A single evolved method scores at least 99 out of 100 on all four benchmarks from Phylo's "Biomni x TusoAI" blog, rebuilt as sealed, reproducible tasks:

- 99.71 on genetic perturbation prediction, Replogle-style screen (R^2 x 100)
- 99.71 on the Adamson-style screen
- 99.71 on the Norman-style screen
- 99.66 on enhancer-gene linking, pgBoost-style (AUC x 100)

The worst score across the sealed test split and an independently regenerated dataset is 99.66, and the acceptance gate passes at threshold 99. Noise caps the best possible scores at about 99.84 (R^2) and 99.8 (AUC), so the evolved method sits within roughly 0.15 points of the ceiling.

How standard methods compare, scored by the same sealed harness:

- Linear/logistic regression on raw features: 78.7 to 92.2
- Distance-weighted kNN: 48.3 to 92.2
- Random forest, 300 trees: 83.7 to 99.4
- Gradient boosting on raw features, the closest analogue of pgBoost itself: 95.2 to 99.6

None clears 99 on all four benchmarks or passes the gate; all stay in the repository so anyone can rerun the comparison.

For context: the original benchmarks need gated multi-gigabyte genomic corpora, GPUs, and paid API keys, so each task was reconstructed as a seeded supervised-learning problem whose generative process mirrors the biology the blog describes, scored on the natural R^2 x 100 / AUC x 100 percentage scale.

An agent called KISS Sorcar did everything in one autonomous session: benchmark reconstruction, a TusoAI-style discovery loop, the evaluation harness, adversarial testing, and reviews. I wrote no code.

Why I trust the results:

1. No test-label leakage. The official evaluation path runs each candidate method in a separate OS process that receives only training data and evaluation features; sealed test labels exist only in the parent process's memory. Hostile methods that walk stack frames or scan the garbage collector find nothing; each attack is a test in the suite.

2. Shuffled-label collapse. Permuting the training labels drops the evolved method to chance (R^2 near 0, AUC near 50), so the 99+ is learned, not memorized or leaked.

3. Adversarial testing and independent review. The 27-test suite includes hostile methods that return wrong shapes, NaNs, or constant outputs, and a SystemExit(0) trick that exposed a real bug in an early acceptance gate, now fixed. kimi-k3 hardened the process-isolation path; gpt-5.6-sol read the data generation, harness, and method code and confirmed the evaluation path clean.

The winning method is a degree-2 polynomial ridge blended with kNN for perturbation, and engineered distance/contact features plus gradient boosting for enhancer linking: 157 lines of code, against roughly 1,400 for the harness, data generation, and tests. When a discovery loop proposes methods faster than humans can review them, the evaluation is where the engineering has to go.

Repository: https://github.com/ksenxx/biomni_tuso
Blog: https://kisssorcar.github.io/blog/tuso-evolved-blog.html
