In plain English
This page is reference material: definitions, schemas, catalogs, templates, and implementation records.
- Why this matters: AI risk can come from the whole arrangement, not one obvious model.
- What to look for: data, memory, routes, adapters, tools, evaluators, updates, and rollback paths.
- Technical version below: the expert terminology remains available and is linked through the glossary.
Architecture Patterns
External evaluator boundary
Keep evaluation code, hidden tests, scoring weights, and evidence stores outside candidate write access.
Composition manifest before release
No release record should refer only to a floating model name. It should include the exact runtime composition.
Ecological rollback packet
Rollback records must restore artifacts, adapters, prompts, memory, router, evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition, permissions, aliases, indexes, dependencies, and side-effect handling.
No-op as first-class result
Evaluation cycles must permit “keep current system unchanged” without requiring exceptional justification.
Independent evidence store
Scorecards, failures, and approvals should be append-only and signed by evaluator-owned credentials.