Training-mean control: automated execution audit

What this source contributed

Records in this release that cite this source. A record can cite several sources.

• 3 results
• 1 evidence claim

Results

Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.

Claims

Statements this source supports, each with where in the source it was checked.

• Evidence summary: Set baselines for UTR translation experiments: Summary Only two sequence-only controls have been measured on the target split, both in Rewire's own runs, so there is no model comparison on it yet. On the 15,003 held-out records of the mRNABench designed-reporter test split, the composition ridge control reached a mean squared error of 1.914 against 2.366 for the training-mean control, and a Pearson correlation of 0.437; the training-mean control has no defined correlation, because its prediction is constant. The published FramePool panels are the only model comparison recorded here: neural models (FramePool and Optimus) reached Pearson 0.743 to 0.966 on the two random-sequence panels and 0.700 to 0.894 on the two human panels, while k-mer random forests, which were evaluated on the random panels only, reached 0.616 to 0.877. Those values are reported in the FramePool study on its own library, one split per panel with no uncertainty interval, and they are not measured on the designed-reporter split, so they are not baselines for it. Both Rewire controls are single runs on one split with no seed variability, and their automated audit checks the run itself, not a published score. (rewire-local-20260920 run outputs: metrics.mse and metrics.pearson for the composition and training-mean controls; uc20260930-source-framepool-s1: S1 Table rows Random 50nt, Random 25-100nt, Human 50nt and Human 25-100nt; uc20260930-source-framepool-s4: S4 Table columns B and C)

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.