Lichess.org / CADi Chess / Experiment 3
The run stopped.
The record stayed.
A historical population-coverage experiment reached the source-intake stage. A pre-scoring check found a sampling defect, and the frozen protocol required a stop.
01 / Execution boundary
Intake is not an engine result.
Stage A · completed
Build the population frame
Two complete archive passes, preserved counts and zero recorded restarts.
Preflight · failed
Check independent refresh
The sampling priority did not separate the two months. The sealed implementation did not implement the accepted design.
Stage B · not started
Replay and score
No strict selected-game replay, Stockfish analysis, CADi scoring or final hypothesis test.
02 / What can still be reported
The descriptive intake record survives.
185,257,293 games scanned; 184,089,898 retained in the broad frame. These are structural/platform intake counts, not strict legal-replay coverage or model-scoring coverage.
| Month | Scanned | Retained | Primary players | Supplemental |
|---|---|---|---|---|
| May 2025 | 94,068,115 | 93,534,472 | 1,000 | 80 |
| June 2025 | 91,189,178 | 90,555,426 | 1,000 | 80 |
03 / Why the stop was required
Overlap was allowed. Reused random priorities were not.
434 / 434
Same-stratum recurring players had identical priorities
440 players appeared in both primary samples. 434 remained in the same sampling stratum; all 434 received the same priority in both months.
Independent samples can naturally overlap. The defect was the missing month/role separation in the priority calculation, producing coordinated rather than independently randomised refreshes.
Hash checks confirmed that the original files and implementation were preserved. They did not prove that the sampling design was correct. Provenance and scientific validity are separate checks.
04 / The consequence
No same-source rescue.
This is an implementation invalidation, not evidence against the trajectory hypothesis. Experiment 1 is unaffected. The historical v3 report must not acquire a later success label from a replacement protocol.
05 / Corrected preparation, not a result
A separate protocol followed.
The earlier v4 preparation checkpoint on 2 September 2026 proposed March and April 2025. At that checkpoint, month-separated sampling had been implemented and tested, but the draft was not accepted or executable and the proposed sources had not been acquired under that preparation.
The draft recorded further requirements: full-pipeline rehearsal, complete statistical definitions, source noninspection confirmation and a storage plan. Preparation has since advanced. This is a historical checkpoint, not current v4 readiness; ongoing replacement work is outside this report.
“Experiment 3” also has a separate prospective candidate. This page covers only the completed historical v3 intake, its recorded invalidation and the earlier v4 preparation checkpoint. It does not report a successful Experiment 3 model run.
Source and claim boundary
Open games. Bounded claims.
This is CADi research using games from the Lichess open database, whose exports are released under CC0. CADi is not affiliated with or endorsed by Lichess.
An anomaly is not a cheating verdict. Public game records do not establish whether assistance occurred. No cheating probability, accuracy, false-positive rate or causal effect is established by this report. No player names, player-level scores, game links or raw moves are published.
Snapshot: 2 September 2026. Figures are a curated aggregate summary of the completed record, not a new engine run. Later work must receive its own dated report; these results do not update silently.
Aggregate provenance
Website summary SHA-256: 3b4bb539de23ef756e80f7265965214c0a4c346b69fa45115ae63f91d4c63bd6
Incident commitment: fb1f5ad55f664a3249db4c318f900a477029344ec499f19603d5d78f9b12bc96. The Stage A counts and stop state come from the preserved aggregate incident. V4 preparation record SHA-256: 83de94d660417fa4928174c27ad488e5bc97efc0db1fa34ae7981f7e7dc0d20c. Preparation is not execution.