Trust and governance
A curated reader path across the trust surfaces that connect evidence, calibration, review, and apply.
This page connects the major trust surfaces in CALIBER so a reader can understand how evidence, review, and apply fit together without starting from deep subsystem architecture.
At a glance
| Surface | What it answers | Start here |
|---|---|---|
| Test sets | What evidence are we scoring against? | Evaluation and test sets |
| Evaluation | How is quality measured? | Evaluation and test sets |
| Calibration | How are better candidates proposed? | Calibration |
| Workflow/runtime approvals | Who may allow a live run to continue? | Workflows |
| Artifact apply / release | Who may move a candidate live? | Review and release flows |
| QA plan | How do runtime QA and engineering validation relate? | QA plan |
1. Evidence comes first
CALIBER is designed around measured change, not silent mutation.
The practical reader path is:
- define or inspect the evidence set
- run evaluation or calibration
- inspect the result
- apply or release only through the explicit control path for that asset family
2. Evaluation is not the same as apply
Evaluation tells you how a candidate performed. It does not make the candidate live by itself.
That distinction matters because different asset families have different live control paths:
- prompts use explicit alias release semantics
- workflows have their own apply/review path
- runtime approvals govern live executions rather than offline candidate motion
3. Calibration is proposal generation, not hidden autopilot
Calibration surfaces generate candidate changes or candidate-ready outcomes using the evidence loop available for that asset family.
Use the deep references when you need the exact execution model:
4. Human authority stays explicit
CALIBER separates:
- evidence generation
- candidate generation
- approval or review
- live apply / release
That is why the governance story must be read across several surfaces rather than only one route family.
5. Recommended reader path
| If your question is... | Read this next |
|---|---|
| What are we measuring against? | Evaluation and test sets |
| How are scores, judges, and results represented? | Evaluation and test sets |
| How does the system generate better candidates? | Calibration |
| How do live workflow decisions work? | Workflows |
| How do engineering tests relate to runtime trust? | QA plan |