What we measured.
One synthetic company, a frozen benchmark, and a conventional retrieval baseline answering the identical questions with the same grader. Every figure below comes from a version-pinned artifact in our repository. Ask, and we will send you the file.
| Answer grade vs paired raw-RAG baseline | 0.694 | against 0.639 |
|---|---|---|
| Retrieval recall@5 | 0.972 | against 0.856 |
| Mean reciprocal rank | 0.958 | against 0.870 |
| Permission leaks | 0 | against the baseline’s 1 |
| Running cost per company | ~$2.58 | per month, model spend |
| Prototype built in | 8 sessions | ~390 tests · $4.26 total API spend |
| Tests enforced in CI today | 1,195 | the gate fails any commit below this floor |
Behaviours we can demonstrate on a call.
- Access rules are enforced inside the query, not filtered after generation.
- Contradictory instructions queue for a human. A newer, lower-authority message did not overwrite a signed policy.
- Superseded rules are kept as history and never presented as current.
- Injected content is rejected at a deterministic boundary, before it reaches a model.
- Skills compile from the graph with zero hand editing, and each role genuinely gets a different surface.
- Change a governed rule, recompile, and behaviour changes — with the generated test packs still green.
What we have not shown.
We would rather you hear this from us than find it in week three. None of it is hedging: each line is a thing we tried to establish and did not.
- We have no customers. None. You would be the first.
- We have never run on a real company’s data. Our test company is synthetic, and real exhaust is messier in ways we expect to hurt.
- Six pre-registered attempts at the correction loop, each sealed before its run, and all six are on the record. Two crashed and measured nothing. Three measured NOT MET. The sixth met all eight of its registered thresholds, and what it showed is narrow enough that the scope travels with it: on one synthetic fixture tenant, for novel-Exception corrections, over a correctable pool of three, at n = 2, a governed correction reached a case the system had never been shown. On real company data it has not been tested. We do not claim the system gets better on its own, because that is not what was measured.
- The component that interviews people for knowledge nobody wrote down is not built.
- Arabic answers are weak. Retrieval finds the right record; the answer quality does not hold up.
- We have tested to 3,000 documents, not 300,000.
- No security certification yet. We can describe our controls and show you the architecture; we cannot show you a badge.
- Our connectors are file adapters. A live mailbox connector is not built.