Measured baseline; candidate admission refused

The local R12 incumbent produced 44 graded completed jobs: 38 passed and 6 failed their deterministic tasks. Four more jobs were cancelled at the bounded traffic deadline; their grades remain unknown. Released intervals closed with retained engine, GPU, host and cgroup telemetry. R13 preserved a separate 48-cell candidate plan but made no benchmark submissions: the managed 49 GiB RAM plus 16 GiB reserve check refused container creation. Exact refusal-time available memory is unknown.

No candidate performance, model-quality comparison, engine-causal result or promotion is supported. Exact incumbent restoration and routed checks passed, followed by one explicit minimal-thinking Pi invocation with 29/29 independent checks. The historical thinking-off failure remains open; an initial reviewer accessor false negative was corrected offline without a live retry. This summary derives from the retained evidence; final inventory and size approval are bound by the publication PR.

Claim ledger: native R12 denominators; closure and cancellation; zero-submission R13 report; admission bounds and unknowns; prospective exclusions; authoritative summary and limitations; scoped restoration and independent Pi validation facts.