Execution Learning System (CALIBRATE)
Purpose: Use this file after execution to compare estimates with reality, update multipliers, and emit reusable planning patterns.
Contents
- CALIBRATE overview
- RECORD data
- COMPARE thresholds
- ADJUST rules
- PERSIST journal format
- Velocity prediction
Overview
RECORD -> COMPARE -> ADJUST -> PERSIST
Calibration keeps Sherpa from repeating static estimates after new evidence appears.
RECORD
Capture after each completed step:
Step: [name]
Estimated: [minutes]
Actual: [minutes]
Size: [XS/S/M/L]
Complexity_Factors: [list]
Risk_Level: [Low/Medium/High]
Agent: [who executed]
Domain: [frontend/backend/infra/test/docs]
Outcome: [clean/rework/blocked]
Notes: [observations]COMPARE
Accuracy Ratio
Accuracy Ratio = Estimated / Actual
> 1.2 overestimated
0.8-1.2 good estimate
< 0.8 underestimatedTarget Range
- target long-run average:
0.85-1.15 - compare patterns across sessions, not just one step
Trend Example
| Session | Avg Ratio | Interpretation |
|---|---|---|
| 1 | 0.72 | many unknowns |
| 2 | 0.85 | improving |
| 3 | 0.95 | stable |
| 4 | 0.92 | healthy |
ADJUST
Base Multipliers
new_technology: 1.5x
unclear_requirements: 1.5x
external_dependency: 2.0x
high_risk: 1.5x
multiple_files: 1.3xAdjustment Rules
- require
3+data points before changing a multiplier - cap each session adjustment at
+/-0.3x - decay toward default by
10%per month - explicit user override beats learned calibration
PERSIST
Record calibration learnings in .agents/sherpa.md.
## YYYY-MM-DD - Calibration: [Project/Epic Name]
**Sessions analyzed**: N
**Overall accuracy**: X.XX
**Key adjustments**:
- [factor]: [old] -> [new] (reason)
**Pattern discovered**: [description]
**Apply when**: [future scenario]
**reusable**: true
<!-- EVOLUTION_SIGNAL
type: PATTERN
source: Sherpa
date: YYYY-MM-DD
summary: [calibration insight]
affects: [Sherpa, relevant agents]
priority: MEDIUM
reusable: true
-->Pattern Library
| Pattern | Typical duration | Common risk | Example note |
|---|---|---|---|
| New API endpoint | 60-90 min |
external dependency | add 1.3x for first endpoint |
| UI component | 45-75 min |
design drift | stabilizes after the second iteration |
| Bug fix | 30-60 min |
unclear root cause | Scout first often saves time |
| Refactor | 60-120 min |
scope creep | strict scope prevents blow-up |
| Test suite | 40-80 min |
flaky dependency | mock early |
Velocity Prediction
Predicted Remaining = sum(remaining calibrated estimates)
Confidence Band = Predicted × [0.8, 1.3]Re-Planning Thresholds
| Current velocity | Action |
|---|---|
> 1.2x estimate |
combine very small remaining steps if safe |
0.8-1.2x |
stay on plan |
0.5-0.8x |
split smaller and add buffer |
< 0.5x |
stop and re-plan |
Quick Calibration
For sessions < 1 hour or < 5 completed steps:
## Quick Calibration
**Steps**: 3 completed
**Avg accuracy**: 0.90
**Action**: no multiplier change yet