Vendor “make it faster” prompts will happily rewrite a loop, swap a data structure, and hand you three “optimizations” without ever running the thing – speculative work that often makes the code uglier and no faster, or faster in the wrong place. This contract inverts that: nothing is optimized until a number is declared and the current system is measured, and nothing is kept unless that same number moved against pasted command output while a named guardrail held. The un-clonable part is the loop – measure, change exactly one thing at the named hotspot, re-measure, revert if flat – plus the rules that catch what casual profiling misses: instrument the steps that look too cheap to matter (that is where cost hides), source your production volume instead of inventing it, prove every number with the command that produced it, and bound the whole thing with a guardrail so “faster” cannot quietly mean “broken”.
Fill in: system_under_testmetric_and_target
Shared as is, for reference. Read it and decide what it will do before you run it; using it is your responsibility, under the terms.