Trust Model
Provenance
What was measured, exactly.
A score without provenance is not interpretable: you cannot tell whether two runs differ because the model changed or because the metric, dataset, or config did. EvalGlass hashes each dimension — a SHA-256 of its canonical JSON — so “what produced this number” is recorded, not remembered.
The ten dimensions
Each dimension is fingerprinted independently. Six are gating: they must match for two runs to support a regression claim. The other four are recorded for inspection but do not block a comparison.
| Dimension | Captures | Gating? |
|---|---|---|
framework | The EvalGlass version that ran. | Yes |
metric_spec | The metric’s declared meaning. | Yes |
evaluator | The evaluator reference and version. | Yes |
dataset | The dataset identity, status, and version. | Yes |
config | The runtime configuration. | Yes |
policy | The data policy in force. | Yes |
example | The specific example measured. | No — different examples still compare. |
evidence | The evidence inputs collected. | No |
authority | The authority inputs at run time. | No |
baseline | The baseline reference, if any. | No |
fingerprint_dimension raises on a value that is not JSON-serializable rather than
guessing. A fingerprint can therefore only ever under-claim comparability, never over-claim it.
The full fingerprint and the per-dimension hashes are written into the
RunRecord, so any later run — or a reviewer — can recompute and check them.