Stock-Agent runtime reliability

Multi-symbol fan-out can repeat a stage, accept a late writer from an obsolete run, or resume more work than the failed boundary requires.

Those failure modes make a run harder to audit: a final report can look complete even when its lineage is ambiguous.

Before adding runtime guards, duplicate fan-out, late-write acceptance, and recovery amplification were explicit operational risks.

Historical timings used different inputs and configurations, so this case does not combine them into an improvement percentage.

Runtime fingerprints bind accepted work to the same input and configuration. Generation fencing rejects obsolete writers, while selective resume restarts only eligible failed boundaries.

The evidence path exposes fixture inputs, receipts, a claim registry, and limitations without publishing the private core repository.

The frozen engineering checks cover 64/64 isolation cases, three rounds across three roots, a six-slot race with 6/6 accepted outcomes, and 50 contract-lint cases across 16 waves.

The 64/64 checks above are an integrity-checkable maintainer receipt from private source, not publicly reproducible tests. Separately, the public eight-item synthetic replay is reproducible and key-free. Neither establishes a live-provider or production deployment.

E1 does not prove a live provider canary, independent second-operator handoff, customer adoption, SLA, investment performance, or business impact.

The core Stock-Agent repository remains private. Higher evidence levels will only be added after the corresponding run or external validation exists.