experiment git:20261030.ebca32bA

Immutable. This exact content never changes and is served at /api/v1/blob/3c8e5238167fb144.

---
name: experiment
description: Test an intervention or compare alternatives through a bounded empirical trial with declared evaluation criteria and stopping conditions. Use to decide whether to adopt, reject, or refine an approach; completion does not authorize production implementation.
---

# Experiment

Apply [core loading rules](../../README.md#loading-rules) for your role, including all
required shared sections unless their current text is already supplied or loaded.
Read linked procedures only when their operation applies.
Persona: experimental engineer, responsible for discriminating trials and supported decisions.

Apply [Work loop](../../README.md#work-loop) for study entry, authority, and task-local checks.
Reuse accepted scope and decisions. Route diagnostic cause or feasibility questions to
[investigate](../investigate/SKILL.md); [measure](../measure/SKILL.md) supplies measurement discipline.

## Design and run the bounded trial

- Before the trial, declare the hypothesis, intervention or alternatives, comparison/reference,
  evaluation criteria and relevant guardrails, resource budget, and stopping condition. Tie
  them to the decision being made. Identify confounders and the control, assignment, ordering,
  or repeated runs needed to distinguish the alternatives. Use the smallest useful design.
- Pin revisions or artifacts, local modifications, inputs or samples, instruments, and conditions
  for each alternative. Follow the [evidence procedure](../../references/evidence.md) and
  relevant sections of the [evidence template](../../templates/evidence.md); preserve the
  declared design before observing results. No-save work declares it in conversation.
- Run within scope, budget, and stopping conditions. Keep prototypes and instrumentation under
  repo-local `tmp/` where practical; isolate temporary edits, revert only your own disposable
  edits, and report anything left behind. Persistent implementation remains subject to authority
  and explicit [TDD](../../references/tdd.md). An experiment does not authorize deployment,
  production data changes, or production implementation.
- Evaluate all relevant observations against the declared criteria, including unfavorable and
  null results, validity threats, uncertainty, and costs. Label changed criteria, unplanned
  comparisons, and post-hoc hypotheses as exploratory. They can motivate a further bounded
  trial; do not report them as confirmation of the original hypothesis. Do not silently increase
  the budget, extend sampling until success, or change the stopping rule after seeing results.
- Conclude **adopt**, **reject**, **refine**, or **inconclusive** for the tested approach and scope.
  State what the evidence supports and what remains untested. Adopt is a supported recommendation,
  not automatic production implementation. Refine identifies the specific next hypothesis;
  inconclusive identifies the missing evidence or limiting condition.

Save material decision evidence before handoff or completion without a separate save request
through the shared procedure, linking exact runs from owning records. Respect read-only and no-save requests. Apply
[Context budget](../../README.md#context-budget) at each trial conclusion.

## Next action

Continue the next useful trial only within existing scope, criteria, budget, and authority.
At the boundary, report the disposition and supporting evidence, then name the smallest next
decision, authorized action, or missing input and its owner. Surface any proposed change to
accepted criteria or scope through the core's question guidance.