Calibration

Estimates improve with feedback. Calibration is the process of comparing estimates to actuals and adjusting your defaults.

Research shows calibration feedback improves estimation accuracy from 64% to 81% (Jorgensen & Grimstad, 2004).

The Feedback Loop

Estimate → Do the work → Log actual → Compare → Adjust → Repeat

1. Estimate before you start (using this skill)
2. Track how long it actually took
3. Log estimated vs actual in a consistent format
4. Compare patterns across 10+ data points
5. Adjust the specific multiplier that's off
6. Repeat — each cycle improves accuracy

What to Track

After completing a task, log:

Field	Estimated	Actual	Delta
Task	Auth service	—	—
PERT Expected	4 hrs	—	—
Actual Total	—	4.8 hrs	+20%
Within PRED(25)?	—	Yes	20% < 25%

Adjustment Rules

After 10+ data points:

Pattern	Adjustment
>60% tasks exceed committed	Increase confidence multiplier or base rounds
>60% tasks below expected	Decrease base rounds or risk coefficient
Human review consistently over	Increase review_minutes values
One task type consistently off	Adjust that type's multiplier
One complexity band off	Adjust that band's rounds only

Change incrementally: ±0.05 on ratios, ±2-3 on rounds, ±0.1 on multipliers. Never adjust more than 20% in one cycle.

Calibration Cadence

PRED(25) Score	Cadence
< 50% (poor)	Weekly
50-65% (developing)	Bi-weekly
65-75% (good)	Monthly
> 75% (excellent)	Quarterly

See PRED 25 for the full accuracy metric. See Reference Stories for calibration anchors.

Sprint Velocity

The calibration system includes mode-aware sprint velocity tracking. See Sprint Velocity for the full guide on:

• Logging sprint data per cooperation mode
• Capacity planning formulas
• Sprint fit checks
• Divergence detection for hybrid teams
• Velocity trend analysis

Full details in calibration.md.