Calibrated Relevance Probability

Definition

A calibrated relevance probability is a reranker output that is a genuine probability — a number in [0, 1] whose value means “this fraction of documents scoring here are actually relevant” — rather than an uncalibrated score whose only meaningful property is its order within one result list.

The distinction matters because ranking and thresholding are different jobs. An ordering-only score answers “which of these is best?” A calibrated probability also answers “is any of these good enough?” — a question a raw score cannot answer without per-corpus tuning.

The Contrast with Cross-Encoder Logits

A Cross-Encoder emits a logit. Its magnitude is arbitrary: it is trained to separate relevant from irrelevant within a candidate list, not to mean anything in absolute terms. To use it as a cutoff you must pick a threshold empirically, per corpus, and re-pick it whenever the corpus, the retriever, or the model version changes.

Uncalibrated score (logit)Calibrated probability
Valid comparisonWithin one result listAcross lists, shards, retrieval legs
ThresholdingRequires per-corpus tuningThreshold is portable
InterpretationOrdinal only”0.9 means ~90% likely relevant”
Cross-model comparisonMeaninglessMeaningful if both are calibrated

What Calibration Buys You

Once the score is a probability, several operations become available that ordering alone does not support:

  • Pruning an overfetched pool. Retrieve deep, then drop everything below a fixed probability — the shortlist size becomes adaptive to how good the matches actually are instead of a fixed top-K. A single rerank call becomes a rerank and a prune.
  • Fusing across retrieval legs. Scores from different shards or from the lexical and vector legs of a Hybrid Search pipeline are on one comparable axis, which is otherwise the whole reason Score Normalization and Reciprocal Rank Fusion exist.
  • Gating a downstream step. A RAG pipeline can decline to answer, or escalate to a wider retrieval, when nothing clears a confidence bar — rather than always feeding the top-K whatever its quality.
  • Honest empty results. “Nothing here is relevant” is expressible.

Measuring Calibration

Calibration is checked by bucketing scores and comparing the predicted rate against the judged rate in each bucket: of the documents scored above 0.9, what fraction were actually labeled relevant? A well-calibrated model tracks the diagonal; an overconfident one sits below it.

This is not what NDCG or MRR measure. Those are pure ranking metrics — they are invariant to any monotonic transform of the scores, so a model can be perfectly ordered and wildly miscalibrated at the same time. Calibration is a separate axis of quality that standard reranker leaderboards do not report.

Where Calibration Comes From

Calibration is a training objective, not a side effect. A Cross-Encoder is trained to separate relevant from irrelevant within a candidate list, which fixes the ordering and leaves the scale arbitrary. Getting a probability instead means optimizing for it directly — Reinforcement Learning for Calibrated Decisions is one stated approach, used for the System One Models that produced the measurement below. Post-hoc calibration of an existing scorer (Platt scaling, isotonic regression) is the classical alternative, at the cost of a held-out judged set per corpus — which is the per-corpus tuning the property was supposed to eliminate.

Two metrics summarise that check. Expected Calibration Error averages the per-bucket gap between predicted and observed rates; Brier score takes the mean squared error of the probabilities directly. They answer different questions and are worth reporting together with log loss, which is the one that punishes confident mistakes hard enough to separate a fuzzy model from a dangerous one.

A worked instance of exactly that separation: on a sentence-classification task, Jev scored Brier 0.156 against a TF-IDF baseline’s 0.102 — bad but not alarming — while log loss came in at 1.849 against 0.335, five and a half times worse (Adapting Jev to Your Domain with GEPA). The second number is the true description of the failure.

In Practice

A worked measurement appears in Hev meets Jev, which uses Jev — a structured-output model that answers a true-or-false question as a probability — as a reranker over BEIR shortlists. On the SciFact subset, documents the model scored above 0.9 were judged relevant 76% of the time, and documents scored below 0.1, half a percent of the time. The author’s framing is that this — not the nDCG@10 figure, which sits level with the purpose-built rerankers — is what the approach offers that a cross-encoder does not.

The claim is narrow and worth keeping narrow: a 0.9 bucket at 76% precision is usable, not perfect calibration, and it was measured on one corpus.

Calibration Is Not Grounding

The limit worth stating alongside the property, put sharply in How to Use Jev - A Practical Guide: a scoring model knows only the state it was handed. It cannot look anything up. So if the candidate set is assembled from a weak source, the model returns a well-calibrated judgment about bad material — the number is honest about the verdict, and the verdict was reached on evidence that should not have been there.

Two distinct failures hide behind one confident-looking score:

What a calibrated probability tells youWhat it does not
JudgmentHow likely this verdict is correctWhether the candidate deserved to be judged
EvidenceWhether retrieval surfaced the right pool at all

This is the recall problem wearing a different hat: a reranker scoring 0.95 on the best of thirty wrong documents is well-calibrated and useless. Calibration makes thresholds portable; it does not make the candidate set correct, and it cannot signal that the right answer was never retrieved. The same guide’s practical rule — retrieve precisely, then judge cheaply — is the mitigation, along with padding the state as little as possible, since accuracy degrades as irrelevant material accumulates (Clean Context).

Calibration Does Not Travel

A third limit, distinct from the other two: calibration is measured on a distribution, not guaranteed as a property of the model. On data unlike what it was calibrated against — a new corpus, a new domain, a shifted query mix — it will still return confident-looking probabilities, and those will be miscalibrated. Nothing in the output signals that it has left familiar ground.

This bites in exactly the case the property is most attractive for. The selling point is a threshold you need not re-tune per corpus; the catch is that whether the threshold still holds is itself a per-corpus question. The one published search measurement covers a single BEIR subset — a long way from evidence that a 0.9 cutoff means the same thing on a product catalogue or a legal corpus.

The practical answer is the cheap one: spot-check calibration against a judged sample whenever the corpus changes, in the same spirit as Out-of-Time Validation. A calibrated score gives you a defensible starting threshold, not a permanent one.

Calibration Is Partly a Property of the Prompt

A fourth limit, and the most actionable one. For a model that takes its question in natural language, how well-calibrated it is depends on how you asked — not only on how it was trained.

Adapting Jev to Your Domain with GEPA separates the two. The same Jev version on the same 300 sentences scored a 10-bin ECE of 0.142 under a hand-written instruction and 0.069 under one optimized against Brier scoreBrier itself falling 44.9%. Nothing about the model changed. Roughly half the miscalibration was attributable to the wording.

Two consequences. First, a vendor’s calibration claim and a third party’s calibration measurement can both be honest and disagree, because they are measuring different instructions (Prompt Sensitivity). Second, calibration is a tunable objective rather than an inherited property: point a prompt optimizer at a probability metric instead of an accuracy metric and you can move it deliberately. The catch is that this makes the prompt a fitted parameter — one that needs a held-out set, and that can learn a corpus’s annotation conventions rather than the truth.