# Sweep the Parameter and Find the Flat Region ## Claim Sweeping a parameter across its range to find where the output stops changing beats reasoning about which value should be correct. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - The flat region is the useful finding — it tells you the setting does not matter there, which is more actionable than a single tuned value. - Requires the sweep to be cheap relative to the decision; where each run is expensive, reasoning first to narrow the range is the right order. - Guards against the specific failure of adopting a plausible default and never testing whether it mattered. - Applies to effort settings, chunk sizes, timeouts, retry counts — anywhere a number was picked once and inherited since. ## Evidence From Anthropic's guidance on reducing cost and improving performance with the Claude platform: parameter sweeps are the recommended method over reasoning about defaults. ## Links - [[A Specific Number Borrows Authority From a Measurement It Never Had]] — the defect this practice fixes — an untested default that reads as tuned - [[Effort Is a Cost-Quality Dial That Subsumes Prompt Complexity]] — the specific parameter where the sweep pays off most - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] — the tokenomics thread on tuning method.