Three Borrowed Shells
I stopped three plausible conclusions from carrying more weight than their evidence allowed.
Evidence-aware AIProduct judgmentQuality diagnosis
An external insight, a source match, and a clean metric each tried to carry a stronger conclusion than the evidence allowed. I rejected additive scope, replaced token overlap with traceable source-span grounding, and audited 175 claims. The audit returned zero quoted claims, but mostly summary-level evidence packs meant the number could not support a verdict about the whole product. Relevance is not provenance, and a real symptom is not automatically a diagnosis.
Relevance ≠ admissionOverlap ≠ provenanceSymptom ≠ diagnosis
Evidence tierPublic reasoning + repository ADRs + bounded audit claim
Boundary: The bounded source-excerpt policy was a downstream consequence later admitted through ADR-0011, not part of the original three-shell judgment.
A Constraint Is Only as Strong as Its Layer
I turned a written rule into an enforced boundary that prevents an unverified manual entry from becoming an ordinary Project Takeaway.
Agent governanceEnforcement architectureSystem design
The repository path is code-traced and test-supported: an ordinary Project Takeaway request is classified and evaluated before persistence, and policy failure returns HTTP 400. The boundary is deliberately source-aware rather than one-size-fits-all. Knowledge convergence remains review context; insufficiently verified signal completion is marked unverified; and manual override requires a separate, explicit, audited exception.
1Documented
2Reviewed
3Constructed
4Sandboxed
Evidence tierCode-traced + test-supported
Boundary: This claim is limited to ordinary Project Takeaway candidate creation. It is not a claim that every invalid write or downstream commitment is blocked.
A Gate You Can Pass Without Understanding Isn't One
I found a comprehension gate that could be passed without comprehension, located the same assumption in my system, and did not rush to build a feature.
Human-in-the-loop evaluationGate diagnosisEpistemic restraint
Reported answer-pattern leakage made an explain-diff quiz passable without demonstrating comprehension. That exposed a distinction in AI Radar: artifact gates can inspect claims, evidence, and inference while still assuming the admitter is competent to judge. Comprehension may resist reliable proxy measurement, so the result remains tracking-only until multiple admitters or operator load makes the failure mode real.
Green quiz→Reported comprehensionGreen quiz↛Demonstrated comprehension
Evidence tierExternal specimen + public reasoning + tracking-only validation record
Boundary: This is a tracked framing gap, not a shipped comprehension feature or a new ADR requirement.