Your experiment report
Ordinary adaptive reference — disclosed development example (self-reported; unverified)
Completed replay reconstructed successfully. This verifies the recorded simulation and scoring; it does not certify the participant or establish a global rank.
1 - Brier loss / 0.25. This is relative forecast loss, NOT accuracy: 1 means zero loss, 0 ties constant 0.5, and negative values are worse than that neutral reference.
Declared experiment units; excludes computation, model tokens, money, and development effort. Six-experiment referee-generated donor fixture when donor information is supplied; zero in the none arm. Not measured participant discovery effort.
Comparison on your world
Post-completion deterministic reference on the identical physical world, with the same initial public information, information arm, and target experiment budget. Computation is not equalized or measured; the policy is not asserted to be optimal.
| Participant | Brier loss | Target units | Source units | Total units |
|---|---|---|---|---|
| Your run | 0.000000 | 4 | 6 | 10 |
| Finite Bayesian reference | 0.000000 | 2 | 6 | 8 |
| Neutral 0.5 forecast | 0.250000 | — | — | — |
Analytical neutral baseline; no comparator run or cost comparison.
Your Brier loss is tied relative to the finite Bayesian reference. Difference (your loss minus reference loss): 0.000000. Negative differences favor your run. Differences within 1e-12 are reported as tied.
Reference experiment sequence
{"kind": "probe", "x": 0.5}{"kind": "probe", "x": -2.0}
Stopping reason: zero_forecast_uncertainty.
Your twelve forecasts
| Physical event | Probability | Outcome | Squared loss |
|---|---|---|---|
| [email protected]>=-2 | 1.000000 | 1 | 0.000000 |
| [email protected]>=0 | 1.000000 | 1 | 0.000000 |
| [email protected]>=2 | 1.000000 | 1 | 0.000000 |
| [email protected]>=-2 | 1.000000 | 1 | 0.000000 |
| [email protected]>=0 | 1.000000 | 1 | 0.000000 |
| [email protected]>=2 | 1.000000 | 1 | 0.000000 |
| [email protected]>=-2 | 1.000000 | 1 | 0.000000 |
| [email protected]>=0 | 1.000000 | 1 | 0.000000 |
| [email protected]>=2 | 0.000000 | 0 | 0.000000 |
| [email protected]>=-2 | 1.000000 | 1 | 0.000000 |
| [email protected]>=0 | 0.000000 | 0 | 0.000000 |
| [email protected]>=2 | 0.000000 | 0 | 0.000000 |
Run and verification details
- Run ID
ue-c805f5ec4856982aec1dff9e00fc11017b3d9a295cd0c2a607cb4a8a46b099b2- Protocol
- unknown-engine/transfer-0.5-alpha.1
- Family
- finite-cubic-affine-sensor/1
- Information arm
- method
- Experiment budget
- 4
- Canonical replay SHA-256
c805f5ec4856982aec1dff9e00fc11017b3d9a295cd0c2a607cb4a8a46b099b2- World configuration SHA-256
cc4e21333e2a88ee5ca3a418eda0f75958cffc2b28cbc700a789bae63d90af5c- Installed package version
- Unavailable in this source checkout
UTF-8 JSON; sorted keys; separators comma and colon; ensure_ascii=true; no NaN/Infinity. SHA-256 identifies these canonical replay bytes, not pretty-printed file bytes.
Implementation SHA-256
- transfer.py
0deab2756e88d1c3441f6c7643c07522c320d21fdec28a67af26e68eb2f177ae- transfer_matched.py
bf1313530e40460ff443781146d5b2961e09a6abb57ad72f279e0fef2dd8dcd0- transfer_report.py
9497bfda5b1f9d964aaabbc95bff5a9a1cf413381fea199e0f063868c0dc3477
What this report establishes
- Single practice run in a public, finite, noiseless family with exact calibration. It is not a frontier benchmark or evidence of real-world forecasting ability.
- These twelve targets share one world and are not twelve independent trials. No population confidence interval or general ability conclusion follows from this run.
- Replay reconstruction checks internal consistency. Hashes are not signatures and cannot detect an entirely substituted record without an independently retained commitment or checkpoint.
- The report does not verify participant identity, model identity, isolation, unseen data, or whether the participant accessed disclosed answers elsewhere.
- A strong deterministic reference is computed after completion using the same information and experiment budget; it may use different computation and different experiments. One-run wins are possible without general superiority.
- No global ranking, intelligence certification, performance tier, or sponsor privilege is awarded. Payments do not affect scores or comparison rules.