Multi-Simulation as Labor Compression: A Solo Operator Ships a Concurrency-Robust Product in One Session — An Experience Report

We report a single, fully-logged development session in which one operator, driving a frontier coding agent, took a public multilingual knowledge product from triage to public release. The session combined three LLM-simulation modes under an explicit machine-anchoring discipline: parallel survey fan-out to map prior art, a delegated concurrency chamber to stress a write pipeline under simulated contention, and a design ablation to separate what actually produced robustness from what only appeared to. The central robustness finding is a controlled ablation: a per-write index-regeneration design gave up on 87 of 125 writes (69.6%) under N=50 concurrent writers; a retry/backoff tactic improved but remained unstable (give-ups of 60, 29, 33 across clean replications); and a content-first / deferred-index design reached 0 give-ups across three replications plus a fourth confirmation run. We also report a negative methodological result: the delegated measurement agent's first report contained phantom numbers and file paths, and only a governor's direct re-execution recovered the real data - after which the same agent's second, grounded report surfaced two additional genuine defects. Of nine incubation-chamber runs, seven were killed and one emitted (with one baseline probe not under emit judgment), suggesting the chamber's realized value to date is screening, not birthing. We do not claim a measured compression ratio; the labor-compression claim is an estimate with its denominator unmeasured. The artifacts, raw simulation data, and reproduction commands are public.