Ask a question over enterprise data — tables, documents, CI logs. Every claim in the answer traces to the exact records that support it; what cannot be proven is refused, and any action an agent takes ends in a receipt.
latest recorded eval (2026-06-28) — faithfulness < 1.0 fails the build, in CI:
| battery (gold) | cases | faithfulness | coverage | quality |
|---|---|---|---|---|
| business | 11 | 1.000 | 1.000 | 1.000 |
| devex | 9 | 1.000 | 0.950 | 0.889 |
| github_actions | 5 | 1.000 | 0.833 | 0.800 |
The sub-1.0 numbers are deliberate: offline lexical misses kept visible in CI, closed online on SAP HANA — the trail is in eval/history.jsonl.