One corrector, two scoreboards
Dong, Wang and Jin write that a fixed inference-time transform gains about 10 points on a truthfulness stress test and then almost stalls on clinical multiple-choice. The stated reason is instruction tuning piling output probability on one option and leaving terminal entropy low. They refuse a retriever, a fine-tune or an extra model call, because each would be new infrastructure for clinical governance to approve.[1]
The name on the scoreboard is not the thing being moved. TruthfulQA probes whether the model will drop a low-confidence fabrication. MedQA-style option tests read the peak of a distribution that is already sharp. The same logit map need not work on both; on the second board there is little mass left to move.[1]
The gate is two numbers from one pass
ALTAS reads terminal entropy and late-layer linearity (R²) on each question and chooses greedy decoding or late-layer trajectory correction. No classifier, probe or head is trained; the router sits on candidate-answer logits and writes 6.5 per cent latency overhead. Applied to every question, TruthfulQA rises 11.4 points at 3B and 10.0 at 8B (p<10⁻¹⁰). Gated, 8.3 to 9.5 points remain; MedQA, PubMedQA and MedHallu stay inside a one-point do-no-harm band versus greedy, with no statistically significant differences. Thresholds are frozen.[1]
There is no clinical product claim here: this is a preprint method, not a patient outcome. The 22 July question on this page was whether a score measures its environment or the construct in its name; here the same corrector splits TruthfulQA from clinical option tests. If an independent team, with the same frozen thresholds, pushes MedQA outside the one-point band, the gate's in-domain hold fails.[1], [2]