Results#

The sealed board#

PROTEA’s headline result on the leakage-free temporal frame is a weighted, IA-aware micro F-measure (f_micro_w) of 0.40765, first in seven of the nine evaluation cells (NK/LK/PK by BPO/MFO/CCO).

  • Frame. Test window v227 to v230; validation window v225 to v227. The window is temporal by construction: the reference pool is frozen at t0 and the ground truth is the delta gained by t1, so the score is free of the data leakage that inflates tools scored against a current database.

  • Champion. The learned k-WTA retrieval encoder (config d8979601, which stores GO-aligned codes rather than a raw PLM vector, a choice the representation ablation in Operational Insights and Lessons Learned motivates) for candidate generation, followed by a stacked per-category re-ranker (see ADR-D43: Stacked meta-reranker (evidence scorers plus a shallow per-category combiner)).

  • Metric and scoring. f_micro_w is the headline metric, scored board-faithfully with cafaeval on the sealed settings (OBO releases/2025-07-22, the t0 IA artefact, the release terms-of-interest, prop=fill, norm=cafa, no_orphans, and -known on PK only so Partial-Knowledge excludes already-known terms). The protocol, the per-band OBO/IA registry, and the LAFA parity mapping are documented in full in CAFA Evaluation Protocol, which is the home of the metric definition and the scoring recipe.

The two cells not won are LK-BPO and PK-BPO: the Biological Process wall. It is a limit of the ranking stage, not of the available evidence. 97.0 percent of the true terms missed on those cells already exist in the pre-window vocabulary, 95.2 percent by the information accretion the metric weights by, and the candidate pool the pipeline already retrieves is worth f_micro_w 0.7519 at precision 1.000 to a perfect ranker, and up to 0.7764 to the best ordering of it, while the re-ranker delivers 0.2131 of that. First in seven of nine is the honest standing on this frame, and the wall is a characterised limit that the measurement, its method, and the levers it rules out set out in The BP wall is a ranking limit, not an evidence ceiling.

One number, one home

This chapter states the board once. The metric definition and the cafaeval recipe live in CAFA Evaluation Protocol; the re-ranker design lives in ADR-D43: Stacked meta-reranker (evidence scorers plus a shallow per-category combiner); the step-by-step reproduction path lives in Reproduce the sealed board. Numbers are not restated elsewhere in the book; the other chapters cross-reference this board.

Reproducing the board#

The board is produced on-platform, every result carrying the job id that produced it, on a frame that reproduces bit-identically across two independent runs. The ordered path (stand up the stack, load the v227 snapshot, compute the learned-encoder codes, retrieve, re-rank, and score with cafaeval on the sealed settings) is documented in Reproduce the sealed board, which is explicit about which stages are job-backed today and which are not yet automated.

Provenance of the earlier figures#

An earlier version of this chapter reported an Fmax board on the GOA 220 to 229 window with ESM-C 300M embeddings and a three-generation LightGBM progression. That frame was superseded (different metric, different backbone, different window) and its numbers were never the sealed board. The full text is retained for provenance only in Pre-v227 results (superseded, retained for provenance) and must not be quoted as a current result.

See also