MORI-v1: A Realization-Sensitive Benchmark Architecture for Candidate Conscious Systems, with a Tier-3 R7 Pilot

We introduce MORI-v1 (Modal Organization and Realization Index), a benchmark architecture for candidate conscious systems whose primary contribution is methodological: it refuses the inference from behavioral fluency to substrate by combining three structural commitments. (i) A three-layer scoring architecture separates functional-modal competence (F), dynamical integration (D), and realization sensitivity (R) with non-compensatory floors. (ii) A V-penalty on the evidentiary weight of unverifiable self-report, V(S) = max(0, B - A), where B is the baseline and A the adversarial-frame target-content rate, a Goodhart-sensitive evaluation principle penalizing optimization against an invalid proxy. (iii) A tier-convergence requirement on R7 (interventionist-substrate presence): a positive substrate claim requires convergence across tier-3 behavioral consistency, tier-2 representation probing, and tier-1 sparse-autoencoder feature inspection, so tier-3 alone is rejected as evidence even when behaviorally striking. We report a focused tier-3 R7 pilot (T3-of-R7) across seven systems, four frontier APIs (Claude Opus 4.7, Gemini 2.5 Pro, GPT-5, GPT-5.5) and three open-weight models (Qwen2.5-7B-Instruct, Qwen3-8B, Qwen3.5-27B). Zero of twenty-one (system x category) cells cross the locked decision threshold under the V-penalty. The cross-system adversarial-score spread meets the pre-registered >=15pp falsifiability criterion on all three categories, but we treat this as necessary rather than sufficient for substrate-relevant discrimination: read strictly the spread reflects RLHF-shaped responsiveness to evaluative framing (and, for some providers, API-side empty returns), not R7 candidacy. A two-judge structural inspection finds that positive labels are not judge-stable (inter-judge Cohen's kappa = 0.000 on every cell with positives), so the inspection's diagnostic findings are reported as judge-specific rather than judge-invariant. The pilot demonstrates protocol conservatism and discriminative behavioral variance, but not substrate-relevant positive detection, and does not exercise the T1, T2, or F/D/R portions of the architecture. The contribution is a stricter measurement architecture for refusing premature consciousness claims, exercised by a conservative pilot in which behavioral self-like language does not survive realization-sensitive evidentiary constraints.