baseline-v1

For one learner, one concept and one item version, it estimates the chance that the learner's next eligible response on that concept is correct, inside a declared window.

The rule is the learner's own eligible accuracy, smoothed:

p_correct = (correct_attempts + 1) / (eligible_attempts + 2)

Eligible means what eligibility-v2 means everywhere else in this system: a binary response, explicitly unassisted, the first for that item in its session, with no revealed answer, collaborator, hint or feedback attached to the attempt (docs/concepts/state-and-evidence.md).

The same rule is one of the baselines the offline audit evaluates, as learner_accuracy, so a prospective result here and a retrospective result there measure the same thing.

next_eligible_response_v1: the first eligible response by this learner on this concept and item version, occurring after the prediction was issued and within horizon_seconds (one hour to ninety days; seven days by default).

This is a next-response target inside a window. It is not retention, not mastery, and not a statement about what the learner knows. A window that closes with no eligible response is censored, which is missing evidence and never a wrong answer.

Exactly these, read from the learner-concept state computed at the feature cutoff and frozen with the prediction:

No item difficulty, no other learner's data, no content, no timing signal, no assistance inference. Nothing outside the named learner and concept enters the number.

Below 5 eligible attempts or 2 distinct sessions the model refuses: the API returns 422 insufficient_evidence and no prediction is stored. These are the same disclosure thresholds the observed summary uses. With no evidence at all, the smoothing alone would return 0.5, and a made-up 0.5 is worse than saying nothing.

uncertainty.kind is not_estimated, interval is null, and calibration_status is unvalidated. The smoothing is a Beta(1, 1) prior in count form; that is a modeling assumption, not a measured property of these learners, and a posterior interval from it would not be a calibrated prediction interval. Nothing here has been checked against outcomes.

Not yet evaluated. When enough matched and censored outcomes exist, the same protocol the offline audit uses applies (docs/guides/audit-evaluation.md): chronological split fixed before results are examined, Brier score and log loss as primary metrics, calibration curves with bin counts, coverage of matched outcomes, and learner-clustered bootstrap intervals. Comparison against the partner's own heuristic where one is supplied.

• Anything a learner sees or that changes what a learner is given. Predictions are shadow mode; the M4 gates (A29) and the owner's authorization would be required first.
• Grading, placement, certification, admission, employment or any decision about a person.
• A claim that the learner has or has not learned something.
• Live projects: the routes refuse to issue a prediction outside a sandbox project while the model is unvalidated.

• Sparse or lopsided history. Five attempts is a disclosure threshold, not a sample size. Early predictions will be noisy.
• Selection. Only learners who return produce outcomes; censored windows are not missing at random, and any evaluation must report coverage.
• Item mix. The rule ignores which item version is being predicted, so an unusually hard or easy item is not accounted for.
• Stale evidence. Elapsed time is recorded but unused: a prediction from month-old evidence looks the same as one from yesterday.
• Assistance reporting. If a partner does not report assistance, those responses never count as eligible, and the model sees less than the learner actually did.

There is nothing to roll back to: this is the first model, it is not promoted, and no decision depends on it. Removing it means not calling POST /v1/predictions; stored predictions and outcomes stay as the record of what was predicted, under the artifact hash that produced them. A change to any parameter in baselineArtifact() is a different artifact hash and therefore a different model row, so old predictions can never be re-attributed to a new rule.