Santa Method

Skill — Santa Method (Multi-Angle Independent Verification)

When this skill activates

After any significant implementation, before shipping, or whenever author-bias is suspected. Use when a single reviewer perspective is insufficient to guarantee correctness — the Santa Method ensures two independent reviewers reach the same conclusion without cross-contamination.

Named after the "making a list, checking it twice" principle: one pass is never enough for high-stakes outputs.

Mandatory actions when this skill is active

Before review begins

1. Identify the artifact under review — code diff, architecture doc, API design, or generated output.
2. Define the review rubric — establish 3-7 criteria with clear pass/fail definitions before spawning reviewers.
3. Prepare sanitized context packets — each reviewer gets ONLY:
   • The artifact itself
   • The acceptance criteria / rubric
   • Relevant domain context (docs, types, schemas)
   • NO access to the other reviewer's notes, findings, or verdict
4. Select reviewer perspectives — choose complementary angles:
   • Reviewer A: Correctness & Logic (does it do what it claims?)
   • Reviewer B: Robustness & Edge Cases (what breaks it?)
   • Alternative pairs: Security/Functionality, UX/Architecture, Performance/Maintainability

During review (parallel execution)

Context Isolation Protocol:

• Reviewer A and Reviewer B operate in separate reasoning contexts
• Neither reviewer sees the other's output until both have submitted
• If using sub-agents: spawn as independent tasks with no shared state
• If single-agent: complete Reviewer A's full assessment, seal it, then begin Reviewer B fresh

Each reviewer MUST produce structured JSON output:

{
  "reviewer": "A",
  "perspective": "Correctness & Logic",
  "verdict": "PASS" | "FAIL" | "CONDITIONAL",
  "confidence": 0.0-1.0,
  "findings": [
    {
      "criterion": "string",
      "score": 1-5,
      "evidence": "specific line/section reference",
      "severity": "critical" | "major" | "minor" | "info"
    }
  ],
  "blocking_issues": [],
  "summary": "one-paragraph assessment"
}

AND Gate Logic (non-negotiable):

• PASS requires: Reviewer A = PASS AND Reviewer B = PASS
• If EITHER reviewer returns FAIL → overall verdict is FAIL
• If EITHER reviewer returns CONDITIONAL → enter convergence loop
• This is an AND gate, never an OR gate — both must independently agree

Batch Sampling (for large outputs):

• If the artifact exceeds 500 lines or 20 logical units:
  • Sample 10-15% of units for full santa review (random + targeted selection)
  • Targeted selection: always include security-sensitive, complex, and high-change-rate sections
  • If sampled sections fail: expand review to 30% or full artifact
  • Document which sections were sampled and which were skipped

After review (convergence and resolution)

If both reviewers PASS:

• Record combined verdict as PASS
• Merge findings into unified report
• Log to .mindforge/reviews/santa-[timestamp].json

If disagreement (convergence loop):

1. Identify the specific criteria where reviewers diverge
2. Narrow scope to ONLY the disputed items
3. Re-evaluate disputed items with additional context (both reviewers now see each other's evidence for disputed items ONLY)
4. Round 2 verdict: apply AND gate again
5. If still disagreeing after Round 2: escalate to human decision with both perspectives presented
6. Maximum 2 convergence rounds — never infinite loops

Convergence rules:

• Round 1: Independent review (no cross-contamination)
• Round 2: Targeted re-review of disputed items only (both see each other's evidence)
• Round 3: NEVER — escalate to human if Round 2 fails
• Each round narrows scope (never broadens)

Output format

Final santa review report:

{
  "method": "santa-method",
  "artifact": "description of what was reviewed",
  "timestamp": "ISO-8601",
  "reviewer_a": { /* full reviewer output */ },
  "reviewer_b": { /* full reviewer output */ },
  "and_gate_result": "PASS" | "FAIL" | "ESCALATED",
  "convergence_rounds": 0-2,
  "disputed_items": [],
  "final_verdict": "PASS" | "FAIL" | "ESCALATED",
  "batch_sampling": {
    "total_units": 0,
    "sampled_units": 0,
    "sampling_percentage": 0.0,
    "sampling_strategy": "random+targeted"
  }
}

Self-check before task completion

Before marking a task done when this skill was active:

• Did both reviewers have fully isolated context? (No cross-contamination)
• Did I apply the AND gate strictly? (Both must PASS, not just one)
• Were reviewer perspectives complementary, not redundant?
• Was convergence attempted on disagreement (max 2 rounds)?
• For large artifacts, was batch sampling applied (10-15%)?
• Was the rubric defined BEFORE review began?
• Is the final verdict supported by evidence from both reviewers?
• Was the review output stored in structured JSON format?