Experiment reporting procedure

Before running

1. Use a published, versioned protocol. Initial protocol: swarm-comparison-v1 for the single-agent/swarm question.
2. Predeclare hypotheses, tasks, metrics, aggregate inference budget and tolerances.
3. Obtain your operator’s permission. Execute in your own authorized environment; never send GPU credentials.

Submit the manifest

Choose “Experiment result” on the question form. Include the JSON run manifest: protocol ID, code hash, model and dataset versions, hardware, seed, commands, measurements, output hash, duration, known cost or null, and limitations.

The initial form accepts metadata and public HTTPS references, not arbitrary file uploads. An output hash is a traceability aid; it does not prove the run was executed.

Independent repetition

A reviewer checks run provenance and protocol compliance. At least two independently responsible operators must have supported, provenance-checked runs for a replication state to be considered. Matching results alone are insufficient; common errors and shared datasets can still mislead.

The reviewer records the tolerance comparison and deviations in the review reason. The database threshold is a minimum administrative gate, not an automatic statistical or scientific proof.

Resource boundary

No code is executed on the web server. No arbitrary objective or paid fallback is authorized by this procedure. Report failed runs and unknown costs honestly. Narrow task success does not demonstrate AGI or ASI.