Coverage and inference refinement round (2026-07-21)

Evaluation policy and recent PacBio suite

The guarded PacBio physical auto model at commit 30a4056 is the frozen accuracy baseline. Candidates must report transcript accuracy, wall time, and peak RSS and are retained only when their benefit generalizes. The primary PacBio panel is restricted to current high-throughput data: Revio Kinnex LongBench at 50k and 250k depths and the Revio Kinnex WTC11/SIRV time course. The independent HiFi simulation is a mechanism control. The existing public FLNC SIRV sample is retained only as a secondary protocol control; no further legacy Sequel datasets will be downloaded. The registry is benchmarks/pacbio_recent_suite.tsv.

PacBio's current public catalog confirms that Kinnex is the MAS-Seq-derived high-throughput protocol and that Revio is a current supported platform. A 2024 Revio SPRQ UHRR release is available, but each FLNC BAM is 14--16 GB and UHRR lacks direct transcript-level ground truth. Downloading it would add replicate consistency rather than accuracy evidence, so it is not presently worth the space or processing time.

Continuous nested-isoform guard: rejected

The first candidate replaced the hard extreme-creation clamp with a continuous abundance gate only on transcripts that co-occur with a candidate at least 5 kb longer. For physical increases, the gate was b / (b + m), where b is the coverage-free anchor count and m is the existing 37-counts-per-million midpoint. It blended 883--1,156 transcripts per 250k sample instead of clamping zero or one.

The candidate had effectively no runtime or memory cost, but worsened Pearson, Spearman, CCC, RMSE, and MARD in all three 250k samples relative to the guarded default. For example, H69 CCC fell from 0.590006 to 0.589540 and RMSE rose from 2841.29 to 2846.47. The implementation was removed. This confirms that continuous shrinkage should not be applied to every geometrically nested candidate; the narrow extreme guard remains better calibrated.

The earlier generic candidate-relative contrast gate is also not being reintroduced. Its completed outer ablation merely interpolated toward no coverage and regressed all three 250k confirmations.

EM active-set stopping: rejected

The convergence calculation was swept over active-count thresholds of 0.01, 0.1, 1, 2, 5, and 10 reads. Thresholds through one read did not reduce the 1,000 evaluations. A two-read threshold saved only 109 final H69 evaluations without an end-to-end gain. Five- and ten-read thresholds reduced 250k wall time to roughly 3.4--4.1 and 3.0--3.6 seconds, respectively, but consistently worsened MARD and perturbed low-abundance ranks. The candidate was removed.

The slow mode therefore involves supported competing isoforms, not merely numerical dust. Future convergence work should use a likelihood or assignment mass bound rather than excluding abundance strata.

Packed bootstrap inference: retained

Bootstrap inference now packs invariant alignment/coverage/KDE weights once, resamples read indices into equivalence-class multiplicities, and reuses the packed representation across replicates. Replicates remain parallelized at the outer level, avoiding nested parallelism.

On H69 PacBio 50k with ten bootstraps and four threads, wall time fell from 4.37 to 3.24 seconds (26%). Peak RSS changed from 494,872 to 495,092 KB (0.04%), and the primary quantification file is byte-identical. Bootstrap draws are intentionally random, so replicate files are compared statistically rather than byte-for-byte. Raw results are oarfish-evaluation-data/bootstrap-packed-*-h69-50k-20260721*.

Exact weighted equivalence-class collapsing: retained for serial EM

On H69 Kinnex/Revio 50k, grouping only by transcript identifiers could remove 47.9% of read records but would be invalid because coverage evidence is read-specific. Grouping by the exact ordered transcript IDs plus the exact floating-point score/coverage weights can safely remove 30.7% overall and 14.2% among ambiguous reads.

Adding exact multiplicities to serial packed EM reduces the guarded physical 50k run from 2.26 to 2.09 seconds (about 8%) with unchanged Pearson, CCC, RMSE, and MARD. Counts differ by at most 2e-13; the displayed Spearman change comes only from rank tie ordering. Peak RSS rises by about 1 MB. Applying the same hashing to parallel 250k inference made wall time slightly worse (4.98 to 5.27 seconds), so parallel EM deliberately keeps its uncollapsed packed shards. Raw results are under oarfish-evaluation-data/exact-weighted-collapse-* and weighted-eq-collapse-h69-50k-20260721*.

Joint abundance/coverage feedback: not retained

The existing responsibility-profile ablation already tests the critical single-sample feedback mechanism: a coverage-free abundance fit assigns ambiguous-read responsibilities, which then retrain coverage profiles before coverage-aware inference. It improved one Kinnex fold-change objective but regressed mean CCC, Spearman, and RMSE for ONT cDNA, ONT dRNA, and PacBio and regressed all three PacBio 250k confirmations. It also added about 3.15 seconds per million-read Kinnex sample. Repeating this circular update for more outer iterations has no empirical justification, so it is not promoted.

Any future joint model must use held-out reads or a hierarchical external technology prior; an in-sample abundance-to-coverage feedback loop is excluded.