Scope
This supporting result comes from the completed path-geometry study, outside the three cohorts emphasized in the paper. The confidence gate was fixed before transfer. It was not fitted to the path-study outcomes.
TypeSafe documents confidence as a statistic derived from the returned probability distribution. It is treated here as a distribution summary and ranking signal. It is distinct from the selected-option probability and is not assumed to equal the probability that an answer is correct.
Observations
The path evaluation contains 1,728 planned root cases from 96 forests under 18 views, with one unavailable outcome. Serving copies are excluded. A fixed threshold of at least 0.9 accepts 261 cases and is correct on 216, leaving 45 errors. Coverage is 15.10%; accepted accuracy is 82.76%.
The reported 95% intervals are 13.31–17.19% for coverage and 72.91–92.62% for accepted accuracy. They resample whole forests and preserve the dependent views. They are descriptive and not simultaneous. Accepted cases are a selected subset; their accuracy is not the unconditional workload accuracy.
A separately learned policy rejects all cases because no development threshold met its prespecified accuracy, case-count, and world-count constraints. Zero coverage has undefined accepted accuracy. It does not demonstrate reliable automated decisions.
Interpretation
A confidence policy needs validation on the intended workload, including its representation and evidence-selection procedure. Comparison with another cohort does not isolate representation as the only difference. Oracle-selected paths, generated task families, and differences in accepted-case difficulty all limit generalization.
The historical source is round3/representation/confidence-gate-path-report.md. The evidence download includes the full 1,824-call path log, frozen plan, gate, protocol, and companion analysis under data/confidence/. It reproduces the gate evaluation without refitting. The 96 serving-repeat calls are retained in the archive and excluded from the 1,728-case threshold analysis. This supporting summary is outside the three-study main inventory.