# Evo 2 7B: completed L4 BRCA1 comparison

This draft evidence bundle contains all 3,893 final L4 scores, the pinned historical H100 reference scores, and the minimum public assay labels needed to reproduce their comparison. It contains no model weights, account information, connection details or private filesystem paths.

Run with Python 3.10 or later; no packages, network access or GPU are needed:

```sh
python reproduce.py
```

The script checks input SHA256 hashes, identical ordered variant IDs, binary labels and score arithmetic before recomputing point AUROCs and Spearman correlation. Optional `--output new-result.json` saves a new receipt. `benchmark.json` records provenance, measured runtime and memory, cost boundaries, and reported bootstrap intervals. The compact script reproduces point estimates, not bootstrap intervals or model inference.

| Measure | Result |
| --- | ---: |
| L4 AUROC | 0.8772196737921563 |
| Historical H100 AUROC | 0.8749886211168324 |
| Absolute AUROC difference | 0.0022310526753238635 |
| Spearman correlation | 0.9828845274833258 |
| Prespecified combined comparison criterion | **Not met** |

The criterion required complete coverage, a passing forward test, Spearman at least 0.99 and absolute AUROC difference at most 0.01. The rank-agreement requirement was not met. This result does not show that L4 is less accurate: its observed assay AUROC is slightly higher.

The historical H100 workflow scored every unique reference window before the alternate windows. The L4 workflow interleaved each previously unseen reference with its alternate. FP8 calibration depends on execution history, so this comparison does not isolate hardware as the cause of the score differences. Runtime boundaries also differ; a direct ratio is not a controlled hardware speed comparison.

The L4 run used 8,192 bp paired windows, batch size one and 5,219 window evaluations: 3,893 alternate windows plus 1,326 shared reference windows. The recorded scoring interval was 17,703.652448 seconds, about 4.92 hours, including incremental score/progress writes. Peak PyTorch allocated memory was 16.784 GiB; sampled total device use peaked at 20.762 GiB on the 24 GB L4. These are distinct measurements.

At the recorded rate, scoring alone was about $0.745 per 1,000 variants for compute, or $0.763 including the allocated disk. The complete successful allocation was estimated at $3.0244 from creation through deletion. That allocation total includes setup, weight download, scoring, export and cleanup; it excludes previous failed allocations and other study work. These are captured-rate estimates, not confirmed settled invoices. Shared reference windows make this dataset's per-variant cost specific to this workload.

Labels come from Findlay et al.'s BRCA1 saturation genome-editing assay: 823 LOF-positive variants and 3,070 FUNC/INT-negative variants. This is one gene and one functional assay, not a patient diagnostic cohort. A positive label means assay loss of function, not a separately adjudicated clinical pathogenicity label. The supplementary source is attributed in `benchmark.json`; this bundle includes only factual variant identifiers, positions and classes. Source notices are preserved without applying the Arc software repository's Apache-2.0 license to third-party data. No additional license grant is assigned to Rewire Bio's generated scores or this script here.

The local research harness failed during monitoring. The existing GPU process continued, its final output was recovered without restarting inference, and deletion was verified. The scientific output passed an independent 32-check terminal audit and a separate point-metric audit; the original failed harness record remains unchanged. This is completed recovered T2 evidence, not completion of the broader study or a clean full-study reproduction.