Reasoning Without Consequence: Irreversible Selection, Curator Limits, and the Hard Ceiling of In-Context Learning

We formalize a structural asymmetry between training and inference in transformer architectures: training instantiates a commitment operator B = K_ret ∘ ι composing irreversible selection with causal persistence; inference does not. In regimes of diagnostic degeneracy — where competing hypotheses produce indistinguishable predictions — this asymmetry becomes operationally decisive. We prove a No-Commitment Theorem for single-layer linear attention: no classifier maintains stable hypothesis assignment under adversarial context extension when hypotheses are diagnostically equivalent. We conjecture the result extends to deep transformers under approximate invertibility conditions, with a pre-registered experimental protocol to adjudicate. A Curator Degeneracy analysis identifies three structural leakage channels (positional asymmetry, compositional depth, subleading statistics) below which models cannot prefer one equivalent hypothesis over another. A three-operator functional decomposition — State Discrimination, Temporal Synchronization, Irreversible Accumulation — identifies the latter as absent at inference time. We state six falsifiable predictions with explicit effect sizes and falsification criteria, anchored by a flagship experiment isolating irreversibility as the operative variable through matched reversible, standard, and semi-reversible architectures. The central claim is scoped and specific: the hard ceiling of in-context learning is not representation capacity but commitment stability — and stable commitment requires irreversible loss that persists beyond the inference episode.