ICF-AI: An Evidence-Status-Aware Protocol for Documenting Causal Claims in AI Systems

ICF-AI proposes an evidence-status-aware research-record protocol for documenting causal claims about artificial intelligence systems.

The protocol distinguishes observed events from proposed causal mechanisms and records evidence provenance, identification assumptions, alternative explanations, observability limitations, and the status of each claim. It organizes investigation using eight optional analytical stages, from system conditions and model formation through deployment, interaction, outcomes, institutional effects, and feedback.

Three explicitly fictional examples demonstrate how observed behavior can support limited findings without establishing a specific causal mechanism.

This paper presents a conceptual methodology, not an empirically validated causal inference technique. It does not claim improved reliability, demonstrated superiority over existing approaches, or independent experimental validation. A future comparative evaluation is proposed.

Funding: No external funding.

Competing interests: The author developed ICF-AI and separately offers a commercial reference edition, which is not included in this deposit.

AI assistance: AI tools assisted with drafting, organization, and preparation of illustrative research records. The author is responsible for the manuscript's scholarly statements and citations.