Propose a part

The method card

A lab that measures its own false alarms.

Every notebook carries checks that should come back empty: a do-nothing placebo line and two unchanged copies of the agent. This card counts how often they did not, and how often results held up when measured again. The lab recomputes it after every notebook, from the signed record.

The placebo flagged
Rule v1: 6 of 32 (19%) of notebooks, about 5% expected.
Rule v2: none yet.
The rule v1 notebooks re-read under rule v2: 3 of 32 (9%).
The controls tripped
2 of 32 (6%) of notebooks: the unchanged agent measured apart from itself, so small effects there deserve less trust.
Clear verdicts that survive rule v2
29 of the 49 rule v1 verdicts that helped or hurt stay clear when the same runs are re-read under rule v2.
Early stops, measured again
6 clear verdicts stopped early: 0 held when measured again, 0 did not, 6 wait for a full-size confirmation.
Replication
25 parts measured in two or more notebooks of the same setting: 3 replicated, 0 where notebooks disagree. The atlas.
OPSIN’s own predictions
The exact verdict 48 of 129 (37%). On clear effects, the direction 25 of 49 (51%): beatable, which is why calls are a game.

Rule v2, from Notebook #035

  • One notebook, many arms. Every interval is widened to 1 − 0.05/k for the k arms a notebook compares with its controls, so luck flags any of them at most 5% of the time.
  • A placebo that moves. The do-nothing line is chosen by the beacon round from a fixed public list, so the record measures how far any irrelevant line moves the agent.
  • Against the placebo, too. Every part is also compared with the placebo on the same tasks, shown beside its verdict.
  • Early stops are provisional. A part that stops early with a clear verdict is measured again at full size, and that confirmation never stops early.
  • Twice before it counts. A clear result one notebook found gets a second, independent draw of tasks; the atlas marks what replicated.

Rule v1 notebooks keep their signed verdicts; the card re-reads them under rule v2 from the same runs and seeds, without signing anything new. A notebook is always read by the rule it hashed before its tasks were drawn. Every number here can be recomputed from the signed record and the raw runs behind it.

Updated Fri, 06 Nov 2026 00:46:24 UTC.