The Fixable Failure Sits at a Decision Interface, Not in the Compute Behind It
Four unrelated results in this window locate the fixable failure at a boundary rather than in the computation behind it: the first entry choice under RLVR, the rollout/training turn boundary, the abstention readout, and the calibrated-probability output. Each also reports a repair whose downstream cost is bounded — coverage restored by late-layer interpolation with no pass@1 penalty, log-probability drift halved at turn boundaries, dose-responsive steering on a decodable direction, and exact decision preservation by construction. Boundary-level diagnosis and repair (entry entropy, drift repair, readout steering, decision-preserving repair) is therefore a distinct capability that post-training and serving stacks should ship as first-class features instead of treating as research tricks.
OpenConfidence 40%
Who it is for
Post-training platform teams, model-reliability and eval tooling vendors, and buyers who currently respond to agent regressions by retraining or scaling.
Why now
The four results come from different groups, methods and even modalities between 2026-09-05 and 09-11, and each supplies both the localization and the bounded-cost repair — which is what makes a shared tooling layer credible now rather than a single paper's technique.
What would kill it
A controlled study showing interface-level interventions do not compose with normal training — for example first-step entropy injection or prefix reintroduction costing pass@1 in other domains — or production trace analysis showing failures are dominated by deep multi-step computation rather than entry, boundary and readout decisions.
Localizes RLVR coverage loss to the first decision rather than downstream reasoning, and shows the repair is cheap: coverage falls up to 67%, entrance-prefix injection lifts completion from 0.018 to 0.212, and late-layer interpolation recovers 37% of coverage with no pass@1 loss because unselected families remain executable.
Localizes a second, different interface — the turn boundary between rollout and training — and shows that treating it explicitly (TITO with drift repair, rollout routing replay for sampled token identifiers and MoE expert choices) halves the log-probability difference from 0.021 to 0.013, which is what makes 300+ turn training stable.
Localizes the abstention failure at the readout: structural invalidity is linearly encoded and nearly orthogonal to the safety-refusal direction, so the defect is routing from an existing signal to the decision, and generation-time steering moves behavior bidirectionally and dose-responsively while random directions do not.
Shows the same structural move outside LLMs: top-1 decision preservation is moved out of calibrator fitting into post-hoc output repair, achieving zero Top-1 Prediction Change Rate by construction while lowering ECE, NLL and Brier — evidence that separating the decision interface from the fitted model generalizes.