Ryo

The model your company uses is frozen. It’s the same one your competitor uses. It knows nothing about your codebase, your conventions, or the decision you reversed in 2024. And next year it still won’t.

The obvious fix is to train it on your own work, but that breaks it. Teach a network something new and it writes over what it already knew. This is catastrophic forgetting. It’s been a known, known for fifty years and it’s why nobody ships a model that keeps learning from your team.

Teach a model something new and it gets worse at what it already knew. Ryo is measured on how much of that original ability comes back.

The first working version recovered 62%. The eighth reaches 107.3% — past breakeven, so it ends up better at the old material than it was before it learned the new thing. Even with the human labels removed, a separate run reached 95.6%. That matters because hand-labelling is exactly what doesn’t scale inside a company.

The comparison is compute-matched: the competing baseline is a full retrain handed exactly the same extra budget. Against it, this version posted the best overall score in the program — 3.32 where lower is better, against 3.98 for the same model run without recovery, whose score is dragged down by its own forgetting. Durability costs about 0.11, and that price did not grow when the model got bigger.

These are research-scale runs. The claim is that the curve is real and holds up under pre-registration, not that the absolute numbers survive unchanged at frontier scale. Proving that is the next phase.

Each version was pre-registered before it was run. Success criteria were committed to code in advance of the result existing, evaluation was mechanical against those criteria, not interpretive. Variants that failed their stated threshold were retired.

Baselines were measured rather than assumed. Outcomes contradicting the registered hypothesis are recorded with the same weight as those confirming it. Particular care was taken to avoid false positives due to measurement errors.

Picture the model your team already pays for, but briefed. Before it writes a line of code, it knows the three internal APIs involved, the house convention for this kind of change, and that somebody tried this exact change in 2024 and reverted it.

Routine work gets served locally and instantly. Only the hard problems go to the expensive model. Everything your team approved that day becomes material to learn from that night. Every commit becomes improved performance and it compounds. A frontier model is the same on day 180 as it was on day 1. This one isn’t.

Methods are withheld. The architecture, the mechanism, training objectives and measurement instruments constitute the substance of the work and remain unpublished.

What appears here is limited to outcomes and applications: sufficient to assess result credibility, insufficient to reproduce the model. This is a deliberate constraint on the page, not an omission.