TCSG Pilot Data Pack v1.0: records of a local feasibility pilot (attempts, GPU telemetry, error register, harness code)

Records of my local feasibility pilot of the TCSG measurement chain, 31 August to 4 September 2026: one workstation with an NVIDIA RTX 5070, GPU energy only, one SWE-bench Verified task (astropy__astropy-13033). The pilot is a harness test. It produced no verified successful goal and no ECSG value, and it is not an AI performance result. Contents: attempt and control tables (24 rows), episode summaries, a synthetic calculator fixture, 125 GPU power samples, the error register (26 harness and process errors, 3 model failures, attributed as recorded), a manifest of the public datasets in my local archive (the datasets themselves are not included), path-scrubbed evidence files, and the pilot harness code and configuration. CHECKS.md reconciles each pilot figure in the paper against the files; CHECKSUMS.sha256 lists a checksum for every file. Personal paths were replaced with placeholders and no values were changed. Limits: GPU energy only; one task, one model family, one machine, one operator; no statistics. Third-party material keeps its own licence. AI use: the TCSG framework, definitions, study design and conclusions are my own work. I used Claude (Anthropic) for editing, formatting and rechecking sources. I built and operated the feasibility pilot with Codex (OpenAI) acting as an execution assistant under my instruction and approval. No AI tool generated data, sources or citations.