santa-method

# Santa Method

Multi-agent adversarial verification framework. Make a list, check it twice. If it's naughty, fix it until it's nice.

The core insight: a single agent reviewing its own output shares the same biases, knowledge gaps, and systematic errors that produced the output. Two independent reviewers with no shared context break this failure mode.

## When to Activate

Invoke this skill when:
- Output will be published, deployed, or consumed by end users
- Compliance, regulatory, or brand constraints must be enforced
- Code ships to production without human review
- Content accuracy matters (technical docs, educational material, customer-facing copy)
- Batch generation at scale where spot-checking misses systemic patterns
- Hallucination risk is elevated (claims, statistics, API references, legal language)

Do NOT use for internal drafts, exploratory research, or tasks with deterministic verification (use build/test/lint pipelines for those).

## Architecture

```
┌─────────────┐
│  GENERATOR   │  Phase 1: Make a List
│  (Agent A)   │  Produce the deliverable
└──────┬───────┘
       │ output
       ▼
┌──────────────────────────────┐
│     DUAL INDEPENDENT REVIEW   │  Phase 2: Check It Twice
│                                │
│  ┌───────────┐ ┌───────────┐  │  Two agents, same rubric,
│  │ Reviewer B │ │ Reviewer C │  │  no shared context
│  └─────┬─────┘ └─────┬─────┘  │
│        │              │        │
└────────┼──────────────┼────────┘
         │              │
         ▼              ▼
┌──────────────────────────────┐
│        VERDICT GATE           │  Phase 3: Naughty or Nice
│                                │
│  B passes AND C passes → NICE  │  Both must pass.
│  Otherwise → NAUGHTY           │  No exceptions.
└──────┬──────────────┬─────────┘
       │              │
    NICE           NAUGHTY
       │              │
       ▼              ▼
   [ SHIP ]    ┌─────────────┐
               │  FIX CYCLE   │  Phase 4: Fix Until Nice
               │              │
               │ iteration++  │  Collect all flags.
               │ if i > MAX:  │  Fix all issues.
               │   escalate   │  Re-run both reviewers.
               │ else:        │  Loop until convergence.
               │   goto Ph.2  │
               └──────────────┘
```

## Phase Details

### Phase 1: Make a List (Generate)

Execute the primary task. No changes to your normal generation workflow. Santa Method is a post-generation verification layer, not a generation strategy.

```python
# The generator runs as normal
output = generate(task_spec)
```

### Phase 2: Check It Twice (Independent Dual Review)

Spawn two review agents in parallel. Critical invariants:

1. **Context isolation** — neither reviewer sees the other's assessment
2. **Identical rubric** — both receive the same evaluation criteria
3. **Same inputs** — both receive the original spec AND the generated output
4. **Structured output** — each returns a typed verdict, not prose

```python
REVIEWER_PROMPT = """
You are an independent quality reviewer. You have NOT seen any other review of this output.

## Task Specification
{task_spec}

## Output Under Review
{output}

## Evaluation Rubric
{rubric}

## Instructions
Evaluate the output against EACH rubric criterion. For each:
- PASS: criterion fully met, no issues
- FAIL: specific issue found (cite the exact problem)

Return your assessment as structured JSON:
{
  "verdict": "PASS" | "FAIL",
  "checks": [
    {"criterion": "...", "result": "PASS|FAIL", "detail": "..."}
  ],
  "critical_issues": ["..."],   // blockers that must be fixed
  "suggestions": ["..."]         // non-blocking improvements
}

Be rigorous. Your job is to find problems, not to approve.
"""
```

```python
# Spawn reviewers in parallel (Claude Code subagents)
review_b = Agent(prompt=REVIEWER_PROMPT.format(...), description="Santa Reviewer B")
review_c = Agent(prompt=REVIEWER_PROMPT.format(...), description="Santa Reviewer C")

# Both run concurrently — neither sees the other
```

### Rubric Design

The rubric is the most important input. Vague rubrics produce vague reviews. Every criterion must have an objective pass/fail condition.

| Criterion | Pass Condition | Failure Signal |
|-----------|---------------|----------------|
| Factual accuracy | All claims verifiable against source material or common knowledge | Invented statistics, wrong version numbers, nonexistent APIs |
| Hallucination-free | No fabricated entities, quotes, URLs, or references | Links to pages that don't exist, attributed quotes with no source |
| Completeness | Every requirement in the spec is addressed | Missing sections, skipped edge cases, incomplete coverage |
| Compliance | Passes all project-specific constraints | Banned terms used, tone violations, regulatory non-compliance |
| Internal consistency | No contradictions within the output | Section A says X, section B says not-X |
| Technical correctness | Code compiles/runs, algorithms are sound | Syntax errors, logic bugs, wrong complexity claims |

#### Domain-Specific Rubric Extensions

**Content/Marketing:**
- Brand voice adherence
- SEO requirements met (keyword density, meta tags, structure)
- No competitor trademark misuse
- CTA present and correctly linked

**Code:**
- Type safety (no `any` leaks, proper null handling)
- Error handling coverage
- Security (no secrets in code, input validation, injection prevention)
- Test coverage for new paths

**Compliance-Sensitive (regulated, legal, financial):**
- No outcome guarantees or unsubstantiated claims
- Required disclaimers present
- Approved terminology only
- Jurisdiction-appropriate language

### Phase 3: Naughty or Nice (Verdict Gate)

```python
def santa_verdict(review_b, review_c):
    """Both reviewers must pass. No partial credit."""
    if review_b.verdict == "PASS" and review_c.verdict == "PASS":
        return "NICE"  # Ship it

    # Merge flags from both reviewers, deduplicate
    all_issues = dedupe(review_b.critical_issues + review_c.critical_issues)
    all_suggestions = dedupe(review_b.suggestions + review_c.suggestions)

    return "NAUGHTY", all_issues, all_suggestions
```

Why both must pass: if only one reviewer catches an issue, that issue is real. The other reviewer's blind spot is exactly the failure mode Santa Method exists to eliminate.

### Phase 4: Fix Until Nice (Convergence Loop)

```python
MAX_ITERATIONS = 3

for iteration in range(MAX_ITERATIONS):
    verdict, issues, suggestions = santa_verdict(review_b, review_c)

    if verdict == "NICE":
        log_santa_result(output, iteration, "passed")
        return ship(output)

    # Fix all critical issues (suggestions are optional)
    output = fix_agent.execute(
        output=output,
        issues=issues,
        instruction="Fix ONLY the flagged issues. Do not refactor or add unrequested changes."
    )

    # Re-run BOTH reviewers on fixed output (fresh agents, no memory of previous round)
    review_b = Agent(prompt=REVIEWER_PROMPT.format(output=output, ...))
    review_c = Agent(prompt=REVIEWER_PROMPT.format(output=output, ...))

# Exhausted iterations — escalate
log_santa_result(output, MAX_ITERATIONS, "escalated")
escalate_to_human(output, issues)
```

Critical: each review round uses **fresh agents**. Reviewers must not carry memory from previous rounds, as prior context creates anchoring bias.

## Implementation Patterns

### Pattern A: Claude Code Subagents (Recommended)

Subagents provide true context isolation. Each reviewer is a separate process with no shared state.

```bash
# In a Claude Code session, use the Agent tool to spawn reviewers
# Both agents run in parallel for speed
```

```python
# Pseudocode for Agent tool invocation
reviewer_b = Agent(
    description="Santa Review B",
    prompt=f"Review this output for quality...\n\nRUBRIC:\n{rubric}\n\nOUTPUT:\n{output}"
)
reviewer_c = Agent(
    description="Santa Review C",
    prompt=f"Review this output for quality...\n\nRUBRIC:\n{rubric}\n\nOUTPUT:\n{output}"
)
```

### Pattern B: Sequential Inline (Fallback)

When subagents aren't available, simulate isolation with explicit context resets:

1. Generate output
2. New context: "You are Reviewer 1. Evaluate ONLY against this rubric. Find problems."
3. Record findings verbatim
4. Clear context completely
5. New context: "You are Reviewer 2. Evaluate ONLY against this rubric. Find problems."
6. Compare both reviews, fix, repeat

The subagent pattern is strictly superior — inline simulation risks context bleed between reviewers.

### Pattern C: Batch Sampling

For large batches (100+ items), full Santa on every item is cost-prohibitive. Use stratified sampling:

1. Run Santa on a random sample (10-15% of batch, minimum 5 items)
2. Categorize failures by type (hallucination, compliance, completeness, etc.)
3. If systematic patterns emerge, apply targeted fixes to the entire batch
4. Re-sample and re-verify the fixed batch
5. Continue until a clean sample passes

```python
import random

def santa_batch(items, rubric, sample_rate=0.15):
    sample = random.sample(items, max(5, int(len(items) * sample_rate)))

    for item in sample:
        result = santa_full(item, rubric)
        if result.verdict == "NAUGHTY":
            pattern = classify_failure(result.issues)
            items = batch_fix(items, pattern)  # Fix all items matching pattern
            return santa_batch(items, rubric)   # Re-sample

    return items  # Clean sample → ship batch
```

## Failure Modes and Mitigations

| Failure Mode | Symptom | Mitigation |
|-------------|---------|------------|
| Infinite loop | Reviewers keep finding new issues after fixes | Max iteration cap (3). Escalate. |
| Rubber stamping | Both reviewers pass everything | Adversarial prompt: "Your job is to find problems, not approve." |
| Subjective drift | Reviewers flag style preferences, not errors | Tight rubric with objective pass/fail criteria only |
| Fix regression | Fixing issue A introduces issue B | Fresh reviewers each round catch regressions |
| Reviewer agreement bias | Both reviewers miss the same thing | Mitigated by independence, not eliminated. For critical output, add a third reviewer or human spot-check. |
| Cost explosion | Too many iterations on large outputs | Batch sampling pattern. Budget caps per verification cycle. |

## Integration with Other Skills

| Skill | Relationship |
|-------|-------------|
| Verification Loop | Use for deterministic checks (build, lint, test). Santa for semantic checks (accuracy, hallucinations). Run verification-loop first, Santa second. |
| Eval Harness | Santa Method results feed eval metrics. Track pass@k across Santa runs to measure generator quality over time. |
| Continuous Learning v2 | Santa findings become instincts. Repeated failures on the same criterion → learned behavior to avoid the pattern. |
| Strategic Compact | Run Santa BEFORE compacting. Don't lose review context mid-verification. |

## Metrics

Track these to measure Santa Method effectiveness:

- **First-pass rate**: % of outputs that pass Santa on round 1 (target: >70%)
- **Mean iterations to convergence**: average rounds to NICE (target: <1.5)
- **Issue taxonomy**: distribution of failure types (hallucination vs. completeness vs. compliance)
- **Reviewer agreement**: % of issues flagged by both reviewers vs. only one (low agreement = rubric needs tightening)
- **Escape rate**: issues found post-ship that Santa should have caught (target: 0)

## Cost Analysis

Santa Method costs approximately 2-3x the token cost of generation alone per verification cycle. For most high-stakes output, this is a bargain:

```
Cost of Santa = (generation tokens) + 2×(review tokens per round) × (avg rounds)
Cost of NOT Santa = (reputation damage) + (correction effort) + (trust erosion)
```

For batch operations, the sampling pattern reduces cost to ~15-20% of full verification while catching >90% of systematic issues.