Predict rubric scores from a calibration

Predict rubric scores from a calibration

Usage

# S3 method for class 'pairwiseLLM_rubric_calibration'
predict(object, newdata = NULL, hard_score = c("median", "mode"), ...)

Arguments

object

A pairwiseLLM_rubric_calibration object.

newdata

Completed CJ result to predict, or NULL for the original items. Same-set scoring accepts the original completed result with unchanged item IDs, exact accepted scores and uncertainty, and matching fit/reference evidence. Item reordering is allowed; collection provenance does not affect scoring. Raw tables, independent cohorts, refits, and Phase B targets are not supported for same-set prediction. For linked_anchors, NULL scores the original reference items; explicit newdata must be a completed Phase B result (or its state) on the stored reference scale. It returns only target items across all spokes, in their input order. Every spoke must have an accepted Phase B refit and committed active hub-spoke evidence.

hard_score

Ordinal hard-score rule: "median" (default) or "mode". The median is the lowest category whose cumulative probability is at least 0.5; equality at a median cutpoint therefore selects the lower category. Modal ties select the lowest category. Ordinal output always retains all category probabilities, both decision rules, and expected level. Both choices give the same deterministic category for percentile scoring.

...

Reserved arguments; currently must be empty.

Value

For percentile scoring, a tibble with item_id, accepted source theta, integer category in 1:K, original-label rubric_score, and extrapolated (outside the fitted CJ range). Under the original-result restriction, extrapolation flags are always false. Stored cutpoints are reused unchanged. No category probabilities or probabilistic summaries are returned. Both ordinal methods include the same five columns, plus probabilities, a list-column of numeric K-vectors named by the original ordered labels; integer median_category and modal_category; and expected_level = sum(k * P(Y = k)) for internal indices k in 1:K, regardless of the original labels' numerical spacing. This expected rubric level is a continuous summary, not a replacement for the ordinal result. category and rubric_score use the requested hard_score rule. Ordinal extrapolated flags use the labeled calibration range, so unlabeled source items can be extrapolated; endpoints are included in the range. Stored standardization is reused. Unfitted objects fail clearly. Linked target output adds set_id (target spoke), source_item_id, global_item_id, and theta_sd (accepted Phase B uncertainty, possibly NA). Its linking attribute contains normalized reference, fit_contract, provenance (including hub/spoke IDs and linking stage logs), diagnostics, and reliability. Provenance estimation/uncertainty methods reflect validated spoke contracts, including accepted-state reuse. Each is a single string when all spokes agree, otherwise the distinct strings in spoke order; per-spoke contracts retain the complete attribution. Failed upstream diagnostics warn and remain available; accepted scores do not assert adequate precision or successful stopping. CJ uncertainty is retained as metadata and is not propagated into category probabilities. Reference predictions and same-set output keep their existing columns and have no linking attribute.

See also

fit_rubric_calibration(), evaluate_rubric_predictions()

Other rubric calibration: evaluate_rubric_predictions(), fit_rubric_calibration()