A high top-1 score is not enough. Researchers need probabilities they can budget against, and a system that refuses when the input leaves the world it learned.
Genomic machine learning often ships two numbers that look similar and mean different things: accuracy and confidence. Accuracy asks how often the top guess is right on a defined evaluation set. Confidence asks whether the probability attached to a guess matches reality often enough to be usable. Abstention asks a third question: when should the system return no answer rather than a decorative one?
For Provotics, those three questions sit at the center of Provenance-2's site-of-origin research core. This note explains the research framing. Exact served-core metrics live on Validation and the model card; prefer those pages when you need a figure for a citation. Research use only. Not a medical device.
A probability is calibrated when it matches reality over many predictions. If a model says it is 90% confident across a batch of cases, then it should turn out right on roughly 90% of them. Overconfident models sound sure when they are wrong. Underconfident models bury usable calls under timid scores. Neither helps a researcher triage work.
On Provenance-2's held-out evaluation, temperature scaling reduced expected calibration error from 0.091 to 0.011, about an eight-fold improvement. That figure is already published on the shorter note What a calibrated probability actually buys you and on the public model surfaces. The practical meaning is simple: the confidence numbers stopped lying in distribution.
A model can post a strong accuracy figure and still be wildly overconfident on the cases it gets wrong. That failure mode is especially costly in genomics, where a confident wrong call can waste follow-up assays or bias a paper's narrative. Treating accuracy and calibration as separate quality bars is therefore not pedantry. It is how you keep a research tool inspectable.
Macro-averaged metrics matter for the same reason. In a 25-site problem, overall top-1 can look healthy while rare sites remain fragile. Provotics reports macro-F1 and related cohort tables on Validation so the large classes do not hide the hard ones.
Calibration helps when the model will answer. Split conformal prediction helps decide whether to answer and how wide the answer should be. Instead of always forcing one site, a conformal procedure can return a small candidate set that contains the true site at a target coverage under in-distribution assumptions, or abstain when nothing clears the bar.
Provenance-2's public research framing targets about 90% conformal coverage in distribution, with abstention as an expected operating behavior rather than a bug. Coverage is an in-distribution property. It does not automatically transfer to sequencing platforms or pipelines the model has never seen. That limit is the whole point of pairing conformal abstention with a separate novelty check.
Conformal abstention answers a question inside the known label space. Novelty detection answers a prior question: does this sample look like something the model has learned? When an input sits far from the training distribution, trusting a calibrated softmax is the wrong move. The honest research behavior is to flag the sample and refuse a forced label.
In practice, Provotics combines distance-based novelty with confidence so unsupported or normal-tissue-like inputs are less likely to be dressed up as a confident site call. The guided demo is illustrative only and does not call the live model; it exists to teach the interface and abstention behavior without pretending to be validation evidence. See the evidence ledger for what is actually published.
In product marketing, "always answers" can look stronger than "sometimes refuses." In research tooling, the opposite is usually true. A refusal is an explicit signal that the current input is outside the claim you can defend. That signal is more useful than a quiet wrong answer with a polished UI.
Abstention also makes evaluation more honest. If you remove hard cases by silence and then quote accuracy only on answered cases without disclosing the abstain rate, you have changed the claim. Provotics tries to keep those operating points visible on validation surfaces rather than burying them in a methods appendix nobody reads.
Those habits transfer beyond Provotics. Any genomic classifier that cannot explain its confidence or its refusal rule is harder to use than its marketing suggests.