Evidence boundary
Skill Gesondian/ai-collab-governance-skills/skills/evidence-boundary
Hermes Loop Engineering skills for self-improving AI delivery governance.
npx -y skills add Gesondian/ai-collab-governance-skills --skill evidence-boundaryAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Use when an AI delivery conclusion depends on partial, indirect, inherited, mock, local, visual, or otherwise limited evidence.
SKILL.md
2.5 KB, 513 tokens by cl100k_base, as published. Nobody here has run it
Evidence Boundary
Core Principle
Evidence is a boundary, not a slogan. A conclusion may only cover what the evidence directly proves.
Evidence Classes
| Evidence | Can Support | Cannot Support |
|---|---|---|
| Route reachable | The route loads under tested conditions. | The workflow is usable or business-correct. |
| Single surface check | That exact surface passed the stated assertion. | All similar surfaces passed. |
| Mock data render | The UI can render a synthetic shape. | Real business rules, permissions, or data lifecycles work. |
| Static scan | Certain text or code patterns are absent or present. | Runtime behavior is correct. |
| Visual screenshot | Layout at one viewport and data state. | Keyboard flow, permissions, or hidden states. |
| Prior receipt | What was previously asserted by that receipt. | New or uncovered behavior. |
Procedure
- Name the claim being made.
- List the evidence references.
- For each evidence reference, state:
- scope covered,
- assertion level,
- fact source,
- freshness,
- blind spots.
- Downgrade any claim that exceeds the evidence.
- Record missing evidence before assigning acceptance or owner conclusions.
Supported Claim Levels
| Claim Level | Use When |
|---|---|
observed | A concrete signal was seen, but no broader assertion is made. |
route_reachable | Navigation reached the tested route. |
surface_checked | A named surface passed named assertions. |
workflow_checked | A multi-step workflow passed with stated data and role assumptions. |
accepted_scope_limited | Acceptance is bounded to listed dimensions only. |
not_supported | Evidence is insufficient for the proposed claim. |
Output Contract
claim:
evidence_refs:
covered_scope:
unsupported_scope:
missing_evidence:
safe_wording:
must_not_claim:
Common Mistakes
- Treating reachability as usability.
- Treating one successful example as class-wide proof.
- Treating mock rendering as real data correctness.
- Inheriting a conclusion from another receipt without checking its scope.
- Saying a capability is shared merely because a shared-looking file exists.
Safe Wording Pattern
Use: "Evidence supports X under Y conditions; it does not yet support Z."
Avoid: "The feature works" when only one route, data shape, viewport, or assertion was checked.
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.