Skip to main content

ADR 0051: Evaluate Authority At The Proposed Operation

Status​

Accepted.

Context​

STAGE 0.6 S3 originally said that an affected mutation must stop when “artifact ownership” is unresolved. That wording could be read at whole-file scope even when a human explicitly authorized one exact field repair, a canonical source named its replacement, and the operation preserved every other byte. It could also allow the opposite mistake: treating authority for one field as authority to regenerate the containing artifact.

The safety-boundary ablation made the ambiguity concrete. Its evaluator declared every mutation to a mixed-provenance prefab unsafe, while all three agents made the same source-backed material repair and preserved the uncertain pose field. Condition-blind review classified the evaluator, not the repairs, as construct-invalid.

A preregistered follow-up corrected the evaluator and separated the two directions across character, UI, and audio artifacts. All eighteen runs made the exact narrow repair and refused broad regeneration. Project-only guidance, the frozen Safety Kernel, and full delivery were all correct. The full skill used more mean activity without a correctness advantage.

Primary access-control and provenance models support a clearer distinction but do not prescribe STAGE's rule. NIST ABAC includes the requested operation in an authorization decision; least privilege limits authority to the assigned task; W3C PROV describes derivation and attribution rather than granting edit permission.

Decision​

  • Evaluate S3 at the scope of the proposed mutation, not at an undifferentiated whole-artifact label.
  • Stop when authority for that mutation or the state it can overwrite is unresolved.
  • Permit a separately authorized, canonical-source-backed narrow operation when it preserves unresolved adjacent state.
  • Never widen authority for one operation into adjacent mutation, normalization, replacement, or regeneration.
  • Require both explicit handling authority and a complete canonical source for broad work that can overwrite mixed-provenance state.
  • Keep project-local contracts primary. This replication does not show that a STAGE treatment improves outcomes when the project already states the boundary clearly.

Consequences​

  • Fail-closed behavior now includes both safe action and safe refusal.
  • Agents need not choose between blanket inaction and unjustified whole-file ownership.
  • Provenance remains material evidence for authority and preservation without becoming authorization by itself.
  • Compact projects can express this boundary locally without adopting STAGE artifacts or loading the full delivery skill.
  • Historical treatment files remain frozen evidence and retain the old 0.6 wording used during their experiments.

Evidence Boundary​

The decision is supported by an evaluator failure, primary-source conceptual lineage, and one eighteen-run synthetic replication across three artifact families on one model and host. It clarifies the method's construct. It does not establish a general treatment effect, productivity benefit, independent adoption, or scientific proof.

Last updated on