Adversarial Review

Input: $ARGUMENTS

Overview

Nothing is a “fact” until it has survived an assassination attempt. The Builder constructs a claim; the Breaker tries to destroy it. Only conclusions that survive attack are trustworthy.

This prevents the passivity of accepting first-obvious answers.

Steps

Step 1: Builder — Construct the Claim

State the claim as precisely as possible:

What exactly is being claimed?
What evidence supports it?
What reasoning connects evidence to conclusion?
What is the confidence level?
What would change if this claim is wrong?

BUILDER'S CASE:
Claim: [precise statement]
Evidence:
1. [evidence 1] — strength: [H/M/L]
2. [evidence 2] — strength: [H/M/L]
Reasoning: [how evidence supports claim]
Confidence: [%]
Stakes: [what depends on this being true]

Step 2: Breaker — Attack the Claim

Systematically attempt to destroy the claim:

Attack the evidence:

Is the evidence real? (Source reliable? Data accurate?)
Is the evidence relevant? (Does it actually bear on the claim?)
Is the evidence sufficient? (Could the claim be false despite this evidence?)
Is the evidence cherry-picked? (What evidence was NOT presented?)
Is the evidence current? (Could things have changed?)

Attack the reasoning:

Does the conclusion follow from the premises? (Valid inference?)
Are there hidden assumptions? (What’s unstated but required?)
Are there logical fallacies? (Ad hominem, straw man, false dilemma, etc.)
Is correlation being mistaken for causation?
Is the reasoning reversible? (Could the same logic support the opposite?)

Attack the claim itself:

Is it falsifiable? (Can it be proven wrong in principle?)
Is it specific enough to test? (Or so vague it can’t fail?)
Does it contradict known facts?
Is there a simpler explanation for the evidence?
What would a knowledgeable opponent say?

Step 3: Score Each Attack

For each attack that landed:

Attack	Target	Severity	Claim Survives?
[attack description]	evidence/reasoning/claim	fatal/serious/minor	Y/N/weakened

Severity levels:

Fatal: Claim cannot survive this attack. Must be abandoned or fundamentally revised.
Serious: Claim is significantly weakened. Needs major revision or additional evidence.
Minor: Claim still holds but with reduced confidence or narrower scope.
Missed: Attack doesn’t actually undermine the claim.

Step 4: Builder — Respond to Attacks

For each serious or fatal attack, the Builder can:

Refute: Show the attack is wrong (provide counter-evidence)
Repair: Modify the claim to survive the attack (narrow scope, add caveats)
Reinforce: Add new evidence or reasoning that survives the attack
Concede: Accept the attack and abandon/revise the claim

BUILDER'S RESPONSE:
Attack: [which]
Response type: [refute/repair/reinforce/concede]
Content: [the response]
Revised confidence: [new % if changed]

Step 5: Final Verdict

After Builder-Breaker exchange:

Outcome	Meaning	Action
Claim survives all attacks	High confidence — treat as reliable	Accept
Claim survives with repairs	Moderate confidence — accept with caveats	Accept with noted limitations
Claim partially survives	Split confidence — some parts reliable, some not	Accept reliable parts, reject or investigate weak parts
Claim falls to attacks	Low confidence — do not rely on this	Reject or fundamentally revise

Step 6: Report

ADVERSARIAL REVIEW:
Original claim: [Builder's claim]
Confidence before review: [%]

Attacks:
| # | Attack | Severity | Builder Response | Result |
|---|--------|----------|-----------------|--------|
| 1 | [attack] | [level] | [response] | [survived/weakened/fell] |

Verdict: [survived / survived with repairs / partially survived / fell]
Confidence after review: [%]
Revised claim: [if modified]

Strongest surviving evidence: [what held up]
Weakest point: [most vulnerable aspect]
What would change verdict: [what new evidence/argument would matter]

When to Use

Before committing to any strategy
When confidence seems too high for evidence
When stakes are high
When you notice you haven’t considered alternatives
→ INVOKE: /stc (steelmanned counterarguments) for strongest possible opposition
→ INVOKE: /aw (assume wrong) for systematic wrongness analysis
→ INVOKE: /cv (convergent validation) for multi-method verification

Verification

Claim stated precisely (not vaguely)
Breaker attacked evidence, reasoning, AND claim
Attacks scored for severity honestly
Builder responded to each serious attack
Final confidence is calibrated (not just original + or - a bit)
Strongest surviving evidence identified

advr