Chain-of-thought represents a decision as selected intermediate claims and their dependencies before the final answer. The application owns this structure: claims, evidence references, assumptions, and decisions that later steps can inspect. It is not a request for the model’s complete private scratchpad; a fluent rationale is not proof of correctness. section 5.2 section 5.9
Make the chain useful by:
- Explicitly representing only important claims whose evidence or invariants can be checked. section 5.2
- Owning the trigger, claim granularity, evidence links, verification rule, and stopping condition. section 5.2
- Skipping it for lookups, formatting, deterministic transformations, and narration that nobody can verify. section 5.2
Payroll calculation
A calculation involving an effective-dated policy, location rule, and rounding convention can expose its policy version, calculation order, and intermediate amount. Another component can then check decisive dependencies that a direct final answer might hide. section 5.2
Explicit claims can reveal missing assumptions and support targeted repair, but add token cost and latency, and can sound convincing while wrong. Keep the chain only when its checkable structure improves accepted outcomes; retain the decision record and external evidence, not a rationale as proof. section 5.2 section 5.9




















