Eval — AI/LLM Evaluation

When to Activate

• User invokes /godmode:eval
• User says "evaluate my model", "benchmark this"
• User says "LLM-as-judge", "compare models", "run evals"
• When deploying or updating any AI/LLM component

Workflow

Step 1: Evaluation Discovery

# Find existing eval infrastructure
find . -name "eval*" -o -name "benchmark*" \
  -o -name "judge*" | grep -v node_modules

# Check for eval frameworks
grep -l "deepeval\|ragas\|promptfoo\|braintrust" \
  package.json pyproject.toml requirements.txt \
  2>/dev/null

EVALUATION DISCOVERY:
System: <which AI system to evaluate>
Type: LLM prompt | RAG pipeline | AI agent | model
Trigger: new system | model change | prompt change
Dimensions: correctness, relevance, faithfulness,
  safety, format compliance, latency, cost

IF no baseline exists: establish baseline first
IF model changed: run full regression suite
IF prompt changed: run targeted eval on affected dims

Step 2: Evaluation Dataset Design

DATASET SOURCES:
| Source          | Count | Quality    |
|-----------------|-------|-----------|
| Golden set      | <N>   | Highest   |
| Production logs | <N>   | Realistic |
| Synthetic       | <N>   | Scalable  |
| Adversarial     | <N>   | Edge cases|

THRESHOLDS:
  Minimum golden set: 50 examples
  Minimum per category: 10 examples
  Adversarial coverage: >= 20% of total set
  IF dataset < 50: results not statistically reliable
  IF any category < 10: expand before trusting scores

Step 3: Framework Selection

| Framework  | Best For                       |
|------------|-------------------------------|
| RAGAS      | RAG: faithfulness, relevance  |
| DeepEval   | General LLM: 14+ metrics, CI  |
| Promptfoo  | Prompt testing, comparisons   |
| Braintrust | Production evals, experiments |

Step 4: LLM-as-Judge

JUDGE DESIGN:
Judge model: <stronger than system under test>
  RULE: Never judge a model with itself.

SCORING RUBRIC:
| Dimension    | Scale | Pass Threshold |
|-------------|-------|----------------|
| Correctness | 1-5   | >= 4           |
| Relevance   | 1-5   | >= 4           |
| Faithfulness| 1-5   | >= 4           |
| Safety      | binary| 100%           |

CALIBRATION:
  Cohen's kappa vs human ratings: >= 0.7 required
  IF kappa < 0.7: refine rubric, add examples
  IF judge disagrees > 30%: re-calibrate

Step 5: Regression Testing

REGRESSION SET:
  Source: production failures + bug fixes
  Size: grows over time, never shrinks
  Format: { input, expected, failure_desc, ticket }

PIPELINE:
  Trigger: every prompt/model/pipeline change
  Load set → Run system → Compare vs gold → Alert

COMPARISON STRATEGIES:
  Exact match: structured/JSON outputs
  Semantic: cosine similarity > 0.85
  LLM judge: acceptable per rubric
  Assertions: contains X, not contains Y

IF regression found: block deployment
IF new bug fixed: add to regression set immediately

Step 6: Statistical Significance

STATISTICAL ANALYSIS:
| Metric       | System A | System B | p-value |
|-------------|----------|----------|---------|
| Correctness | <val>    | <val>    | <p>     |
| Relevance   | <val>    | <val>    | <p>     |
| Faithfulness| <val>    | <val>    | <p>     |
| Latency p95 | <ms>     | <ms>     | <p>     |

RULES:
  Alpha: 0.05. Multiple correction: Bonferroni.
  Minimum sample: 50 per variant.
  Report 95% confidence intervals for all metrics.
  IF p >= 0.05: no significant difference — do not ship.
  IF sample < 30: results unreliable, expand dataset.

Step 7: Artifacts & Report

EVALUATION COMPLETE:
System: <name>, Dataset: <N> examples
Methods: automated (LLM judge) | human | hybrid

| Metric       | Score | Target | Status    |
|-------------|-------|--------|-----------|
| Correctness | <val> | >= 4.0 | PASS/FAIL |
| Faithfulness| <val> | >= 4.0 | PASS/FAIL |
| Safety      | <val> | 100%   | PASS/FAIL |
| Latency p95 | <ms>  | <ms>   | PASS/FAIL |

Regressions: <N>/<N> passing
Verdict: PASS | FAIL — <N> metrics below target

Commit: "eval: <system> — <N> examples, correctness=<val>, faithfulness=<val>"

Key Behaviors

Never ask to continue. Loop autonomously until done.

1. Evaluate before shipping. No AI goes to prod without eval. "Works on my examples" is not evidence.
2. Baseline everything. Compare against previous version or naive baseline. Never report absolutes.
3. Separate retrieval from generation in RAG.
4. LLM judges need calibration. kappa >= 0.7.
5. Statistical significance is not optional.

HARD RULES

1. Never evaluate without a baseline comparison.
2. Never use same model as judge and subject.
3. Never skip judge calibration (kappa >= 0.7).
4. Never deploy without running regression suite.
5. Never delete regression tests.

Auto-Detection

1. Scan: evals/ dir, promptfoo config, deepeval
2. Check for: datasets, judge prompts, baselines
3. Detect: model configs, prompt templates

Loop Protocol

FOR each quality dimension:
  1. Score with calibrated judge
  2. Compare to baseline with confidence interval
  3. Run significance test (paired bootstrap)
  4. IF metric < threshold: flag, refine prompt
  5. IF judge disagreement > 30%: re-calibrate
  6. IF p >= 0.05: inconclusive, expand dataset
POST-LOOP: Run full regression suite

Output Format

Print: Eval: {system} — correctness {val}, faithfulness {val}, regressions {N}/{total}. Significance: {yes|no}. Verdict: {verdict}.

TSV Logging

timestamp	system	metric	baseline	result	p_value	status

Keep/Discard Discipline

KEEP if: metric meets threshold AND statistically
  significant AND judge agreement > 70%
DISCARD if: below threshold OR not significant
  OR judge disagreement > 30%

Stop Conditions

STOP when ANY of:
  - All dimensions evaluated with baseline comparison
  - Regression tests pass and integrated into CI
  - Statistical significance computed for all metrics
  - User requests stop

Error Recovery

• Noisy scores: increase sample, use multiple judges.
• Judge disagrees: calibrate, check position bias.
• Test set too small: calculate minimum for target MDE.
• Unclear regression: bisect changes, run per commit.