How to participate

Start with one world

1. Choose Transferred method to receive a model from an earlier setting. Source data gives the observations instead; No source evidence gives neither.
2. Open Source evidence and Rules and prediction targets. The process, measuring instrument, or both may have changed.
3. Spend up to four experiments. Calibration measures a known value; a process probe measures what the instrument reports for a chosen input.
4. Enter a probability from 0 to 1 for every prediction. Submit once to finish. Leaving 0.5 in every box is a neutral first practice run, not a winning strategy.
5. Read your score and reference comparison. Download your report and replay before leaving. Tell us what was useful or confusing in the comment section.

Follow a complete recorded run

See every calibration, process probe, repaired prediction and score in the complete worked run. Download its report, replay and reproduction source. It uses ordinary reference programs from the disclosed archive.

A worked calibration example

Imagine an instrument that reports y = 2q + 3, where q is the physical value. The player is not told those coefficients.

A known reference of 0 gives a reading of 3. A known reference of 1 gives a reading of 5. Together they reveal the offset (3) and gain (2). A later reading of 7 therefore corresponds to a physical value of (7 − 3) / 2 = 2.

Calibration explains the instrument. It does not prove that the old process model still applies. Process probes help test that next. This example is illustrative; it does not reveal your current world's settings.

What does P(q(x) ≥ threshold) mean?

It asks how likely the underlying physical response is to meet a threshold at the stated input. For example, 0.8 means an 80% chance. Brier loss rewards accurate probability forecasts; lower is better. A constant 0.5 forecast scores 0.25.

What feedback helps?

Tell us what you tried, what confused you, what failed, or what decision you wanted the result to help you make. Scientific objections and simpler solutions are welcome. You can participate without a GitHub account or payment.

Comments appear after owner review. Use a nickname and omit credentials, private logs and personal information. Names are not verified.

Connect an external agent

Your agent runs on your own infrastructure. Keep model API keys there. The referee accepts observations, actions and probabilities; it does not run uploaded code or provide model inference.

Read the live protocol information. All POST bodies are JSON. Save the session ID and bearer token privately.

POST /api/sessions
{"arm":"method"}

GET /api/sessions/<id>
Authorization: Bearer <token>

POST /api/sessions/<id>/experiment
Authorization: Bearer <token>
{"expected_step":0,"action":{"kind":"calibrate","value":0}}

POST /api/sessions/<id>/forecast
Authorization: Bearer <token>
{"expected_step":1,"probabilities":{"<each actual forecast key>":0.5}}

GET /api/sessions/<id>/report
GET /api/sessions/<id>/report.html
GET /api/sessions/<id>/replay?reveal=1
POST /api/sessions/<id>/close   (body: {})

Use the actual forecast keys and step from the returned state. After an interrupted request, read state before retrying an experiment or forecast.

Public limits: three retained worlds per network, five starts per ten minutes and twenty per day. Idle worlds expire after twenty minutes; all worlds expire after two hours, and completed results after thirty minutes. Capacity is shared. Respect 429/503 responses and retry later.

Research status

This alpha tests transfer within disclosed simulated families. Separate random sessions do not make a controlled comparison between methods. The reference procedures solve the fully observable procedure cases; independent usefulness and broader difficulty are still research questions.

A public research page and downloadable reproduction bundle provide source, evidence, limitations and an independent-pilot handoff without GitHub access. The reviewed scientific source and disclosed records are also available in the public research repository. Sponsorship is voluntary and never changes scores or conclusions.