From trajectory to dataset
Every record is an evidence candidate; not every record is training data.
Four different outcomes
A successful tool call, the intended effect occurring, independent verification of that effect and eligibility for training are separate decisions. Each requires its own evidence. training_eligible defaults to false.
Auditable data preparation
Consistent snapshot → usage-rights and verification selection → sensitive-data redaction → canonical state → label review → deduplication and leakage checks → split by task family → schema and semantic checks → versioned dataset. The initial design excludes image binaries from training data.
Separate training and evaluation
Offline training aims to produce a candidate from versioned data and a pinned recipe. Held-out test tasks must not leak into training or tuning. Baseline and candidate are compared with the same environment, permissions and budget. Fewer Supervisor calls alone are not an improvement.
Promotion and rollback
Even a candidate that passes evaluation requires an authorized promotion decision. Activation checks identity, compatibility and health; prior artifacts and configuration are retained for rollback. This site performs no training, registry service operations or adapter loading.