Found during 0.24.0 release testing. Reproduces on published 0.23.11 — pre-existing.
The Artifacts/Observations line in check (and secure) prints the raw classifier intent label next to benign artifacts:
- benign hardened SOUL →
Artifacts SOUL.md soul · malicious · no inferred capabilities (1 constraint) while the top-line score is 94/100
- benign clean skill →
SKILL.md skill · suspicious · …
- output of the tool's own
create-skill → SKILL.md skill · malicious · …
The classifier 0.5.0 is documented (semantic-compiler notes) as over-flagging benign/OOD input at max confidence, and the established rule is that its raw verdict is never surfaced standalone — the benign-context override / analyst escalation path exists precisely for this. The Observations line bypasses that: a reader sees "malicious" printed against a clean artifact that the same output scores 94–98/100, a direct internal contradiction.
Fix direction: the Observations line should print the post-override assessment (or the artifact type only), never the raw intent class. If the raw class is wanted for debugging, gate it behind --verbose with an explicit "raw classifier affinity (over-flags benign)" qualifier.
Found during 0.24.0 release testing. Reproduces on published 0.23.11 — pre-existing.
The Artifacts/Observations line in
check(andsecure) prints the raw classifier intent label next to benign artifacts:Artifacts SOUL.md soul · malicious · no inferred capabilities (1 constraint)while the top-line score is 94/100SKILL.md skill · suspicious · …create-skill→SKILL.md skill · malicious · …The classifier 0.5.0 is documented (semantic-compiler notes) as over-flagging benign/OOD input at max confidence, and the established rule is that its raw verdict is never surfaced standalone — the benign-context override / analyst escalation path exists precisely for this. The Observations line bypasses that: a reader sees "malicious" printed against a clean artifact that the same output scores 94–98/100, a direct internal contradiction.
Fix direction: the Observations line should print the post-override assessment (or the artifact type only), never the raw intent class. If the raw class is wanted for debugging, gate it behind --verbose with an explicit "raw classifier affinity (over-flags benign)" qualifier.