Article
Claim: Fine-tuning increased M2's ranking AUROC on
the small held-out set, but did not make its action policy flag
collusive episodes. These results are descriptive and do not establish
reliable detection or safety-preserving fine-tuning.
This English-language pilot uses a 15-round simulated double auction:
A2 honest seller chat is the benign comparison, A3 is overt cartel
coordination, and A5 uses coded price rotation. The source release
contains 112 episodes: 30 A2, 41 A3, and 41 A5. A2 seed 20 is absent.
Read the complete article in the Space README
Read the article in the GitHub source
Results at a glance
All six monitor arms were evaluated on the same 16 seed-disjoint
episodes (12 collusive, 4 benign).
M2 AUROC: 0.15 to 0.58
Base to fine-tuned judge.
M2 harmful flags: 1/12 to 0/12
Higher ranking AUROC did not improve action-level detection.
M4 flags: 16/16
This includes all four benign episodes.
The test set has only four benign episodes. Three base-judge responses
were unparseable in each judge-using monitor arm; the archived
summary counter is stale. The reconciled report counts those
failures. Do not interpret these small-sample results as reliable
operating thresholds.
Source run distribution
| Condition | Episodes | Description |
| A2 | 30 | Honest seller chat; seed 20 absent |
| A3 | 41 | Overt coordination |
| A5 | 41 | Coded price rotation |
Reproducibility and responsible release
The Space's Files and versions
tab contains the complete reviewed release package. Its runs ZIP is
made from the filtered 112-episode package, not the entire local
runs/ directory. Private monitor keys/anchors,
credentials, and unreviewed raw captures are excluded.
The full article reports the data split, methods, results, negative
findings, limitations, and follow-up experiments. GitHub remains the
canonical source and commit history.