Design the experiment

Contract: references/interaction-contract.md. Calibration, templates, provenance: references/experiment-manifest.md.

Trigger

• A planned comparison: prompt A vs. B, M0 → M1, retrieval on/off, a new guardrail.
• "Write a hypothesis." / "Create a pre-analysis plan." / "Name the stopping rule."
• A sweep about to run with no plan for how the winner gets confirmed.

Do

1. Convert the change into one tradeoff-aware hypothesis: directional, minimum effect, naming what may not regress, plus population, proxies, and scorers. This is the release gate stated in advance.
2. Name independent, dependent, and control variables (reference.md). Any control that matters must be a schema field, or it cannot be held constant or sliced by.
3. Pin what hides inside apparently atomic variables: the serving path behind a model string, the tool manifest behind "the agent." If more than one thing changes per arm, the design cannot attribute the effect to any single factor — isolate first.
4. Choose pairing (same items across arms — the default, far more sensitive) and K = 3–5 runs for anything gating.
5. Verify the treatment is implementable in every arm. Publish the per-arm implementation table, and exclude arms that cannot receive it from the treatment-effect claim rather than recording them as "no benefit."
6. Pre-specify stopping, exclusions, exploratory vs. confirmatory, multiplicity, and estimators — the elicitation regime dictates the estimator.

Avoid

• Do not choose the analysis after seeing outcomes, or relabel a sweep as confirmatory.
• Do not fix a defect mid-comparison. Restart under the corrected config, or finish as designed and record the defect — a mid-run correction leaves no trace in the results table.
• Do not run unbounded parallelism; it is the most productive source of fake failures.
• Do not analyze results here.

Check

• One hypothesis with direction, minimum effect, and guardrail bounds.
• All three variable classes enumerated; every control exists as a field.
• Serving environment and tool manifest pinned per arm.
• Pairing and K stated; stopping rules fixed; exploratory vs. confirmatory labeled in advance.
• Per-arm treatment implementation table drafted; manifest reproducible.

Risk

• Flexible stopping, undocumented arms, and uncontrolled changes make an apparent improvement indistinguishable from selection noise.
• A treatment with no uniform implementation quietly becomes several treatments.
• Metered ceilings fail in the worst available way — partway through one arm, leaving a partial result that resembles data.
• Confounds enumerated on a whiteboard stay there unless they become fields.

Braintrust

Write H1 into the experiment description so the pre-registered claim travels with the results — the cheapest integrity control available. Mark the confirmatory run explicitly in its name and description.

This is the stage that has to get the shared mechanics right, because they are decided here and cannot be repaired afterwards. From references/platform-mechanics.md: §3 pinning, since pairing must be chosen before running and cannot be applied retroactively; §4 metadata for every independent variable, serving path and tool manifest included, and every control named in step 2; §5 naming, or the arms never line up; §6 separate sweep experiments, losers kept, so the denominator survives to the report; §7 hygiene for concurrency and smoke-test deletion.

Each of those is cheap now and impossible later. That asymmetry is the whole argument for designing the experiment before running it.