Save a standalone Bayesian ranking as a reusable rubric reference

Prepare a completed reference ranking before scoring any new samples. This supports reference rankings fitted from pooled historical comparisons as well as a single collection of comparisons. The original scores and comparisons are saved together so later linking can check that it uses the same reference.

Usage

prepare_linked_rubric_reference(
  cj,
  evidence,
  set_id,
  items = NULL,
  trait = NULL,
  provenance = list()
)

Arguments

cj

A completed fit_bayes_btl_mcmc() result. With several refits, the last refit is used. It must record evidence identity at fit time. Older results must be refitted from their original comparisons first; they can still be used for ordinary same-set rubric calibration.

evidence

The exact comparison rows used for that refit, in the same results-table format supplied to fit_bayes_btl_mcmc(). For a subset fit, supply only that subset. See build_btl_results_data().

set_id

A single reference-set identifier, as in prepare_link_input().

items

Optional data frame with item_id and global_item_id, mapping every fitted item to a unique ID across sets. When omitted, fitted item IDs are also used as global IDs.

trait

The trait being ranked, such as "organization". Required when the completed result does not already record it.

provenance

Optional named list of serializable study/source metadata, such as source_commit. Stored separately from computed identity fields.

Value

A serializable pairwiseLLM_linked_rubric_reference. Supply it to fit_rubric_calibration() with calibration_design = "linked_anchors". items holds global IDs, source IDs, frozen scores and SDs. For linking, hub supplies the set/item mapping; points, posterior_draws, and evidence supply the E1, E2, and E3 hub inputs, respectively. Attach source to each input to carry the same computed reference_hash. judge contains the shared bias/lapse settings from the accepted fit. Diagnostics and reliability are retained; failed diagnostics warn as in ordinary calibration.

Details

The supplied comparisons must match the evidence recorded when the fit was made. Compatible item names alone are insufficient. Changed comparisons require a new fit and a new reference. Save the result with saveRDS() and restore it with readRDS(); no live sampler is needed for later prediction. Hash-engine provenance is recorded for new references. Older references without it still require every identity hash to verify. Hashes from before rlang 1.3.0 use a different algorithm; see save_link_session() for error and recovery behavior. A completed fit with unverifiable fit-time evidence identity cannot be reused automatically to rebuild a reference.

This constructor supports standalone fits with shared judge parameters. Import-ready adaptive Phase A artifacts continue to work directly with fit_rubric_calibration(). Neither route changes the original reference scores to match the distribution of a new set.

See also

prepare_link_input(), fit_rubric_calibration()

Other rubric calibration: evaluate_rubric_predictions(), fit_rubric_calibration(), predict.pairwiseLLM_rubric_calibration()