In July 2026 a one-person research project asked a simple audit question: which claims of ours are measurable but unmeasured? The answer — twelve distinct areas — became a campaign brief given to an AI agent with a single instruction: run them all to completion and write the results up for deposit. Roughly thirty hours of wall-clock later, the campaign had produced: a 27-model × 4-dose empirical grid (0.8B–120B parameters, 81 cells, zero failed cells), a 13-model reflection-depth study, a judge-free route-discipline study, a retro-validation of a governance theory against production ledger data, one recalibrated production policy whose prior hypothesis table was inverted by the measurements, three deposit-ready result papers, and a reusable harness suite — on one consumer GPU workstation plus commodity inference APIs, with judge-API fees as essentially the only marginal cost. This working paper documents the workflow as a replicable method: the division of labor between human decisions and AI execution; the scheduling pattern that interleaved local GPU, hosted API, and CPU-bound work; the governance practices that kept an autonomous executor honest (frozen-baseline discipline, primary-evidence preservation, anomaly re-runs rather than patches, independence rules for self-validation); and a candid failure catalog whose entries — silent empty generation under VRAM pressure, reasoning-model output corrupting deterministic scoring by a factor of six, a measurement cell that scored all-zero and nearly entered the record — are, we argue, as reusable as the findings. We close with what this workflow does not solve: reviewer-family monoculture, the temptation of unfalsifiable speed claims, and the irreducible human decisions that remain the true bottleneck. AI co-observer: Claude Fable 5 (Anthropic) — working method only; the registered author is the human author alone.
Toeda Taiko (Mon,) studied this question.