Training

From frontier APIs to fine-tuned open weights you own.

Allocate's long game is simple: you start on frontier models through the Gateway, and as your experience database grows, the platform trains models you own. Training is the top of the RecursiveDB compaction path, not a separate product.

How a tenant's model gets trained

• Your history is compiled into replayable environments and a private benchmark. The benchmark's held-out slice is sealed from every optimizer.
• Adapter training runs on verified experience from your database, gated on beating your benchmark before it serves traffic.
• Full post-training distills your optimized setup onto an open-weight base model of your choice from the catalog, then reinforces it against your own replayed environments.

Turning it on

Training is a switch, one per agent. Until it is on nothing trains: no round runs, no adapter is built and no cost lands. Ask us to switch an agent on and the weekly round picks it up from then. The switch names the routes that train, so one agent can train on its member chat and leave its notes out. Each round reads only your own store, trains when there is enough new experience to train on, and writes the records, the win rate and the cost into your daily digest. A week with too little new experience logs one line and skips. Switching off stops the next round; the adapters you already own stay yours.

The guarantees

• Every promotion is eval-gated: it serves only if it measurably beats what came before on your data.
• Every promotion is reversible with a pointer flip.
• Training inherits your region and your boundary. Rollouts run inside your Private Inference Cloud, and nothing about your model is pooled.
• The weights are yours: stored in your scope, exportable at any time.

Training is early. If you are building toward owned weights, talk to us and we will scope the path from your current traffic.