Company open source
Action State open source · Apache-2.0
Terms Pack compiler
Part of EvidenceBook
Obligations & Outcomes to Actions
Records tell you what happened. They don't tell you whether it was good. This project closes that gap.
You describe what you're obligated to deliver, or what outcome you're trying to achieve, in ordinary language: "resolve the customer's request in one interaction," "never pressure an employee," "close critical vulnerabilities within the window." The compiler turns that into a small set of evaluation axes with frozen criteria, wires them to your capsule log, and generates a skill that runs every day, judges each case, and seals the results. A second skill aggregates the daily results into a monthly report.
Where an outcome can be counted, it's counted. Where it must be judged, an AI judges under a disclosed method, and humans calibrate a blind sample every week. The output is a report your customer, your auditor, or your own team can open, drill into case by case, and verify was never altered.
What's in the repo
- →The Terms Pack compiler — obligations and outcomes to evaluation axes to skills
- →Reference evaluation and aggregation skills
- →The report format, itself sealed as capsules
- →A worked example on a public customer-service benchmark (an airline support agent)
- →Profiles for common obligation types, growing over time
License & governance
Apache-2.0. Maintained by Action State. Contributions welcome.
This project is company open source and is not foundation-bound. That's a deliberate choice, not a shortcoming — it lets us move the tooling forward while the standards it sits on top of move to independent governance.
View on GitHub ↗
Worked example — regulatory obligations
QuadX AI mapped UK BNPL / consumer-credit obligations (the FCA Handbook) to verifiable evidence artifacts — grading each obligation by who can actually check it: operator-asserted (you take the firm's word), participant-verifiable (a party to the transaction checks it cryptographically), or third-party-verifiable (a stranger, such as an auditor or a regulator, verifies it from the artifact and a public anchor, trusting neither the agent nor the operator). It cites the same independent cross-verification runs as the interop index above.
The FCA clause mapping is QuadX AI's FCA Obligations Mapping v1.0 (22 July 2026): QuadX AI reviewed and corrected the Action State Group strawman and verified the FCA clause references against the FCA Handbook as at July 2026. It codes the rules as written, not as enforced — informative, not legal advice and not an FCA position. That's the point of the compiler on a real regulatory surface: "shown met" becomes a checkable claim with a stated depth, not a promise.
Read the FCA obligations mapping (PDF) →