AFP Specification
Purpose
AFP defines a reliability-oriented workflow for AI systems that produce consequential or reusable outputs. The current public specification is intentionally small and testable.
Core workflow
- Assumptions — state material assumptions that could change the result.
- Evidence — attach important claims to inspectable evidence where available.
- Counter-evidence — actively check for evidence or conditions that could overturn the working conclusion.
- Decision — distinguish facts, inference, assumptions, and unknowns before converging.
- Validation — test the output against explicit acceptance criteria and known failure modes.
- Revision — when validation fails or evidence changes, revise and run the relevant checks again.
SafeLoop
SafeLoop is the feedback cycle around that workflow:
Produce → Inspect → Challenge → Validate → Revise → Re-test
The loop is complete only when the result either passes its stated checks or is explicitly marked as unresolved.
Reliability requirements
A conforming AFP workflow should make it possible to answer:
- What evidence supports the important claims?
- Which assumptions materially affect the result?
- What could falsify or block the conclusion?
- Which checks were performed?
- What changed after a failed check or new evidence?
- Can the relevant test be run again?
Scope boundary
AFP does not guarantee factual correctness, safety, or compliance by itself. It is a workflow structure for making those properties easier to test and audit. Domain-specific controls, qualified human review, and external rules remain necessary where applicable.
Status
Public specification status: working specification. This page defines the current minimum structure. Formal schemas, machine-readable states, and conformance tests are planned but are not claimed as complete.