Prospective construction-portfolio-v0 source, freeze and neutral engineering checks. No evaluated attempt has started when this source archive is prepared. The six complete policies compare stock native, equipped ordinary, two/three-stage sustained direct construction, exact whole-candidate selection and integration. Candidate pools are acquired independently for each policy, using the same generator and budgets. There is no shared-pool accounting or paired fixed-pool causal estimate. All native stages use the original task-owned Luna proxy and common150-request ceiling; candidate acquisition, failures, retries and finalization are included once. No grading information is made available to an actor. The study has180 original attempts across ten heavily exposed development identities and three repetitions. It is not clean validation. Reproduction requires the unchanged Harness-Bench checkout at commit b8d89634c6f445fcc68ceb6744a59bef6cdf4863, official baseline, Docker and the dedicated Luna credential. Image layers are external. C-Clean reconstruction inputs and the public supplement URL are included; current public source matching does not establish historical image identity. All five equipped policies additionally refresh the reference cleanup snapshot at task-round entry and bind cleanup to the actual current task prompt. These common integration repairs are explicit departures from an unmodified reference. No closed study was changed or rerun. A fork includes regular workspace/runtime files and native conversation, not service or process state, ACLs, xattrs or ownership. Dedicated service networks and unsupported initial trees fall back to ordinary native construction with remaining time; candidate failures stay visible. Selection installs sealed candidate bytes/state/response; integration can edit current files. Trusted stopped-tree transfers have a fixed window ending60seconds after the original deadline; no inference or checks use this window. Expired decision time retains the first available candidate. Disposable checks are network-disabled and cannot change sealed candidates. See frozen CHOICE, STUDY and INTEGRATION for exact commands, limits, failure rules, once-only full-barrier analysis and the fixed24-delivery review. All preparation uses authored neutral fixtures and scripted inference, and supplies no empirical capability result. No task artifacts, provider prompts, assessments, model credentials or active outcomes are included.
