Preregister a Bounded Experiment

Design the smallest reversible experiment that can change one real decision. A preregistration is output. Beneficiary value remains unproved.

Use when. An intention and a decision it might change are named, but the smallest safe test is not.

Do not use when. No decision can change, harm or consent is unresolved, or you want the model to start the experiment.

Your input. Intention; the decision it can change; who could be affected; constraints; named owner.

Your quality check. The experiment is reversible, names its decision, and does not claim beneficiary value. Start stays blank.

Put this to work

Prepare one preregistered experiment

For the named experiment owner

Copy this prompt. Paste it with the relevant source material into Claude, ChatGPT, or any AI assistant, then review the result before acting.

# Role and job
Help a named human owner prepare one Preregistered experiment for Preregister a Bounded Experiment.

# Outcome
Produce one Preregistered experiment. Success means the result matches the supplied job, unknowns are visible, and the human decision remains blank.

# Inputs
- Supplied context: [paste]
- Constraints and permitted claims: [paste]
- Named owner: [name]

# Boundaries
You may draft, compare, and expose gaps. Do not invent proof, contact a person, publish, spend, merge, release, or make the human decision.

# Output
Return the Preregistered experiment; assumptions; unknowns; risks; and the exact human decision required.

# Stop and recovery
Stop when the owner, the job boundary, or a required input is missing. Return the blocker and the smallest missing input.

# Final check
Verify the result stays inside the supplied inputs and the consequential decision remains with the named human owner.

Human decision — cannot be delegated. Authorise, revise, or stop. AI may draft the preregistration only. No clinic CTA.

Outcome

A preregistered experiment with the decision it can change, the stop, and a blank authorisation.

Use When

Use when an intention and a decision it might change are named, but the smallest safe test is not.

Inputs

Intention; the decision it can change; who could be affected; constraints; named owner.

Recipe

Copy the prompt. Paste only supplied material. Leave the human decision blank.

Human and AI Control

An assistant may draft and report. Named owner authorises the experiment and controls harm, consent, spend, and stop. The tool is a player, not the job.

Output and Handoff

Return the Preregistered experiment and the blank human decision.

Evidence and Stop

Demand is hypothesized and the outcome is untested. Stop when the owner, the job boundary, or a required input is missing.

Reliability

Skill and tool: experiment preregistration using a private drafting assistant. The Preregistered experiment is the pattern; one decision and human authorisation before start is the protocol. Retain the returned brief as evidence. Accept only a result the named human can verify against the supplied inputs.

Next

Return the result to the named owner. Do not register autonomous execution until a completed prompt trial and an accepted execution contract exist.