Release Pipeline
Promote the digest you tested through four environments that differ in data and keys, not code, and let a statistic on a fixed suite, not a green build, decide whether a prompt, a tool or a model version ships.
Last updated
After this section you can
- Promote one digest through four environments that differ in data, keys and limits
- Size an eval gate for a 120-case suite and say which difference it can and cannot detect
- Pin every action and keep the eval key small, so a pull request cannot read a production key
30
Promotion and the Eval Gate
Promote one digest through four environments, and let a statistic on a fixed suite decide what reaches users.
Key idea
Promote the digest you tested through four environments that differ in data, keys and limits, never in code. The pipeline is a cost ladder that ends at a gate: a statistic on a fixed suite, not a green build, decides whether a prompt, a tool or a model version ships.
One digest moves through four rungs, and the eval gate decides the last hop
Amber is the decision. The digest is built once and promoted, never rebuilt. Pans sideways on a phone.