Models improve through datasets, evals and canaries — never silently

Model improvement is how One3 gets better without breaking trust. Nothing changes in production without a versioned release.

Production traces are sampled and failures are clustered. Humans label examples. A dataset version is created with a datasheet, a PII scrub and a licence check. An experiment runs — prompt optimisation, or LoRA fine-tuning (SFT or DPO) on the build machine. The candidate is tested on an eval suite for quality, safety, cost and latency. Governance approves it, it enters the model registry, and it runs as a canary on 5–10% of traffic. Outcome metrics are monitored and the model is promoted or rolled back.

Tenant data is never used to train shared models without opt-in. Provider terms are checked before any model output is used as training data. AI Lab fine-tuning is limited in Year 1 to at most one classifier or extractor, and only if labelled data justifies it.

Every step on this page is an Action: a typed verb with a risk tier (R0 read-only to R4 irreversible or financial), a permission check and a reversibility class. Agents and humans call the same Actions, so the same approval rules apply whether a person clicks a button or a Digital Employee acts.