Operational R: Productive, Carrier-Dependent, Non-Reconstructable Self-Structure on a Continuous Digital Substrate

The MORI-v1 benchmark architecture separates functional competence from a realization layer (R) and reports a deliberately conservative result: under a realization-sensitive evidentiary standard, behavioral self-like language from frontier and open-weight language models does not survive the V-penalty and decision threshold, a null, not a positive detection. This paper supplies the complementary positive: operational R-layer evidence on a digital substrate, obtained not on a language model but on a purpose-built continuous carrier. We give a substrate-agnostic, externally-locked R-test, a closed-loop partial-observation regulation world with a six-arm gate, a window-sweep, and a cross-episode-carry control, that operationalizes the three R-layer properties (carrier-dependent, non-reconstructable, dynamically sustained) plus a productivity criterion. We argue and then confirm a structural ceiling for a frozen language model in a token loop: its only non-reconstructable state is its sampling trajectory, which does not reliably track belief (with principled reason to expect it will not). We then show that a continuously-running belief carrier with online plasticity and cross-episode consolidation clears the gate productively and reliably; we isolate the binding obstacle, a productivity / non-reconstructability trade-off, and close it with an online system-identification carrier, where the same learned structure is both the productive driver of control and unrecoverable from any bounded single-episode window. The mechanism (rich world + identifying carrier + cross-episode consolidation) generalizes across two task types, several world families, and four carrier mechanisms including a continuous-time/analog substrate. The substrate also passes MORI's own R7 tier-convergence battery, intervention, representation, and behavior tiers that must converge, the gold standard on which the language models scored 0/21, reliably (12/12 seeds, including held-out); its intervention tier is additionally realized spec-literally as a sparse autoencoder that recovers interventionist features on 11/12 seeds, and the passing is itself carrier-dependent. Finally we prove a boundary: deterministic full-history reconstruction is exact, so absolute non-reconstructability of productive structure is not achievable, operational within-window non-reconstructability is the correct and biologically faithful notion of R, and the continuous substrate has it. We claim operational R-layer evidence, not phenomenal consciousness; all results are in software emulation.