Trust Model
Host Ownership
What EvalGlass owns, and what you own.
This split is what keeps a green result honest. Because the framework never writes its own gold answers, approves its own thresholds, or calibrates its own judges, it can never quietly grant itself the authority to gate. Authority only ever enters from a host-owned record you wrote and reviewed.
The ownership line
| EvalGlass owns — managed framework | You own — host truth |
|---|---|
| The effect-free Evaluation Core: score meaning, aggregation, authority resolution, and the Verdict Engine. | Datasets and references (the gold), and their validated / proposed / retired status. |
| The Runtime Harness that reads files, runs replays, calls judges, and persists results. | Rubrics, judge calibration records, and approved thresholds. |
| The EvalGlass Skill that vendors managed files into your repo at integration time. | Baselines and the decision to promote one, plus host evaluator code. |
| The vendored runtime under evals/_evalglass/. | Data policy decisions and every domain judgement about whether an output is acceptable. |
Generated does not mean approved
The Skill can scaffold host-owned files to get you started, but it deliberately scaffolds them as non-authoritative. Nothing the framework generates can gate until you validate it.
- Scaffolded gold lands as proposed, not validated.
- Suggested thresholds land as proposed, not approved.
- Judge rubrics land uncalibrated.
- The authority.json record starts empty, so a first run is informational.
Runtime outlives the tooling
Because authority lives in host-owned files and the managed runtime is plain vendored code, the evaluation keeps working after the Skill and any coding-agent context are removed. The framework is an instrument you own a copy of, not a service you depend on.