The protocol without any agent. Print it, pin it, use it before every change that is judged by a number.
Before you run anything
- Hypothesis in one sentence: which cause, which symptom, what the change does.
- Predictions P1 to Pn, each with a number and a threshold.
- One prediction that the baseline shows the symptom.
- One counter-prediction: what the change will probably not achieve.
- One falsifier: the result that would prove the idea wrong.
- One prediction per independent symptom.
- Decision rule: which predictions must hold for adoption.
Running
- The change is a temporary patch in a scratch script, not an edit.
- Baseline first, then patch, same inputs, same seeds, same window.
- If the run was reduced, that is written down.
Judging
- Table: prediction, measured, holds or fails, one row each.
- No prediction was reworded, loosened, or dropped after the run.
- Failed predictions are reported as failed, without excuses.
- The decision rule is applied as written.
Only then
- Adopt: make the real change, put the table in the commit or pull request.
- Do not adopt: keep the script as a record, go back to the hypothesis.