bench

# bench Build a usable experiment from the person's question. Read [the project contract](references/project.md) when authoring or revising the source. Keep the procedure, independent unit, factors, levels, response unit and intended decision explicit. Use the studio to make the work inspectable. 1. Inspect `bench/project.json` and `bench/DESIGN.md`. Preserve existing collected runs, approved protocol, original files and change history. A new brief is open ended; the coffee plan is only an example. Record necessary assumptions and choose an achievable initial run budget. 2. Generate a balanced factorial with independent repetitions and complete blocks. If adding controls, bracket and intersperse them. A two-level design plus shared centers cannot separately identify multiple quadratic effects; inspect matrix rank before collection. 3. Build with `node tools/build.mjs` and give the person a real collection sheet. Do not invent observations. Use blank responses for unfinished runs. Labels, units and ids must survive CSV round trips. A source edit must leave the browser's source-save bridge operational. 4. Import original bytes with `prepareImport` / `commitImport`, or use `recordMeasurement` with a reason for a direct observation or correction. Do not quietly change units, sample identity, settings or response values. Exclude only with a defensible recorded reason; keep the value and compare the sensitivity fit using every recorded training measurem...