A services-agreement RL environment.
An environment contract for revising one versioned services agreement against a bounded playbook, matter facts, governing law, and output requirements.
Design premise
The environment must distinguish a completed review from a convincing draft.
The verifier has to check the trace, final artifacts, and task state together.
Environment contract / v0.1
What the task contains.
These are the fields the implementation and test fixtures still need to make concrete.
Task + legal boundary
Represent the customer or provider / review one versioned services agreement / bounded record / governing law + as-of date
State + observations
Agreement / matter facts / versioned playbook + fallbacks / current redline + issues list
Actions + transitions
Read / search the bounded record / compare / redline / annotate / submit / each edit updates the document and trace
Reset + termination
Reset to the frozen start state / stop on submit, action limit, or timeout
Output artifacts
Redline + issues list / each issue links to the changed language or an explicit decision not to change it
Reward + held-out measures
Training reward from artifact checks and a calibrated rubric / frozen held-out measures stay outside optimization
Verifier tests
Known-good / known-bad / alternative-correct / exploit cases / redline and issues-list consistency
Version + split + provenance
Task, environment, agreement, playbook, and corpus versions / train, dev, and held-out fixed before rollout
Open work
What has to exist before release.
- Frozen task fixtures for both represented-party positions
- Known-good, known-bad, alternative-correct, and exploit cases
- A calibrated expert rubric with false-pass and false-reject review
- Baseline runs across named models
- Failure analysis, limitations, versioning, and release terms
This design note contains no benchmark or training result.