One number, three claims

Probability theory supplies rules for combining uncertain propositions. If two mutually exclusive events have probabilities 0.3 and 0.2, their union has probability 0.5. Those rules do not, by themselves, say what a probability is. The long debate over interpretations exists because the same formal calculus can represent physical frequencies, rational credences and decision-relevant uncertainty.[1]

The frequency reading: which collection?

On a frequency reading, a probability earns meaning through a repeatable class. “The probability is 0.70” says that, in a suitably defined sequence or population, the event occurs at about that rate. The discipline lies in the denominator. A model validated on all payments may appear calibrated while failing for a newly introduced channel, a rare country pair or a different adjudication policy. There is no frequency without a rule for deciding which cases belong together.

Frequency is attractive because outcomes can answer back. Collect many forecasts near 0.70, wait for the labels, and compare the observed rate with 0.70. Yet the test works only after the event, horizon and reference class are fixed. “Rain tomorrow” is weaker than “at least 0.01 inches at this forecast point during the next twelve hours”. “Customer will default” is weaker than “a contractual default will be recorded within twelve months, under this product definition, using information available before decision time”.

A frequency claim is therefore a statement about a measurement process, not merely about a model output. Change the grouping rule, label policy or time window and the empirical meaning changes even if the printed number does not.

The belief reading: relative to which evidence?

On a belief reading, probability measures an agent's graded commitment given an information state. Ramsey's account treated the probability laws as constraints on consistent partial belief, and later subjective traditions developed the connection between credence, preference and choice.[12] The phrase “given an information state” does most of the work. Two careful agents can assign different probabilities without either violating probability theory when they possess different evidence, use different models or place different prior weight on possibilities.

This reading handles singular events naturally. There will be only one result of tomorrow's election, one outcome of a particular safety review and one fate of a specific shipment. A credence can still be coherent because it describes what is reasonable to believe before the outcome is known. The single outcome cannot prove that a 0.70 credence was irrational. A realised failure after a 0.70 success forecast is compatible with the forecast. Accountability has to accumulate across related judgements or examine whether the evidence and update rule were defensible.

The decision reading: what would the number license?

A decision reading asks what choices are justified when uncertain outcomes carry different consequences. It is often expressed through a fair price, betting rate or expected-utility calculation. Proper scoring rules can elicit probabilities by making honest reporting optimal under stated assumptions, an idea developed in the probability-assessment literature.[11]

This is operationally powerful and easy to misuse. Observed willingness to act combines beliefs with utility, risk attitude, liquidity, institutional duty and available alternatives. A person who declines a wager at seventy per cent odds may distrust the evidence, dislike gambling, face a cash constraint or assign special value to avoiding loss. An action threshold is therefore not a clean probability unless the consequence model is separately controlled.

Probability and utility must remain distinct even when they meet in a decision. A forecast should not be silently retrained to mimic a policy threshold and then presented as if it estimated an event rate. Otherwise a budget change, review-capacity change or regulatory rule can alter the “probability” without any change in the event or evidence.

Translation between readings is conditional

A mature forecast can support all three readings, but the translations require evidence. A meteorologist may hold a 0.70 credence because a model and observations support it. Across many forecasts issued under comparable conditions, the 0.70 reports may also realise at roughly seventy per cent. A user can then combine that probability with the costs of cancelling an event or being caught in rain. The shared number is useful because each transition is visible.

Problems begin when the transitions are assumed. A model's softmax output is often called a belief although it may not be calibrated as a probability of correctness. A historical event rate is sometimes attached to an individual even when the individual's evidence differs materially from the reference class. A business threshold may be embedded during training and later described as risk. These moves take a result from one layer and silently promote it into another.

The correct question is not which interpretation wins in the abstract, but which obligations the current use requires. Scientific estimation may prioritise repeated performance and model criticism. A one-off intelligence assessment may need coherent credence and explicit evidence. An operational control needs those virtues plus a loss model, authority and recovery. The probability contract keeps the obligations additive rather than interchangeable.

flowchart TD
  A["The 70 per cent problem"] --> B["One number, three claims"]
  B --> C["How probability earns<br/>accountability"]
  C --> D["When good-looking<br/>probabilities fail"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#5B5CF6,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

How probability earns accountability

The three readings do not live in sealed philosophical rooms. Operational methods connect them. Calibration compares stated probabilities with observed rates. Proper scoring rules make truthful reporting advantageous in expectation. Decision theory combines probability with consequences. Each bridge answers a different challenge, so passing one test cannot substitute for passing the others.

Calibration: does 0.70 behave like 0.70?

Imagine sorting past cases by forecast. Among cases assigned probability 0.70, about seventy per cent should realise the event. More generally, a forecast is calibrated when the observed frequency conditional on the forecast value matches that value. Dawid gave this idea a precise sequential form for probability assessments.[4]

Calibration turns a belief report into a public empirical obligation. It does not require every 0.70 case to succeed. It asks the collection of 0.70 cases to realise the event at the stated rate. The obligation is local to the event definition, information boundary and population over which the comparison is made.

This distinction changes how an individual error should be investigated. When a 0.90 forecast fails, the immediate question is not whether the forecaster lied. The failure may be one of the ten expected exceptions in a hundred such cases. Investigation should instead ask whether the case belonged to the declared population, whether the evidence cut was respected and whether similar 0.90 forecasts fail too often. Calibration audits a forecasting practice; it does not convert probability into certainty case by case.

The converse matters as well. A successful 0.05 forecast does not prove hidden skill. Rare events occur. Treating every surprising success as validation encourages post-hoc storytelling and makes weak systems appear clairvoyant. The unit of evidence is the predeclared collection and scoring process.

A score that sees distance

A calibration table needs many cases at each probability. A scoring rule gives every forecast a loss. For a binary event, the Brier score averages squared distance between the forecast and the realised outcome. Brier introduced the probability score for forecast verification in 1950.[2]

The Brier score is strictly proper for binary probabilities: under the assumed distribution, expected loss is minimised by reporting the forecaster's actual probability. Proper scoring rules generalise this incentive to richer predictive distributions.[5] Properness disciplines reporting, although it cannot ensure that the forecaster's evidence, model or population is appropriate.

Different proper scores express different sensitivities. Squared loss gives a finite, symmetric penalty and decomposes cleanly for diagnosis. Logarithmic loss punishes assigning tiny probability to an event that occurs, and an asserted probability of zero produces unbounded loss. That severity can be useful when unsupported certainty is the central risk. It can also let a few extreme cases dominate a small evaluation. Score choice is therefore part of the measurement design, even when all candidate scores are proper.

A proper score also assumes the reward is the reason for the report. If a forecaster benefits from influencing the downstream action, reputational incentives and strategic behaviour can overwhelm the formal payment rule. In machine systems, an analogous problem appears when the training objective rewards acceptance, click-through or task completion while the output is later interpreted as an event probability. Properness belongs to a specified elicitation game, not to a number detached from the game.

Why calibration is not enough

Suppose half the cases realise the event. A system that prints 0.50 for every case is perfectly calibrated over the full sample. It is also useless for ranking or selective action. A better system may assign 0.10 to one group and 0.90 to another, with each group realising its stated rate. Both are calibrated; only the second resolves high-risk from low-risk cases.

Murphy's decomposition separates these properties.[3] For discrete forecast groups, the Brier score can be written as:

Consequences turn a probability into a threshold

Consider two actions: intervene or do nothing. A false positive intervention costs C_FP; a missed positive event costs C_FN. If the event probability is p, intervene when the expected false-positive cost is no greater than the expected false-negative cost. Under this simple loss model, the threshold is:

When good-looking probabilities fail

Probability failures often survive a dashboard because the dashboard tests the number under conditions different from those of use. The failure is then blamed on “model confidence”, although the broken object may be the event definition, label process, reference class, information cut or consequence model.

The reference class is chosen, not discovered

A new case belongs to many classes: all transactions, card-not-present transactions, first-time beneficiaries, a region, an amount band, a device pattern and combinations of these. Each class can have a different event rate. Too broad a class hides local error; too narrow a class produces unstable estimates. The correct class is the one that preserves the mechanism relevant to the claim and has enough evidence for a defensible test.

Subgroup calibration is therefore a design question. It requires deciding which distinctions could change the outcome process or decision. The answer should be declared before inspecting every possible slice, otherwise a team can search until it finds either reassurance or failure. Where many intersecting groups matter, uncertainty intervals and sample thresholds should accompany the calibration estimate.

A probability has an information timestamp

Suppose training labels include the result of a manual investigation, but the live forecast is issued before that investigation. A validation pipeline that accidentally exposes investigation-derived features can produce excellent probabilities for an impossible information state. A milder version occurs when two forecasts called 0.70 are conditioned on different evidence. Grouping them together tests neither belief cleanly.

The information cut is part of the forecast's identity. Record what was knowable, permitted and actually supplied when the probability was issued. This matters for retrospective replay, because later evidence can make an old forecast look foolish or obvious. A fair evaluation reconstructs the world and evidence available at that decision moment.

Actions can change both outcomes and labels

Forecast evaluation is simplest when prediction does not affect the event and every outcome is observed. Consequential systems often violate both conditions. A fraud intervention can prevent the loss it predicted. A medical treatment can change the patient's outcome. A review queue can produce high-quality labels for selected cases while unreviewed cases remain weakly labelled. The system then observes a mixture of natural outcomes, intervention effects and missing counterfactuals.

Suppose every high-risk payment is held and investigated. Confirmed fraud may be measured accurately in that group. Low-risk payments are released and only a fraction of later fraud is reported. A naive reliability curve can make the model appear underconfident at the top and exceptionally safe at the bottom, even when the difference comes from observation. Random audits, delayed outcome linkage, inverse-probability weighting or a causal evaluation design may be needed, depending on the workflow.

This creates a boundary between prediction and policy learning. A probability of the outcome without intervention is counterfactual once the system acts. A probability of the outcome under the current policy changes when the policy changes. The contract must say which quantity is intended. Otherwise recalibration can chase the effects of yesterday's intervention policy and destabilise tomorrow's.

Calibration error is itself a measurement

Continuous predictions rarely repeat exactly, so practical metrics place them into bins. Expected calibration error then depends on bin count, bin boundaries, class conditioning and which probabilities are included. Nixon and colleagues showed that these choices can change the ranking of recalibration methods in deep-learning experiments.[8] Kumar, Liang and Ma further showed that popular methods can understate miscalibration and developed estimators with stronger guarantees under their assumptions.[9]

A dashboard should therefore expose the reliability diagram, sample counts and estimator choices, rather than displaying one ECE value as an intrinsic property. In sparse high-consequence regions, a wide uncertainty interval is more honest than a smooth curve inferred from a handful of cases.

Calibration can be gamed without lying

A surprising theoretical result sharpens the negative control. Foster and Vohra showed that randomised forecasting procedures can achieve asymptotic calibration against very broad outcome sequences.[6] The result is mathematically valuable, yet it warns against treating calibration as proof that a system understood the generating process. A procedure can satisfy a long-run calibration criterion through its forecasting strategy without recovering the causal structure a user imagines.

Calibration is best treated as a necessary operating property for probability forecasts, joined by resolution, proper scoring, robustness under shift and decision value. Modern classifiers may also be accurate while miscalibrated. Guo and colleagues documented this pattern for several neural-network architectures and datasets, and found temperature scaling effective in many of their experiments.[7] That finding supports recalibration as a practical tool, not as a guarantee beyond the tested distributions.

Build the probability contract

A production system should reject a probabilistic value that arrives without enough structure to evaluate or use it. The contract below is deliberately stricter than a model API response. It separates the claim from the authority to act on it.

The contract can be implemented as a typed object with invariants. The event key must resolve to an approved outcome definition. The valid horizon must end after issue time and before the calibration evidence expires. Feature timestamps must not cross the information cut. The population key must select an evaluation report with a minimum sample and declared uncertainty. The interpretation must be one of a small controlled vocabulary rather than free text.

These checks do not decide whether the model is good. They prevent the system from asking an undefined question. A probability lacking an outcome key cannot be verified. A value lacking an information timestamp cannot be replayed fairly. A value lacking a population cannot inherit calibration evidence. A value lacking an expiry can continue to influence decisions after the process changes.

Contract validation should fail closed before consequence calculation begins. This is an important authority boundary. Model code may propose values and supporting metadata. A deterministic component verifies that the proposal describes an approved probabilistic claim. Only then may a separate decision service combine it with current policy and costs.

The fields are related. An outcome label created by the action itself can invalidate a simple calibration test. A reference class defined after seeing errors can turn evaluation into selection. A threshold owned by the model team can conceal business value judgements. The contract makes those dependencies reviewable.

Worked example 1: a minimal weather forecast

Consider the statement “30 per cent chance of measurable rain at point P between 06:00 and 18:00 tomorrow”. The event is binary, the point and horizon are explicit, and the evidence cut is the forecast issue time. The interpretation is a probability forecast intended to match frequencies across comparable point-period forecasts. The test groups similar forecasts and compares them with observed precipitation, while a proper score also rewards useful variation.

After one dry day, the forecast has not been refuted. After a large collection of 0.30 forecasts, an observed rate near 0.30 supports reliability for that population. A farmer deciding whether to irrigate still needs crop value, irrigation cost and tolerance for missed rain. The forecast contract ends where the farmer's decision contract begins.

Worked example 2: a production-shaped payment review

Worked scenario, using synthetic values. A monitoring model assigns probability 0.18 to the event “the payment will be confirmed unauthorised within thirty days under the adjudication policy in force at issue time”. Its information cut is the pre-payment event stream. The calibration population is card-not-present transfers in the same channel and amount band, excluding cases whose labels remain unresolved.

A review costs one unit and a missed unauthorised payment costs eight units in the simplified operating model. The numerical threshold is 1/(1+8), about 0.111, so 0.18 qualifies for review. That calculation does not authorise a hold. A deterministic policy checks customer consent, jurisdiction, review capacity, statutory timing and whether the model's calibration evidence is fresh. A human or authorised rule owns the effect.

Later, the outcome observer records whether adjudication confirmed the event. It also records selection: reviewed cases may receive better labels than released cases. The calibration analysis must correct or at least disclose that observation process. Without this step, the system can look safer simply because it examines its high-scoring cases more thoroughly.

Receipts make the contract operational

At issue time, store a decision receipt containing the probability contract, model lineage, evidence references, policy version and chosen action. After execution, store an effect receipt and an independent readback of the resulting state. When the label matures, attach outcome evidence without rewriting the earlier receipt. This sequence preserves the distinction between what was believed, what was authorised, what happened and what was learned later.

Replay then becomes a controlled experiment. A reviewer can ask whether the old probability was reasonable under the old evidence, whether today's model would differ, whether the old policy chose the expected action and whether the effect occurred exactly once. Without immutable time boundaries, retrospective evaluation quietly gives the past system knowledge it never had.

Run the calibration lab

The laboratory uses five synthetic strata with exact event rates 0.10, 0.30, 0.50, 0.70 and 0.90. It compares a sharp calibrated forecast, a constant base-rate forecast and an overconfident transformation. Change the consequence ratio and watch decision cost move independently of calibration.

Read the columns diagnostically. Reliability near zero means each quoted value matches its realised rate. Resolution above zero means the forecast partitions cases into groups with different rates. Brier score combines both with outcome uncertainty determined by the event rate. Decision cost depends on the declared action rule, so it can change while all forecast-quality columns remain fixed. That separation is the point of the experiment.

The overconfident predictor preserves the ordering of cases. A ranking metric could therefore prefer it just as strongly as the sharp calibrated predictor. Its probabilities are still wrong, and the error changes which cases cross a threshold. The base-rate predictor shows the opposite failure: its probability is globally right and its ranking is useless. A serious evaluation needs both controls.

levels = [0.10, 0.30, 0.50, 0.70, 0.90]
sharp(case) = case.true_rate
base_rate(case) = 0.50
overconfident(case) = sigmoid(1.8 * logit(case.true_rate))

brier = mean((forecast - outcome) ** 2)
reliability = sum(weight * (forecast - observed_rate) ** 2)
resolution = sum(weight * (observed_rate - prevalence) ** 2)
threshold = false_positive_cost / (false_positive_cost + false_negative_cost)
intervene = forecast >= threshold

Audit a probability statement

Use the instrument below before a probability enters a consequential workflow. A “no” in the first six rows blocks action. Later rows determine how much authority, monitoring and recovery the use case requires.

The audit maps naturally onto a controlled AI system. Intent states the decision being supported. Identity records the issuer and consumer. World state defines the event and later outcome. Context fixes the evidence cut. Reasoning identifies the estimator and calibration transform. Authority remains outside the model. Action uses a typed contract. Evidence and outcome support evaluation. Release and recovery define expiry, rollback and replay.

Run the gate at three cadences. At design time, challenge the event, reference class and loss model. At release time, freeze calibration evidence, tolerances and authority. In operation, monitor population movement, missing labels, local reliability, action rates and consequence drift. A threshold can become wrong even when the model remains calibrated because review cost or missed-event harm changed. Conversely, a stable threshold can become unsafe when calibration drifts.

Withdrawal should be an ordinary state, not an emergency improvisation. When evidence expires or a sentinel breaches tolerance, the system can abstain, revert to a validated baseline, widen human review or disable the affected subgroup. The contract supplies the exact scope of withdrawal, avoiding a choice between ignoring the problem and shutting down every use.

A probability becomes safer when it loses the power to smuggle these layers inside one floating-point value. The architecture should make the number easy to compute and hard to misuse.

stateDiagram-v2
  [*] --> Claim
  state "The frequency reading, which<br/>collection?" as Claim
  state "The belief reading: relative<br/>to which evidence?" as Test
  state "Thought experiment 1: the<br/>outcome is fixed, the evidence<br/>moves" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Compact glossary

Calibration

Agreement between stated probabilities and observed rates in a declared population.

Resolution

Ability to separate cases whose realised event rates differ from the overall base rate.

Proper scoring rule

A rule whose expected score is optimised by reporting the forecaster's actual predictive distribution.

Reference class

The set of cases treated as comparable for a frequency or calibration claim.

Information cut

The evidence knowable and permitted at the exact time a probability was issued.

Decision threshold

The probability at which one action's expected consequence becomes preferable to another's.