AI Training Claim Ledger Framework

AI Training Claim Ledger Framework

Every material training claim should record the exact sentence, claim type, model and version, source, section or line, publisher or independent status, method, metric, scope, uncertainty, missing evidence, currentness, allowed paraphrase, prohibited inference, reviewer, and refresh trigger.

Separate architecture, data, compute, training, post-training, evaluation, safety, license, deployment, and commercial claims. A source that supports one layer should not be used to infer another.

The ledger gate asks whether the public sentence says more than the evidence. If it does, narrow the sentence, add the missing source, label the uncertainty, or remove the claim.