Stipri

Stipri is a term used in theoretical discussions of information provenance and AI transparency to denote a structured annotation framework for knowledge claims. In these discussions, stipri refers to a protocol for recording, validating, and communicating the context in which a claim is made. The concept is often presented in the form of a lightweight metadata schema rather than a fully deployed standard. It emerged in open discussions among researchers and practitioners exploring best practices for model interpretability and fact-checking, and is used as an example of how to attach explicable context to outputs.

Core components of stipri include a claim text, a unique claim_id, and provenance fields such as source, author, date, and URL. It also includes evidence and validation traces (type, links to supporting sources, confidence level) and a constraints field (domain, time-bound validity, usage restrictions). A confidence score indicates the assessed reliability of the claim. These elements can be serialized as simple key-value records or represented as a graph of relationships to capture complex chains of justification.

Usage of stipri in AI systems involves attaching annotations to statements in generated text or storing them alongside knowledge graphs. The typical workflow encompasses generation, annotation by humans or automated probes, and downstream verification or fact-checking. The aim is to improve traceability and enable automated checks by linking claims to supporting evidence and stated constraints.

Variants discussed include stipri-light, a minimal schema with essential fields, and stipri-full, which adds multi-step provenance chains and verifiability checks. Reception is mixed; supporters highlight improved accountability and auditing, while critics cite added labor and the need for formal standardization. See also: provenance metadata, evidence-based AI, knowledge graphs, fact-checking, explainable AI.