Measurement first
Design the measurement before you build the tool.
We do not build a fix and then wonder whether it helped. We build the instrument that can tell, then build the fix against it. A measurement designed after the fact rationalizes whatever was already built; a measurement designed first constrains it.
The test: Before building the tool, can you state how you will score it?
Why
Most AI tooling ships on intuition because nobody defined the score. "It feels better" is not a result. Drag that cannot be scored cannot be removed. You cannot tell a real improvement from a plausible one, and you cannot tell when to stop. The measurement is the part that makes the rest honest.
How we apply it
- sysml-bench fixes the task, the conditions, and the scoring before any tool change. The findings in the published results exist because the measurement existed first.
- This reference gates itself on measurements:
just tokens-check,just agent(vacuity), andjust voiceall score the work before it ships.