Scoring Validator

A probability, not a promise

In the dashboard, the scoring role has the tabs Overview, Assets, Tasks, Stake and rewards, and Activity. A Scoring Validator (SV) publishes a probability of default for covered assets and keeps those predictions current. A prediction of 5% does not promise repayment. It says that default should be uncommon across comparable observations. The protocol evaluates a collection of predictions. It does not treat each default as misconduct.

The operator also declares a methodology commitment and reference with Declare your scoring method. With these, others can see when the methodology changes. The contract does not audit the underlying model.

From predictions to outcomes

1. The SV publishes and updates predictions for covered assets with Publish a score.
2. A cohort closes, fixing the predictions used for that cohort. The dashboard calls a cohort a Scoring window (Close scoring window).
3. The configured forward horizon and payment allowance elapse.
4. The contract evaluates eligible scheduled-payment outcomes (Evaluate scoring window).

The implemented outcome test reads whether relevant payments were resolved on time. It does not rely only on when a keeper (a service that submits routine transactions) happened to mark a payment missed. Assets with no scheduled payment in the horizon are not in the eligible sample.

Real Testnet uses a 365-day forward horizon. Governance sets the horizon, publication cadence, and evaluation deadlines. A cohort is not ready just because its publication window has closed.

What calibration measures

The test uses a Brier score: the average squared difference between a predicted probability and the observed outcome (default or no default). Confident predictions that turn out wrong add more error.

The protocol compares the observed error with the error that the predictions themselves imply, and it allows for sample size. Lower error is better, but perfect calibration does not need zero error. The sample must meet a minimum size. If the protocol skips a small sample, that is not a successful calibration result.

Evaluation can proceed in bounded chunks of on-chain computation, so a large cohort does not have to fit into a single transaction.

Operational responsibilities

Keep your predictions current. Declare methodology changes through the supported process. Track when each cohort becomes ready for evaluation. An overdue evaluation can restrict your rewards and new volume. It is different from a completed evaluation that had too few eligible observations.

A calibration breach starts the penalty process. A missed publication cadence and an unannounced methodology change each have their own enforcement path.

Follow the Brier example to calculate an observed score, expected score, standard error, and breach threshold. The parameter reference records the testnet horizon and minimum sample.

Limits of the signal

Calibration depends on the protocol's settlement records and defined horizon. It does not assess every legal, liquidity, valuation, or operational risk of an instrument. Also, the current exit model does not hold an SV's collateral until every future cohort is evaluated. So if an evaluation comes late, the collateral may no longer be available to debit.

Prepare scoring operations

Follow Become a Business Validator for admission, role collateral, and monitoring preparation. A scoring role does not give token-management authority. It also does not commit an insurer to cover the asset.