counter-position

Counter-Position Protocol

Philosophy

An opinion is NOT valid until you can effectively argue BOTH sides.

This is not debate for debate's sake. It is the core epistemological discipline: if you cannot steelman the opposition, you have not understood the problem deeply enough to decide. Confirmation bias is the failure mode this protocol exists to prevent.

Do NOT finalize any architectural decision until counter-position has been run. If you cannot construct a strong argument against your preferred position, your confidence is not earned.

Auto-Invocation Triggers

Signal	Why
Any architectural decision	Architectural choices are expensive to reverse
High confidence + low evidence	Confidence without evidence = confabulation
Single surgeon dominates consensus	Groupthink risk — the dissenting view needs construction
Sentinel risk >= high	High-risk changes deserve adversarial scrutiny
"Obviously" or "clearly" in reasoning	These words often mask unexamined assumptions

The 4-Step Protocol

You MUST create a TodoWrite task for each step:

Step 1: State the Claim

What is being asserted? Write it as a clear, falsifiable statement.

Good: "We should use Redis sorted sets for the priority queue because they provide O(log N) insertion with built-in ordering."

Bad: "Redis is the right choice." (Too vague to counter — what aspect? Compared to what?)

Step 2: Steelman the Opposition

Construct the strongest possible argument AGAINST the claim. This is not a strawman — it is the best case the other side could make.

Rules:

• Argue as if you genuinely believe the opposition
• Use the strongest evidence available, not the weakest
• Address the claim's actual strengths, not irrelevant weaknesses
• If you cannot construct a strong counter-argument, that IS evidence the claim is strong

Step 3: Test the Opposition

Check the counter-argument against actual code, evidence, and data. This is where theory meets reality.

Evidence Grade	Standard	Action
Empirical	Tested against running code/system	Trust — this is ground truth
Documented	Referenced in docs, specs, prior decisions	Trust — verified by prior work
Analogical	Pattern from similar systems/contexts	Consider — useful but not conclusive
Theoretical	Reasoning from principles without direct evidence	Investigate further before deciding
Anecdotal	Single observation or claim without verification	Do NOT decide based on this alone

Counter-position decisions MUST be backed by Empirical or Documented evidence. Theoretical/Anecdotal evidence triggers additional investigation.

Step 4: Decide with Evidence

Only after testing both sides, form a position:

• If evidence clearly favors one side: state the decision with evidence citations
• If evidence is ambiguous: present BOTH sides to Aaron with your assessment of the evidence quality
• If evidence contradicts your initial position: update your position. Corrigibility > confidence.

Surgeon Integration

Neurologist (Qwen3-4B)

Ask with code-specific framing:

You are part of a SOFTWARE DEVELOPMENT protocol called '3-surgeons'.
You are the local LLM (Qwen3-4B). We write CODE, not perform medical surgery.

CLAIM: [the assertion being tested]
COUNTER-ARGUMENT: [the steelmanned opposition]

YOUR ROLE: Classify which side has stronger pattern support based on the
code context provided. Look for: similar patterns in the codebase,
precedent from prior decisions, technical constraints that favor one side.

CODE CONTEXT: [relevant code snippets]

Why domain anchoring: Without explicit code-context framing, Qwen3-4B confabulates into medical neurology. Confirmed empirically: 3/3 false positives without anchoring, 7/7 actionable responses with anchoring.

Cardiologist (GPT-4.1-mini)

Ask to cross-examine the evidence for both positions:

Cross-examine these two positions for a SOFTWARE DEVELOPMENT decision:

POSITION A: [the claim]
EVIDENCE A: [supporting evidence with grades]

POSITION B: [the counter-argument]
EVIDENCE B: [supporting evidence with grades]

Check for: logical consistency, cognitive biases (confirmation, anchoring,
availability), confidence calibration, and whether evidence grades are
accurately assigned.

The cardiologist's strengths to exploit here:

1. Logical consistency scanning — detect contradictions between claims and evidence
2. Cognitive bias detection — flag confirmation bias, anchoring, availability heuristic
3. Confidence calibration — flag overconfident claims with weak evidence

Anti-Patterns

• Strawman opposition: Constructing a weak counter-argument you can easily defeat. If the opposition is easy to dismiss, you have not steelmanned it.
• Cherry-picking evidence: Selecting only evidence that supports your preferred position. Both sides get equal evidence scrutiny.
• Anchoring on first opinion: The first analysis (often Atlas's) anchors subsequent thinking. Counter-position exists to break this anchor.
• Skipping Step 3: Theoretical arguments feel convincing. They must be tested against code reality before deciding.