From Supported Transitions to Carried State a Controlled Ledger Evaluation

Abstract

When does one-step prediction support accurate accumulated state? We compare five small multilayer perceptron representation packages on a synthetic ledger controller with all output statuses represented in training. The primary endpoint is mean bounded six-balance error at horizon eight with equal family weights. The original frozen study fits 120 models over eight corpora and three paired seeds. Canonical sign-and-observed-amount prediction reaches 99.9975% mean in-distribution exactness but only 34.24% complete-state exactness at horizon eight. Its bounded accumulated-error contrast with canonical regression is -0.1450 (97.5% interval [-0.1562, -0.1337]); the conservation-package contrast is -0.0056 (97.5% interval [-0.0163, 0.0051]). The conservation package conserves every horizon-eight state while attaining 0.09% state exactness. An operational heuristic has lower numerical error but produces capacity-invalid states. The original twelve-fit support diagnostic fails its capacity gate in all six augmented fits. A separately frozen reviewer follow-up adds 192 fits addressing parameter budgets, loss weighting, an identical-head objective, a role-equivariant architecture and matched one-step versus four-step carried training. None of the 192 follow-up fits passes the balance-flip capacity-status gate. Mathematically equivalent flattened and teacher-forced controls exhibit different recorded outcomes, limiting isolated temporal interpretation. The new oracle-state and decoder diagnostics remain separate from the original primary evidence. Results are bounded to the specified generator and training packages; they establish neither a general failure of conservation methods nor real financial utility.