Some tasks should just be quick
Spend extra reasoning on ambiguity and costly mistakes.
Retrospective reconstruction of the week beginning 2025-03-10, written on 16 November 2026. This is not a claim of publication at the time.
Claude 3.7 Sonnet and GPT-4.5 had arrived in late February. Comparing their capabilities also raises a less exciting product question: how long should this particular task take?
Pulling a clearly labelled date from a document should be quick. Reconciling contradictory migration requirements deserves more attention. Treating both as the same kind of job either slows down routine work or rushes a decision that will be expensive to undo.
Start with a few real examples from each group. Check whether extra effort improves the result. Include failures and review time in that comparison. A longer answer is not evidence that the extra wait helped.
Keep the routing simple until there is a reason to complicate it. A fast path for straightforward transformations and a more careful path for ambiguity may be enough. Every additional decision rule is something else to maintain.
If the person has to wait, tell them what is being checked. Then return the result. Nobody needs a performance of deep thought while a system takes thirty seconds to extract a field that was already labelled.