Single-sample accuracy refinement (2026-07-23)
This round evaluated two annotation-free refinements in sequence against the adaptive rank-preserving model. Both accuracy and wall-time/peak-RSS costs were gated. Only the second refinement was retained.
Equivalence-component mass conservation: rejected
The first experiment built transcript components from candidates co-occurring on ambiguous reads. After abundance blending, each component was rescaled to preserve its pre-blend corrected abundance. A broad implementation incorrectly rescaled singleton components and visibly lost rank accuracy. A restricted version conserved only multi-transcript components representing at least 0.25% of the library.
The restricted model improved synthetic Pearson by 0.00015, CCC by 0.00016, and RMSE by 0.91 with essentially unchanged Spearman. It also improved H69 cDNA RMSE by 9.4 and H69 dRNA RMSE by 5.1. However, it reduced H69 cDNA Pearson by 0.00141 and CCC by 0.00040, and worsened both PacBio test cases. Building and rescaling the components added approximately 0.5--1.1 seconds and 9--17 MiB RSS on the human screens. The accuracy/cost trade was not acceptable, so the implementation and CLI variants were removed.
Artifacts:
oarfish-evaluation-data/rank-blend-family-screen-20260723oarfish-evaluation-data/rank-blend-family-multitarget-screen-20260723
Agreement-calibrated alignment likelihood: retained
The retained model sharpens a read's alignment-score likelihood only when the score-winning alignment also has both the least unexplained terminal clipping and the longest aligned transcript span. The exponent is strongest for score ties and is regularized toward one as candidate multiplicity grows:
alpha = 1 + 4 * (second_score_probability / top_score_probability)^2
/ sqrt(candidate_count - 1).
Probabilities are updated in place as p <- p^alpha. This combines alignment
score, read-level censoring evidence, aligned span, and candidate multiplicity
without an annotation or truth labels. It avoids the failure of the earlier
held-out-likelihood temperature learner, which selected the flattest candidate
for every sample.
On the 28-case primary panel, relative to adaptive rank preservation alone:
| Metric | Mean delta | Wins |
|---|---|---|
| Pearson | +0.000018 | 17/28 |
| Spearman | +0.000252 | 22/28 |
| CCC | +0.000052 | 21/28 |
| RMSE | -0.795 | 21/28 |
| MARD | -0.000021 | 23/28 |
The six Spearman regressions were all small (largest absolute regression 0.00050). PacBio improved most consistently: mean Spearman +0.000316 (10/11), Pearson +0.000061, CCC +0.000072, RMSE -0.432, and MARD -0.000039.
Outer validation also passed:
| Dataset | Pearson delta | Spearman delta | CCC delta | MARD delta |
|---|---|---|---|---|
| SIRV E2 dRNA | +0.000414 | 0 | +0.000380 | -0.000787 |
| Independent ONT | +0.000166 | +0.000768 | +0.000171 | -0.000058 |
| Independent PacBio | -0.000081 | +0.000430 | -0.000075 | -0.000065 |
SIRV E0 rank correlations were unchanged. A stronger exponent coefficient of eight improved PacBio further but reduced synthetic and dRNA Spearman; it was rejected in favor of the conservative coefficient of four.
Two same-binary paired repeats on synthetic dRNA, H69 cDNA, and H69 PacBio measured a mean 3.1 ms increase in the coverage/calibration phase. Mean peak-RSS changed by -27 KiB (measurement noise), confirming that the in-place scan adds no material memory. Whole-run and EM timing varied in both directions between repeats and showed no reproducible overhead.
--alignment-calibration auto is the default. It activates only for automatic
transcriptome coverage inference and abstains for genome projection, explicit
non-auto coverage modes, and an explicitly supplied --score-prob-denom.
agreement forces the model and none disables it.
Artifacts:
oarfish-evaluation-data/alignment-agreement-rank-full-20260723oarfish-evaluation-data/alignment-agreement-outer-20260723oarfish-evaluation-data/alignment-agreement-cost-{none,on}-20260723oarfish-evaluation-data/alignment-agreement-strength8-screen-20260723