Field note /
Independent review cannot be self-gated by the decision it checks
A semantic decision skill worked in controlled qualification, then made zero calls across six real manager decisions.
How the claim changed
Belief
If an agent is taught when to request independent semantic evidence, the skill can remain optional while still checking consequential decisions.
Evidence
The Jev decision skill activated on controlled fixtures, including Choice and Noul/Score calls with typed evidence.
Contradiction
Across six frozen real manager tasks, the agents made zero Jev calls. A later exploratory comparison found only three of six selected proposals matched the public gold proposal, so optional uptake could not be treated as a reliable safeguard.
Correction
Agent-Workflow now prototypes a provider-neutral semantic checkpoint outside the choosing agent: record the draft first, project neutral primary evidence, invoke the provider, and require explicit reconciliation when evidence disagrees or remains uncertain.
Current claim boundary
This establishes a mechanism for removing semantic self-gating, not evidence that the checkpoint improves correctness. The six known tasks are replay cases only; effectiveness requires a new disjoint preregistered cohort.