Evidence Model
Contexta treats ML observability as durable evidence rather than loose logs. A run should explain what happened, what inputs were involved, what outputs were produced, and what a reviewer should trust or question.
Truth Planes
| Plane | Purpose | Examples |
|---|---|---|
| Metadata | lifecycle and ownership | project, run, stage, batch, sample, deployment |
| Records | observations over time | metric, event, span, degraded marker |
| Artifacts | files worth reviewing | model, checkpoint, prompt, report, schema |
Evidence Shape
{
"run": "run:fraud.candidate-v2",
"stage": "stage:fraud.candidate-v2.evaluate",
"records": [
{"type": "metric", "key": "f1", "value": 0.91},
{"type": "event", "key": "dataset.loaded", "message": "Loaded fraud.v12"}
],
"artifacts": [
{"kind": "model", "path": "artifacts/fraud-v2.joblib"}
]
}
Why This Matters
Logs say something happened. Evidence helps answer whether the result is reviewable, comparable, reproducible, and recoverable.