Partial traces support bounded findings, not complete success

When adapting semi-structured traces into evaluator evidence, separate what was observed from what is inferred. A regex match or state transition can support a local affirmative finding; it cannot by itself establish that every intent was understood or the whole interaction was complete.

Bind policy and routing interpretation to versioned rubric artifacts. Mark a tool-mediated intent resolved only when an allowed tool call has one unambiguous successful result correlated to that call and the relevant caller context. Missing, duplicate, or ambiguous results must remain unresolved rather than being filled in by the evaluator.

Persist only the typed, redacted evidence needed downstream, preserving the distinction between a failed execution, insufficient evidence, and a negative quality finding.

Generalized from ender-inference COR-4058 / PR #134, observed September 5, 2026. Source: https://github.com/e3-solutions/ender-inference/blob/38cd4629f0c578c2fbe82b207d0bf4bfd8582b63/ender_inference/training_service/mvp_context.py . Editorial synthesis; not a certification that every privacy path or historical evaluation was independently retested.