/guardian

Learning Feedback Loop Reference

Purpose: Calibrate Guardian from Judge, Zen, Launch, and squash outcomes without overfitting to too little evidence.

Contents

• Feedback sources
• JUDGE_TO_GUARDIAN_FEEDBACK
• Calibration rules
• .agents/guardian.md storage
• Improvement metrics
• Squash learning
• AUTORUN and Launch integration

Feedback Sources

Primary sources:

• JUDGE_TO_GUARDIAN_FEEDBACK
• ZEN_TO_GUARDIAN_HANDOFF
• LAUNCH_TO_GUARDIAN_HANDOFF
• squash outcome feedback

Use feedback to update:

• prediction confidence
• quality score calibration
• pattern exceptions
• team preferences

JUDGE_TO_GUARDIAN_FEEDBACK

## JUDGE_TO_GUARDIAN_FEEDBACK

**PR**: #123 - feat(auth): add OAuth2 support
**Review Status**: CHANGES_REQUESTED

### Guardian's Predictions vs Actual Findings
| Predicted Issue | Predicted Severity | Actual Finding | Match |
|-----------------|-------------------|----------------|-------|
| Null pointer oauth.ts:45 | HIGH | Confirmed | TRUE_POSITIVE |
| Race condition token.ts:33 | MEDIUM | Not found | FALSE_POSITIVE |
| - | - | XSS in callback.ts:72 | FALSE_NEGATIVE |

### Prediction Accuracy for This PR
- Judge precision: ...
- Judge recall: ...

Calibration Rules

Prediction confidence

• initial confidence: 70%
• true positive: +5%
• false positive: -10%
• false negative: +10% sensitivity on similar patterns
• bounds: minimum 30%, maximum 95%

Threshold calibration

Trigger threshold recalibration when:

• 3 consecutive same-direction adjustments occur

Quality score calibration

Target prediction bands:

• A+ predicted: 95+
• A predicted: 85-94
• keep average deviation < 15%

No-adjust guardrail

Do not recalibrate major heuristics from too little evidence:

• fewer than 3 relevant examples
• contradictory feedback without a clear dominant pattern

.agents/guardian.md Storage

Store project-specific learning in this shape:

# Guardian Calibration Data

## Last Updated
2026-03-06

## Pattern Exceptions
- naming:
  - ...

## Team Preferences
- squash:
  - ...

## Historical Accuracy
- prediction_accuracy: ...
- false_positive_rate: ...

Improvement Metrics

Targets:

• prediction accuracy > 75%
• false positive rate < 20%
• average score deviation < 15%

Squash Learning

Use squash calibration only when:

• 5+ data points show repeated factor misalignment

Examples:

• pairwise score too permissive on multi-author chains
• file-overlap weighting overfits noisy commits

AUTORUN Integration

AUTORUN may:

• apply existing learned thresholds
• update confidence scores from explicit downstream feedback

AUTORUN must not:

• override stable heuristics from tiny samples
• silently change team-wide thresholds without surfacing it

Launch Integration

Use Launch data to learn:

• merged PR outcomes
• review turnaround
• split strategy success
• squash strategy effectiveness across PRs