Run evidence
Observed and managed Runs retain visible events, artifacts, outcomes, corrections, and observation gaps.
Read-only Hugging Face mirror
Turn completed Runs into reviewable capability changes.
A local-first workbench for retaining Runs, diagnosing evidence-supported issues, comparing one bounded AGENTS.md or Skill change, and publishing or rolling it back with human approval.
- Release / status
- main includes 0.3.0 protocols; latest packaged preview is 0.2.0
- Source snapshot
5beb058d946a- Runtime
- Static product tour; full Fastify/SQLite product is local only
What the project owns
The workbench closes the loop after a Run without becoming a second Agent or granting the Agent approval authority.
Bounded diagnosis
Issues keep instruction, Skill, tool, environment, permission, validation, model, and unknown causes distinct.
Governed evolution
Four-cell comparison, exact diff, explicit approval, conflict-safe publication, and rollback history.
Operational sequence
The sequence is explanatory. Install and execute the actual project locally from its canonical repository.
01
Retain
Keep the completed Run and its visible evidence gaps.
02
Diagnose
Connect one recurring issue to exact evidence and counterevidence.
03
Compare
Evaluate one bounded capability-file candidate in isolated worktrees.
04
Approve or reject
A human reviews the exact diff and objective verifier cells.
05
Publish or roll back
Hash checks prevent overwriting a target that changed concurrently.
Product view
Synthetic or presentation-only images from the canonical repository.
Evidence boundary
This page preserves the project's published limits instead of turning engineering checks into outcome claims.
Local-first
The full service binds to loopback and stores product data locally.
No hidden reasoning
Only locally exposed events are retained; gaps remain explicit.
Bounded study
The Wiki study is three replicates, one synthetic grader, and two model versions in one family.
Inspect the artifacts
Local files below are curated public mirrors. GitHub remains authoritative for development, releases, and issue history.