Frozen delivery-backtranslation-v0 source and prospective launch record. The 120-attempt construction/assessment barrier remains active; this archive contains no task deliveries, provider prompts, active operation traces, grades, or credentials. The ten development identities were heavily exposed; this is exploratory, not clean held-out validation. Reproduction requires the unchanged Harness-Bench checkout at commit b8d89634c6f445fcc68ceb6744a59bef6cdf4863, official baseline sources, Docker, and the dedicated evaluation credential. Recreate the C-Clean reference from the attached reconstruction inputs and the public agent_harness_2 v1 supplement URL in ACQUISITION.json. Use the public reconstruction script with --verify-only. The supplied execution-evidence Dockerfile requires the external ecology-execution-evidence-base:e2cb5a5f5b5c image and installs strace; neither that base image nor image layers are included here. Build it as ecology-execution-evidence:prep-v0, then build the local C-Clean Dockerfile over the reconstructed source. The fourteen manifest hashes identify the latest public source, and the PREPARATION document records its earlier preparation status; the current STUDY adopts this reference. The local image is not claimed identical to a historical paper image. Run freeze/launch only in a new study directory; preserve the original once-only barrier and accounting rules. Full commands are in the frozen INTEGRATION.md; native.py holds the shared notice, one completion continuation and Luna reader. core.py captures actual changed current artifact text, keeps native-entry source distinct, withholds it from inverse readers, and supplies it to joint reviewers. Exact evidence, caching, stale-version checks and partial failures remain visible. runtime.py exposes review/read/manifest to ordinary Python. Final construction and source adjudication remain with the actor. Helper inference is charged inside total construction cost. Preparation checks are engineering evidence, not benchmark capability evidence.
