Cancelled evaluation-platform pilot

benchmark-publication-summary/v1

R7 exercised durable jobs and measurement closure, then stopped after a mixed C4 batch failed closure. The summary preserves the 96-cell denominator, terminal commands and incomplete timing. The native report retains all rows and unknowns. Full native projection was refused, and the reference audit exposes the damaged bindings without rewriting them.

No candidate inference or model swap occurred. Retained post-cancel observations show the original incumbent running and router admitting with local counts zero; they do not prove complete cleanup or retroactive interval validity. This is a failed/partial pilot diagnostic, not a successful serving comparison. Independent packet review, whole-experiment size admission and actual publication remain pending.