No material speed gain: both candidates missed the predeclared 10% E2E
requirement against both controls. Closed by retaining the original
configuration; all negative and variable-output measurements remain visible.
Private identity: public reproducibility is limited deliberately. Closed
for publication with explicit redaction and no named-checkpoint qualification.
Protected originals and a private source mapping remain intact.
Observe-mode workload: absent canaries, variable output and no matched
no-speculation arm prevent controlled speculative-speedup claims. Future
finalist work would require those controls; no further live experiment is
authorized by this publication.
Recovery-runner gates: the isolated trial runner gained resource and
process/speculation checks before the DFlash run, with independent review and
44 focused passing tests. Those source changes are not shipped here.
Runtime inspection/download output: bounded read-only installed-source
inspection was needed; the managed download command mixed progress with its
JSON result. These tooling gaps are retained in protected trial records and
are deferred from this documentation change. No unsupported lifecycle command
was introduced as a published workflow.
Restoration: completed after each experiment with exact private identity
checks and direct/routed postchecks. No candidate promotion remains pending.