Reference: the decision system in detail
| Chapter | Use it for |
|---|---|
| Architecture | policy as data, the engine, evidence, the read models and the alternatives that were rejected |
| Approvals | the three approved actions, the token, the displayed premise versus the recomputation |
| Limitations | known gaps, including model-misreported premises on approval prompts |
These documents stay on GitHub at the pinned revision. They are operational, or long measurement logs whose lessons the case studies already teach:
docs/EVALUATION_ANALYSIS.md: a dated evaluation log (one model, one date);docs/DEMO.md: a scripted demo against a running stack;docs/SECURITY.md,docs/EVALUATION.md,docs/OBSERVABILITY.md,docs/VERIFICATION.md;docs/UPSTREAM.md: how this repository derives fromsimple-agent-template.
Note
Two claims in the externally linked documents are broader than the code. Read them with the lessons in mind.
SECURITY.mdandDEMO.mdsay a past decision "can be reproduced", but lesson A6 explains why it cannot fully be: versions are hand-maintained labels, and the fund facts used are not recorded. Fund facts are also refreshed from the data file at every boot without a history row.