Receipts
We fixed the rule, the configuration, the endpoint and the test in writing, committed that file, and only then ran the seasons it would be judged on. This page is what came back. It says we do not beat the free Sleeper consensus. It is published because a record you only show when it flatters you is not a record.
| Position | Our top-3 | Consensus top-3 | Our avg finish | Theirs | Verdict |
|---|---|---|---|---|---|
| QB | 8 | 9 | 7.94 | 9.47 | NO EDGE DEMONSTRATED |
| RB | 6 | 9 | 18.24 | 11.71 | NO EDGE DEMONSTRATED |
| WR | 9 | 9 | 26.06 | 19.97 | NO EDGE DEMONSTRATED |
| TE | 9 | 13 | 12.32 | 9.03 | NO EDGE DEMONSTRATED |
| All | 32 | 40 | outright #1s 12 to 12 | BEHIND CONSENSUS | |
Each pick is matched against the consensus pick at the same position in the same week and graded on the same rule: the player who actually finished #1 in PPR. 136 paired picks over 2 seasons neither of which had been run here when the rule was written.
An earlier version of this record claimed 23 top-3 finishes against the consensus’s 18, and said the edge lived at running back. Two audits refuted it and we withdrew it.
- It leaked. The availability gate read a roster snapshot written after the games were played — a post-kickoff answer key. Removing it cost 1.8 points of running-back average finish and 3.8 at tight end.
- It was chosen on the season it was scored on. Re-running the same four configurations on a different season inverted the ranking almost exactly. The margin was selection noise.
- The running-back edge did not replicate. On the configuration selected out of sample it became 7 podiums against 5, with a worse average finish and nearly double the busts.
The withdrawn document is kept in the repository as a tombstone rather than deleted, so the record of what was claimed, and why it was wrong, is not quietly lost.
Three things we measured that stop us making claims other people make freely:
- Last season’s tendencies mostly do not carry over. Only pre-snap motion persists strongly year to year. For play-action, deep rate, screen rate and the coverage families, a team’s own prior number predicts next season worse than the league average does.
- Points allowed to a position is close to noise at two positions. Year-over-year persistence is 0.08 at quarterback and 0.05 at receiver — neither distinguishable from zero. We no longer let those numbers carry an argument.
- “Elite against man coverage” is mostly “elite”. For most split types a player’s overall rate predicts next season’s split better than the split itself does. Only target depth survived the control.
Every number in a brief has to exist in that week’s measured dossier or the brief is rejected and rewritten before you ever see it.