Use CALIBER
Calibration
The practical calibration path: choose the asset family, generate candidates, inspect evidence, and move to explicit review or apply.
Use this page for the practical calibration path: choose the asset family, launch candidate generation, inspect the evidence, and move to explicit human review or apply only where the product supports it.
At a glance
| Task | Start here | Deep reference |
|---|---|---|
| understand whether calibration exists for an asset | check the per-asset model | Calibration architecture |
| generate better candidates | launch the asset-specific loop | Calibration architecture |
| inspect evidence and targets | connect back to evaluation and release | Evaluation and test sets, Review and release flows |
| recover a lost or stuck job | use the operator runbook | Operations runbook |
1. What calibration is for
Calibration is proposal generation backed by evidence. It is not hidden autopilot. Different asset families expose different candidate-generation and apply semantics, so always reason about the target asset first.
2. Common tasks
| You want to... | Read this next |
|---|---|
| calibrate prompts or skills | Calibration architecture |
| understand how evaluation feeds the loop | Evaluation and test sets |
| inspect the live-control boundary after calibration | Review and release flows |
3. Common failure modes
| Symptom | First thing to check |
|---|---|
| Calibration produced a candidate but nothing went live | apply/release is a separate control path |
| The candidate improved one slice but hurt another | the evidence set or judge mix is incomplete |
| The job vanished or stalled | use the runbook recovery path instead of retrying blindly |