Chain-of-Thought and Reasoning Models Explained

Reasoning prompts and reasoning-style models spend more tokens or internal compute to search through intermediate steps before producing an answer. That extra work can improve math, planning, code repair, and multi-constraint tasks, but it does not turn the model into a theorem prover.

The running example is a refund-policy assistant deciding whether ticket R-2041 qualifies for a partial refund. The useful product contract is not the private scratchpad; it is the final answer, cited policy, confidence, and escalation rule.

Modern reasoning systems often separate internal deliberation from the visible answer. That separation matters because raw scratchpads can contain false starts, sensitive prompt details, or persuasive but unverified claims. A short explanation can still be useful, but it should summarize evidence rather than dump every internal step.

For R-2041, expose the policy clause, decision, and unresolved question. Keep the private exploration private. If the model changes its mind during reasoning, the final response should include the verified reason, not the whole wandering path.

Trigger: the assistant receives a policy excerpt that says refunds are available within 14 days, but the ticket date is in a different timezone. Symptom: the answer includes a confident multi-step explanation and approves the refund. Root mechanism: the reasoning path optimized coherence around the wrong date conversion.

Recovery is to verify intermediate facts with deterministic code or a rule engine before allowing the answer. Prevention is to separate fact extraction, rule evaluation, and response generation so the model can explain a checked decision instead of inventing one.

Create ten refund examples with one hidden trap each: timezone, missing receipt, excluded item, policy version, or duplicate request. Run the model once with a short answer prompt and once with a reasoning-summary prompt. Compare only the final decision, evidence, latency, and escalation rate.

Pass when the extra reasoning improves accuracy on hard cases without increasing unsupported approvals. If it only produces longer explanations, route the task through tools or a smaller deterministic workflow instead.