wyrd

name	wyrd
description	Hypothesis validation and closed-loop learning
user-invocable	true
disable-model-invocation	true
allowed-tools	["Read","Write","Glob"]

Wyrd

Hypothesis validation and closed-loop learning.

"What will be is what survives the test."

Usage

/wyrd                     # Show current confidence state
/wyrd calibrate           # Recalculate from rejection history
/wyrd test hypothesis.md  # Validate hypothesis against codebase
/wyrd learn               # Extract patterns from recent rejections

Philosophy

Wyrd (Old English: fate, destiny) reveals truth through testing.

• Fate emerges through testing — Truth is revealed, not assumed
• Rejection is data — Each "no" teaches the system
• Confidence calibrates — Predictions improve with feedback
• Reality anchors claims — Validate against real code and tests

Workflow: Show State (/wyrd)

1. Read grimoires/rune/wyrd.md
2. Display confidence calibration table:

   ## Wyrd State
   
   | Effect | Base | Adjustment | Current |
   |--------|------|------------|---------|
   | Financial | 0.90 | -0.05 | 0.85 |
   | Destructive | 0.90 | 0.00 | 0.90 |
   | Standard | 0.85 | +0.05 | 0.90 |
   | Local | 0.95 | 0.00 | 0.95 |
   
   ### Learning Metrics
   - Total hypotheses: 47
   - Validation rate: 78%
   - Rejections this sprint: 2
   

3. Show active hypotheses if any
4. Show recent pattern influences

Workflow: Calibrate (/wyrd calibrate)

1. Read grimoires/rune/rejections.md
2. Count rejections by effect type with decay:
   • Full weight: 0-7 days
   • Half weight: 8-30 days
   • Quarter weight: 31-90 days
   • Zero weight: 90+ days
3. Calculate adjustment factors:
   • -0.05 per weighted rejection
   • Cap at -0.30
4. Update grimoires/rune/wyrd.md

Workflow: Learn (/wyrd learn)

1. Read recent entries from grimoires/rune/rejections.md
2. Detect patterns (3+ similar rejections):
   • Same effect type
   • Same change direction (e.g., timing reduction)
3. Promote detected patterns to grimoires/rune/patterns.md
4. Optionally create Sigil taste entries:

   Pattern detected: 3 timing reductions for Financial effect
   Average adjustment: 800ms → 520ms
   
   Create taste entry? [y/n]
   

Confidence Formula

confidence = base_confidence + taste_adjustment + rejection_adjustment

Where:
- base_confidence: Effect-specific default
  - Financial: 0.90
  - Destructive: 0.90
  - Standard: 0.85
  - Local: 0.95

- taste_adjustment: +0.05 per Tier 2+ taste match

- rejection_adjustment:
  - -0.05 per similar rejection in last 30 days
  - -0.10 if same component rejected before
  - Cap at -0.30

Integration with Other Constructs

Construct	Wyrd's Role
Sigil	Wyrd feeds rejection learnings → Sigil records as taste
Glyph	Wyrd provides confidence → Glyph shows in hypothesis
Rigor	Wyrd invokes Rigor during self-validation phase

Rules Loaded

Core Rules

• rules/wyrd/00-wyrd-core.md - Philosophy
• rules/wyrd/01-wyrd-hypothesis.md - Hypothesis format
• rules/wyrd/02-wyrd-learning.md - Learning protocol

Confidence & Calibration

• rules/wyrd/03-wyrd-confidence.md - Confidence calculation
• rules/wyrd/08-wyrd-recalibration.md - Recalibration protocol

Rejection & Learning

• rules/wyrd/04-wyrd-rejection-capture.md - Explicit rejection handling
• rules/wyrd/05-wyrd-file-modification.md - Implicit edit detection
• rules/wyrd/06-wyrd-change-analysis.md - Physics change analysis
• rules/wyrd/07-wyrd-pattern-detection.md - Pattern detection algorithm

State Files

File	Purpose
grimoires/rune/wyrd.md	Current confidence state
grimoires/rune/rejections.md	Rejection history (append-only)
grimoires/rune/patterns.md	Extracted patterns

Learning Pipeline

User Rejection ("n")  ──┐
                        ├──→ rejections.md ──→ Pattern Detection ──→ Tier Promotion
User Edit (implicit)  ──┘
                                    │
                                    ↓
                            3+ similar rejections
                                    │
                                    ↓
                            patterns.md → taste.md (Tier 2)

File Modification Monitoring

After Glyph generates a file:

1. Start 30-minute monitoring window
2. Detect physics-relevant changes via git diff
3. Prompt: "Record as taste? [y/n]"
4. Log to rejections.md (implicit_edit type)
5. Check for pattern detection