script via terminal:

cat <<'PAYLOAD' | /path/to/adversarial-reasoning/scripts/adversarial-critique.sh
QUESTION:
<the original question>

POSITIVE CANDIDATE:
<the answer from Stage 5>

ASSUMPTIONS:
<numbered list of assumptions this answer relies on>

REJECTED ALTERNATIVES:
<for each rejected candidate: what it claimed and why it was rejected>

INFERENCE CHAIN:
<the step-by-step reasoning that produced the positive candidate>
PAYLOAD

Parse the JSON response. Each critic will return:

• survival: survives / damaged / fatal
• issues[]: typed failure labels with targets, critiques, severities
• hidden_assumptions_found[]
• edge_cases_tested[]
• overall_note: weakest link assessment

The response also includes consensus (combined verdict) and agreement (unanimous/partial).

Rules for handling verdicts:

• If consensus is fatal — the candidate has a structural flaw. Return to Stage 5 with the specific issues and reconstruct.
• If consensus is damaged — the candidate has problems but may be repairable. Address each serious issue. If repaired, note what changed.
• If consensus is survives — proceed to Stage 7. Still note any minor issues.
• If critics disagree — investigate the specific points of disagreement. The truth is usually in the details.
• If all critics unavailable — self-critique with extra skepticism and mark the answer as unverified_by_independent_critique.

Do NOT dismiss critics' findings without explanation. If you disagree with a specific issue, state why with the same rigor you'd apply to any other claim.

Stage 7: Final Answer

Deliver:

• The best surviving answer (post-critique)
• The most important rejected alternative and why it failed
• The main assumption on which the conclusion depends
• Confidence level: high / medium / low
• What would change the answer (what new information would flip it)
• Critic consensus and whether any issues were addressed

Behavioral Rules

1. Never accept first-plausible-answer syndrome. At least one competing candidate must be considered.
2. Criticism must be typed. No "this seems wrong" without a failure label.
3. The final answer must be contrastive. Why right relative to nearby wrong.
4. Rhetoric is not evidence. Elegant prose does not substitute for valid inference.
5. Candidate sets must be disciplined. Plausible, not exhaustive.
6. Ambiguity surfaces early. Before analysis, not after.
7. Conclusion matches support. Never claim certainty when only plausibility is established.
8. Independent critique is real input, not theater. Engage with it substantively.
9. Distinguish fatal vs repairable failures. Some candidates are wrong at the root; others need one condition added.
10. Never merge incompatible frameworks silently. If two candidates use different definitions, keep that explicit.

Lightweight Mode

For problems that warrant adversarial care but not the full 7-stage treatment:

1. State the answer
2. State the strongest counter-argument
3. Shell out with a compact payload (question + answer + one alternative)
4. Incorporate critique
5. Deliver with confidence level

Failure Modes of This Skill

Watch for:

• Performative critique with no logical bite (from any model)
• Too many low-value candidates flooding the process
• Mislabeled failures
• Mistaking skepticism for rigor
• Over-pruning genuinely valid but unusual answers
• False precision in confidence statements
• Treating any single critic's verdict as infallible (every model has blind spots — just different ones)