A milestone worth marking: the first contract-review assistant we built has reached production. And true to the principle we lead with, it shipped with its evidence attached — not a demo and a promise.
Here is what “evidence attached” meant in practice for that release:
- Groundedness 0.96 — every answer traceable back to its cited sources, not the model’s imagination.
- Eval suite 48 / 50 — the gate held at threshold; two known edge cases were flagged rather than waved through.
- Guardrails: green — policy and safety checks passed on every output, with refusals logged.
- Two cases to a human — low-confidence outputs routed to a reviewer before anything was delivered.
None of those numbers matter much on their own. What matters is that they travelled with the release, visible to the client, and that the gate would have stopped the ship if they hadn’t held. That’s the difference between a system you can stand behind and one you can only hope about.
It’s an internal milestone, not an award — we’re marking it because it’s the first time the whole glass-box discipline ran end to end on a live engagement. The next one starts the same way: scope in writing, evidence attached, and control kept by you.