10.6 An audit record, not private chain of thought

Chain-of-thought prompting can improve task performance by eliciting intermediate natural-language steps [ Wei et al. , 2022 ] . Those steps are useful artifacts, but their causal faithfulness to a model’s internal computation cannot be assumed. Generated explanations can omit or rationalize features that influenced an answer [ Turpin et al. , 2023 ] .

CoLT avoids depending on that assumption. Its public record contains only artifacts that can be typed and checked:

• the maintained declaration and versioned realization;

• externally inspectable claims, evidence, sources, and tool results;

• obstruction witnesses and their observers;

• registered repair proposals and preserved invariants;

• validation results, uncertainty, and admission decisions; and

• hashes, timestamps, model versions, or other provenance needed to reconstruct the transition.

A model may generate candidate rationales to populate this record, but the record earns authority from verification and admission rather than from a claim that the rationale exposes hidden cognition.