Glossary

Technical primitives

Cluster

A specific specialisation path defined by a model family, benchmark target, and customer-controlled execution boundary.

Composite score

A weighted score that summarises benchmark performance, robustness, and operational requirements. Promotion gates compare the candidate against the configured threshold and frontier baselines.

Evidence bundle

A review packet attached to a candidate model. It should include benchmark version, score movement, lineage, latency, cost, rollback posture, and approver state.

Fleet boundary

The customer-controlled environment where training and inference execute. Shared public or owner-safe evidence should stay redacted to summaries, scorecards, and review labels.

Iterate ledger

The recorded state trail for a specialisation run: evaluation, diagnosis, specialisation, assessment, blocked states, and promotion decisions.

Promotion gate

The approval boundary a candidate must clear before production use. Gates combine benchmark score, operational checks, rollback posture, and human sign-off.

Scenario

A realistic task with domain input, expected evidence, pass criteria, and common failure modes.

Specialist model

A model accepted for a narrow workflow because it beats the configured benchmark and carries enough evidence for approval and rollback.