Retiring the non-logistic coverage models (2026-08-03)

Summary

Every coverage model oarfish offers was run head-to-head on Panel B — six NanoSim/TKSM simulations with exact read-level truth — using the current binary, serially on an idle node. --model-coverage (logistic) beats all of them.

Results

Spearman against exact truth, scored over an identical reference transcriptome:

sample	none	logistic	auto	endpoint	hybrid	degradation
nanosim-NA12878-cdna	0.840612	0.887130	0.885205	0.820957	0.877806	n/a
nanosim-NA12878-drna	0.873359	0.920082	0.917575	0.848241	0.911196	0.917575
nanosim-H9-cdna	0.876814	0.910829	0.910073	0.857114	0.905001	n/a
nanosim-H9-drna	0.911101	0.936289	0.934951	0.878742	0.931588	0.934951
tksm-RSII	0.932431	0.938823	0.940940	0.928092	0.941208	n/a
tksm-SQ2	0.950101	0.947869	0.949717	0.951190	0.952083	n/a

Paired against logistic:

arm	mean Spearman	Δ vs logistic	better on	mean wall
`none`	0.897403	-0.026101	1/6	111.3 s
`logistic`	0.923504	—	—	115.0 s
`auto`	0.923077	-0.000427	2/6	144.8 s (+25.9%)
`endpoint`	0.880723	-0.042781	1/6	116.0 s
`hybrid`	0.919814	-0.003690	2/6	117.2 s
`degradation`	—	-0.001923	0/2	109.9 s

Notes on reading this table:

endpoint is worse than no coverage model at all (0.880723 vs none's 0.897403), losing to none on 5 of 6 samples.
degradation is byte-identical to auto on its two applicable samples. It is what auto dispatches to for ONT direct-RNA. It requires --seq-tech ont-drna and errors on cDNA and PacBio, so it is only evaluable on 2 of 6 samples; its higher raw mean is an artifact of those two samples having the highest absolute Spearman in the panel.
Coverage modelling itself is clearly worthwhile: logistic - none is +0.026 on 5/6 samples. This retirement removes alternatives to logistic, not coverage modelling.

The PacBio caveat, stated explicitly

On both TKSM PacBio samples, hybrid and auto beat logistic, and on TKSM SQ2 even none beats it (0.950101 vs 0.947869). This is the same pattern seen throughout this evaluation series: logistic underperforms on PacBio.

That is not sufficient to keep the stack — n=2, no kernel wins overall, and the gains are 0.002-0.004 — but it is a real, repeatedly-observed signal and the most likely place a future coverage model would find headroom. It is recorded here rather than dropped.

What was removed and what was kept

Dependencies were verified rather than assumed. binomial_probability initially appeared orphaned to a module-path search, but is reached from single_cell.rs via a crate-root re-export; deleting it would have broken single-cell quantification. single_cell.rs itself does not touch the adaptive stack.

	lines
removed	388	`util/endpoint_probability.rs`
removed	271	`util/hybrid_probability.rs`
removed	545	`util/degradation_probability.rs`
removed	274	`util/pacbio_endpoint_probability.rs`
removed total	1,478
kept	81	`util/logistic_probability.rs` (logistic path)
kept	76	`util/normalize_probability.rs` (logistic path)
kept	224	`util/binomial_probability.rs` (single-cell path)

Relationship to the rest of this series

This completes a sequence in which every addition to the coverage model made after logistic was measured against it on exact truth and failed:

change	Δ vs logistic on exact truth	outcome
rank blending	-0.0878 worst family	removed
dominance pruning	-0.0437 worst family	removed
alignment calibration	-0.000143	removed
censoring	-0.000233	removed
splice-junction endpoint	+0.000257	archived, not adopted
`endpoint` kernel	-0.042781	removed here
`hybrid` kernel	-0.003690	removed here
`auto` / `adaptive` kernel	-0.000427, +25.9% cost	removed here
`degradation` kernel	-0.001923	removed here

The one change that did survive was not a feature at all: recalibrating --score-prob-denom from 5 to 3, which improved 32 of 34 samples at zero structural cost (docs/score-prob-denom-recalibration-2026-08-03.md).

The root cause is documented in docs/coverage-auto-rebenchmark-results-2026-08-03.md: these features were selected against none/endpoint/hybrid/adaptive on a panel whose truth is matched-Illumina, and on which none outranks logistic. The strongest baseline was never in the comparison set.