consensus

Consensus

When to Use

Consensus is the lightweight decision tool -- a quick confidence-weighted vote rather than a full cross-examination. Use it when:

• Validating a claim -- "Is this approach safe?" "Does this pattern follow best practices?"
• Checking assumptions -- before building on an assumption, verify it with both models
• Pre-commit verification -- "Will this change break backwards compatibility?"
• Resolving simple disagreements -- when you have two plausible options and need a tiebreaker
• Quick sanity checks -- "Is this the right data structure for this use case?"

Do NOT use for: complex architectural decisions (use cross-examination), health checks (use probe), or risk assessment (use sentinel).

How to Invoke

CLI

3s consensus "Redis LPUSH is O(1) for single elements"

MCP Tool

consensus(claim="Redis LPUSH is O(1) for single elements")

How It Works

Each surgeon receives the claim and responds with a JSON assessment:

{
  "confidence": 0.85,
  "assessment": "agree",
  "reasoning": "Redis LPUSH is documented as O(1) per element..."
}

The system then calculates a weighted consensus score:

weighted_score = sum(confidence_i * assessment_value_i) / sum(confidence_i)

Where assessment values are: agree = +1.0, uncertain = 0.0, disagree = -1.0.

Interpreting the Weighted Score

Score Range	Meaning	Action
+0.7 to +1.0	Strong agreement	Proceed with confidence. Both surgeons agree the claim is valid.
+0.3 to +0.7	Moderate agreement	Proceed but verify. One surgeon is less confident or uncertain.
-0.3 to +0.3	Uncertain / Split	Do not proceed blindly. Consider escalating to full cross-examination.
-0.7 to -0.3	Moderate disagreement	The claim is likely wrong or oversimplified. Investigate further.
-1.0 to -0.7	Strong disagreement	The claim is rejected by both surgeons. Re-examine your assumption.

Reading Individual Assessments

The result includes each surgeon's individual vote:

Cardiologist: agree (confidence=0.90)
Neurologist:  uncertain (confidence=0.60)
Weighted score: +0.50
Total cost: $0.0008

Key things to notice:

• Confidence asymmetry -- if one surgeon is very confident (0.9) and the other is not (0.3), the confident vote dominates the weighted score. Check whether the less-confident model has a valid concern.
• Assessment disagreement -- one agrees, one disagrees. This is the most valuable signal. Why do they disagree? Escalate to cross-examination to find out.
• Both uncertain -- neither model has enough information. You may need to provide more context or rephrase the claim more specifically.

When to Escalate to Cross-Examination

Escalate from consensus to full cross-exam when:

1. Weighted score is in the uncertain zone (-0.3 to +0.3) AND the decision is consequential
2. Surgeons explicitly disagree (one agrees, one disagrees) regardless of confidence levels
3. Both are uncertain with low confidence (<0.5) on a claim you need to act on
4. The reasoning conflicts -- even if both agree, check if their reasoning is compatible

Do NOT escalate when:

• Strong agreement on a low-stakes claim
• The uncertain zone is acceptable for a reversible action
• You already have external evidence (documentation, tests) that resolves the claim

Cost

Consensus makes 2 LLM calls (one per surgeon). The Neurologist call is free (local model). The Cardiologist call costs approximately $0.0004-0.001 depending on claim length. This is roughly 1/5 the cost of a full cross-examination.

Example Workflow

# Step 1: Quick consensus check
3s consensus "Switching from SQLite to PostgreSQL is safe for our <100 user base"
# Result: weighted_score=+0.45 (moderate, but uncertain zone edge)

# Step 2: Score is borderline -- escalate
3s cross-exam "Should we switch from SQLite to PostgreSQL given <100 concurrent users?"
# Result: Full analysis reveals Neurologist flagged WAL mode contention
# that Cardiologist missed, while Cardiologist noted migration complexity

The consensus caught that this was not a clear-cut decision. The cross-examination revealed exactly why.