Solutions
A release gate for every team shipping agents.
From a single prompt change to a whole RL checkpoint, add a clear ship / limit / block decision before changes reach users.
One layer, every change
Every prompt, tool, or model change is a release.
Whoever ships it, the question is the same: does this hold beyond the tests we can see — and did it earn the score or game the verifier? We answer it the same way for every team.
Agent product teams
Ship prompt and tool changes without guessing
Every prompt tweak, model swap, or new tool is a release. Run it through the gate first for a clear ship / limit / block call — not a vibe check.
- Catch regressions a public eval can't see
- Block changes that overfit before users feel them
- An immutable, reviewable record behind every release
Assurance Card · example
- decision
- BLOCK
- reasons
- ood_regressed · hidden_regressed
- public
- 0.740 → 0.910
- hidden
- 0.732 → 0.611
- ood
- 0.701 → 0.488
- record
- redacted · reviewable
✗ candidate not promoted
RL & fine-tuning teams
Certify the verifier before your model games it
GRPO and RFT reward the shortcut the moment one exists. Certify the verifier's gameability up front, then re-grade every checkpoint without an LLM judge — Isomorphic Perturbation Testing (IPT) catches solution-level hardcoding, a trajectory scan catches test/grader tampering.
- Drops into your RLVR/RFT reward as a deterministic grader — no LLM judge
- Re-checks wins on unseen task variants + structural hardcode detectors + trajectory tamper scan
- Runs in your trust domain — reference and tests stay local
Eval, safety & enterprise teams
Prove what shipped, without exposing what's private
Give security and compliance an immutable, redacted record for every release — with a named-human sign-off on high-risk limits.
- Redacted assurance card on every decision
- Private evaluation boundaries; human approval gate
- Bring-your-own-key and private deployment paths available
What every team gets
The same release gate, wherever you ship
Gate policies
Declare which checks block, limit, or stay advisory — and a LIMIT on a high-risk agent waits for a named-human sign-off before anything ships.
CI & training-loop gate
Call the API, drop the grader into your reward loop, or run the GitHub Action to fail the build on a shortcut.
Works with your stack
Runs in your trust domain with no LLM judge in the loop — bring your own models, frameworks, and keys.
