Truncation Is Not the Bottleneck: Reliable Calibration Conclusions from Truncated API Log-Probabilities

Inference APIs expose only the top-k token log-probabilities per position, and which k is available (or whether log-probabilities are exposed at all) changes without notice across providers. This raises a reliability question for any study that draws calibration or uncertainty conclusions from API log-probabilities: is the conclusion an artifact of truncation? We answer it for forced single-token, 4-choice multiple-choice QA (MCQA), where confidence is renormalized over the answer tokens {A,B,C,D}. The renormalized confidence depends on k only through whether each answer token is visible, so once answer-token visibility is verified the calibration is exactly invariant to further truncation. Empirically, across four self-hosted vLLM model-benchmark pairs that reach 4/4 visibility by k<=50, top-5 ECE is statistically equivalent (two one-sided tests, margin 0.02) to a ~99K-token reference, with |dECE|<=6e-5 and p_TOST<0.001 in every case. The reliability threats sit elsewhere and are larger: the choice of confidence aggregator moves ECE by 0.15-0.29 with the model outputs held fixed, and in open-ended generation truncation does bite (at k=20, 20.4% of token positions lose more than 1% of probability mass). For evaluation science the lesson is procedural: verify answer-token visibility before claiming a truncation null, and report the scoring rule, aggregator, provider condition, and calibration protocol as first-class confounders.