# GB1 benchmark results Executed November 7, 2026 UTC with Python 3.12.14, `additive-ridge-v1`, alpha 10. The [protocol](PROTOCOL.md) was fixed before evaluating test outcomes. The [machine-readable report](results.json) includes every split ID, fitted coefficient, input checksum, and per-run metric. [Test predictions](predictions.csv) are retained for independent inspection. No hyperparameter search or subsequent model selection was performed. ## Within-scaffold interpolation Each of three seeds draws 384 training variants and 1,000 disjoint test variants from the 149,361 measured records. Values below are means across seeds. They are not confidence intervals or independent biological replicates. | Method | Spearman rank correlation | MAE in relative-enrichment units | Mean measured outcome of top 24 | |---|---:|---:|---:| | Additive ridge | 0.371 | 0.145 | 0.664 | | Nearest training sequence | 0.333 | 0.112 | 0.495 | | Training mean | Undefined (constant) | 0.130 | 0.125 | | Existing composition heuristic | 0.143 | Not comparable units | 0.287 | | Random-selection expectation | — | — | 0.0777 | Ridge Spearman ranges from 0.323 to 0.420. Its top-24 mean ranges from 0.481 to 0.925, and 4, 4, and 6 of the 24 selected variants exceed WT (=1) in the three runs. All random-regime test sequences have residue/mutation-count support. The ridge model improves average rank correlation and shortlist outcome over these baselines, but **its MAE is worse than both the nearest-sequence and mean baselines**. Therefore the numerical predictions are not established as accurate assay-value forecasts. This mixed result is a reason to evaluate ranking and calibration separately, not to claim general superiority. ## Mutation-distance extrapolation stress test Training uses 384 variants with at most two mutations; testing uses 1,000 variants with three or four mutations. Ridge's mean Spearman drops to 0.282, with MAE 0.320. Nearest-sequence Spearman is 0.245 with MAE 0.285; the legacy heuristic has Spearman 0.141. The ridge top-24 mean ranges from 0.194 to 0.758. All 1,000 test variants in every stress run exceed the training mutation-count range. **The production shortlist gate excludes them.** Their raw predictions are computed here only to characterize a failure mode, not to demonstrate supported product recommendations. No evidence here establishes generalization to other scaffolds, assays, or protein families. ## Concrete errors In random-regime seed 20261107, variant FWAG has measured enrichment 6.509 but the model predicts 0.438. In seed 20261108, IYGL measures 3.708 while the prediction is -0.014. In seed 20261109, FWCA measures 7.555 while the prediction is 0.566. These are four-site codes at the study's positions 39, 40, 41, 54. The unbounded additive fit can produce negative predictions for this nonnegative assay and misses some highly beneficial combinations. Individual errors must not be hidden by an average correlation. All six fits converged (20–33 sweeps). Exact train/test sequence overlap was zero. Feature support was derived only from each training set; no pretrained weights or imputed outcomes entered the analysis. The three seeds reuse one historical dataset, whose labels may have study-specific preprocessing and measurement noise. The same scaffold on both sides is intentional and is disclosed. These results are **retrospective scientific evaluation**, separate from the software tests for parsing, security, storage, routing, and numerical correctness. No prospective wet-lab validation, time-savings study, customer willingness-to-pay test, or proof of therapeutic benefit has been completed.