Key terms¶
| Term | Meaning |
|---|---|
| Agent | A service that may propose SQL from a natural-language question. |
| Baseline | The version that already works. |
| Candidate | The version that needs a release check. |
| Context | The metadata or prompt content that changed. |
| Dataset identity | A content-based description of the exact data used in a run. |
| Evidence | A fact returned by a trusted runtime or metadata source. |
| Independent oracle | A separate calculation used to check the agent result. |
| Lineage | A record of how data moves between assets. |
| Manifest | The YAML or JSON input that describes one release run. |
| MCP | The tool boundary used to read DataHub evidence. |
| OpenTelemetry span | One timed unit of work in a trace. |
| Receipt | A redacted record of the run and its decision. |
| Replay case | One test question with expected behavior and an oracle. |
| Snapshot | The frozen data object used for a reproducible run. |
| Trace ID | The ID that links spans from one operation. |
PROMOTE |
Required checks passed. A human may consider release. |
REVIEW |
Evidence is incomplete or a required check needs human review. |
QUARANTINE |
A critical check failed. Do not release the candidate. |
Evidence versus explanation¶
An explanation says what should be true. Evidence says what the running system returned.
For example:
- “The service should emit a trace” is an explanation.
- “The receipt contains this trace ID from the OpenTelemetry SDK” is evidence.
Themis records the second kind. It does not invent a trace ID.