Repository navigation
Evaluation
AgentCannabis keeps different kinds of evidence separate. A validator can prove repository structure, an install check can prove package availability, and a captured evaluation can show how a bounded workflow behaved on a given fixture. None of these alone proves current law, site applicability, or professional approval.
Run the release gate from the repository root:
.\scripts\validate-all.ps1The wrapper runs structural validation, source-link validation, and the deterministic review-gate test suite. Preview public package metadata with:
.\scripts\publish_skill_repository.ps1 -DryRunEvaluation reports live under docs/development/evaluations/:
-
CC-10-forward-test.md— five reference skills and captured output review; -
CC-35-integrated-workflows.md— integrated workflow cases and scoring; and -
CC-36-adversarial-evaluations.md— adversarial boundary cases and scoring.
Scenario files inside skill packages are specifications. They are not execution evidence until outputs are captured, scored, fixed where needed, and retested. The reports describe their fixture scope, reviewer method, and limitations.
The public release gate verifies all 18 professional skillsets with
gh skill install from a release tag in fresh destination projects. Install
success proves packaging availability and discoverability; it does not prove
legal correctness, current source applicability, or future model behavior.
Release v1.0.5 records passing structural validation, source records,
deterministic review gates, root routing templates, GitHub skill packaging dry
run, and all 18 public professional skillset installs.
When reviewing a report, check:
- whether the fixture is fictional and the inputs are complete;
- which sources were available and what dates or applicability limits they carried;
- whether the output preserved unknowns and escalated regulated or hazardous decisions; and
- whether the result was reproduced after a fix.