Architectural Signatures in Multi-Agent Reasoning: Prompt-Independent Generation and Instrument-Dependent Evaluation

Three control architectures (single-shot, parallel ensemble + synthesizer, sequential four-role chain) processing one complex institutional case are shown to carry stable closure-operation signatures. Two confounds are defeated: prompts are regenerated by an independent adversarial generator and re-run in triplicate (27 outputs, 3/3 per signature); a pre-registered blind external judge classifies the nine anonymized outputs 9/9 by operation (chance p ≈ 5.1×10⁻⁵), including 6/6 on the critical pair. The key finding is the instrument inversion: the same external judge that barely discriminates the conditions under a quality rubric (Δ = 0.262) succeeds perfectly under an operation rubric — content converges across architectures, operations do not. Third deposit of the CODHZ research program.