Context
Claude Code is currently intentionally Observe.
MARGINAL must not broaden authority merely because an integration exposes a mechanism that appears capable of blocking an action.
Authority must be earned from evidence.
Contribution wanted
Define the measurable evidence requirements that would have to be satisfied before any Claude Code action family could become eligible for Earned Enforcement.
This is not an implementation task for broad enforcement. It is a specification + validation task grounded in real Claude Code behavior.
Required evaluation dimensions
At minimum characterize:
- lifecycle/action coverage;
- interception reliability;
- outcome attribution quality;
- false-positive rate;
- false-stop rate;
- recovery behavior after an incorrect recommendation;
- governance overhead;
- task-quality preservation;
- evidence freshness;
- model/version drift;
- integration/version drift;
- minimum local evidence window;
- promotion conditions;
- demotion/revocation conditions.
Evidence discipline
For each candidate action family, distinguish:
- what MARGINAL can observe;
- what MARGINAL can prove;
- what MARGINAL can safely intercept;
- what remains unknown.
Do not combine these into one capability claim.
Non-negotiable authority rules
- Shared/Commons evidence does not grant local authority.
- Commons observation counts do not increment local promotion counters.
- Popularity is not proof.
- Candidate, supported, validated or promoted Commons evidence cannot bypass local trust gates.
- Unknown or failed outcomes cannot strengthen enforcement authority.
- Missing evidence must fail open.
- Enforcement eligibility must remain scoped to the exact action family for which evidence exists.
- Model/version drift must be able to revoke or reduce authority.
- Integration behavior drift must be able to revoke or reduce authority.
Acceptance criteria
Deliver a concrete, testable specification that defines:
- minimum evidence required before promotion;
- minimum observed coverage;
- acceptable outcome certainty;
- false-stop constraints;
- quality-preservation requirements;
- governance-overhead accounting;
- evidence freshness requirements;
- promotion blockers;
- demotion/revocation triggers;
- behavior when evidence is incomplete or contradictory.
Where possible, encode the requirements as executable tests or machine-readable policy rather than documentation only.
Valid outcome
A valid contribution may conclude that the current Claude Code evidence surface is insufficient for Earned Enforcement.
That is preferable to overstating authority.
Privacy
Do not require raw prompts, source code, command text, tool output, private repository identity or persistent user identifiers as evidence for promotion.
Context
Claude Code is currently intentionally Observe.
MARGINAL must not broaden authority merely because an integration exposes a mechanism that appears capable of blocking an action.
Authority must be earned from evidence.
Contribution wanted
Define the measurable evidence requirements that would have to be satisfied before any Claude Code action family could become eligible for Earned Enforcement.
This is not an implementation task for broad enforcement. It is a specification + validation task grounded in real Claude Code behavior.
Required evaluation dimensions
At minimum characterize:
Evidence discipline
For each candidate action family, distinguish:
Do not combine these into one capability claim.
Non-negotiable authority rules
Acceptance criteria
Deliver a concrete, testable specification that defines:
Where possible, encode the requirements as executable tests or machine-readable policy rather than documentation only.
Valid outcome
A valid contribution may conclude that the current Claude Code evidence surface is insufficient for Earned Enforcement.
That is preferable to overstating authority.
Privacy
Do not require raw prompts, source code, command text, tool output, private repository identity or persistent user identifiers as evidence for promotion.