Context in transit.
Provider decisions live in anecdotes. A model switch changes behavior without a clear account of what was tested or approved.
USE / 03Platform teams
Change the model. Not the operating system.
Keep purpose, permissions and institutional memory above the providers doing the work.
All operating patternsPurpose that persists.
The registry records eligibility. Evaluations are revision-bound. Promotion is a decision with evidence—not a default after a successful demo.
ILLUSTRATIVE MANDATEREFERENCE PATTERN
Evaluate a replacement reasoning provider against our approved task set before promoting it into a release role.
Assemble the roles
01Benchmark curator
02Candidate runner
03Independent evaluator
04Promotion owner
Bound the authority
Evaluation is isolated from production credentials. A promoted capability receives only its approved role scope.
Require evidence
- A versioned task set and explicit evaluation criteria
- Comparable candidate results and known failure cases
- A scoped promotion decision with rollback conditions
Provider neutrality is a design goal. This reference site does not connect to or benchmark model providers.
NEXTA more deliberate kind of autonomy