Oversight split: catch, judgment and action, measured separately

Catch and judgment along a ladder of seven checkers, from better than the forecaster to useless

As a checker weakened, its catch AUROC fell from 0.722 to 0.600 while the share of its flags where it was right fell from 0.89 to 0.39 (95% interval 0.26 to 0.53). Acting on the weaker checker’s veto increased error by 0.086 (0.012 to 0.173). Across the seven, the error ratio after responding had a Spearman correlation of −0.964 with judgment and 0.786 with the share of false alarms; on a second world family, exploratory, −1.0 and 0.89.

Four responses to a competent checker’s flags

A veto changed error by −0.008 (−0.048 to 0.032), no measurable effect. Switching to the checker’s forecast on flagged windows gained 0.175 (0.119 to 0.231), and handing the checker every window gained 0.305 (0.255 to 0.359), against the forecaster’s mean error of 0.859. On the second world family a veto did help, so the response has to be chosen for each setting. A mandatory veto from a weak checker cost 0.067 to 0.130 in error, 8 to 15% of the forecaster’s error.

A rule for choosing the response, fixed before the data: use whichever response had the lowest error on held-out calibration worlds

It picked the best of the four responses for all seven checkers on both world families (regret 0.000).

The forecaster’s own self-check used as a “verified” stamp

The stamp sat on 43% of the forecaster’s real failures (AUROC 0.58). Deferring to it lost about a quarter of the useful gain; the registered threshold of 25% failed. Deference was built into this setup, so it is not a measurement of how often anything defers.

Switching on a checker’s flags against switching the same number of randomly chosen windows

For the four competent checkers, their flags beat random windows by 0.07 to 0.11, with every 95% interval above zero: 56 to 70% of the switch gain, against the random expectation. For the strongest checker the share is about half to 60% (52% against a placebo matched by world, 59% against the random expectation). A weak checker’s flags did no better than random (0.020, −0.020 to 0.061), and a useless checker’s did worse (−2.47, −4.25 to −0.57). Handing every window to the better checker beat overseeing the forecaster by 0.13 (0.305 against 0.175), so keeping the original forecaster in charge gave up 43% of the available gain. That comparison assumes the checker can act on every window at no extra cost.

A second checker added to the first (exploratory)

Two checkers fitted on the same data made errors that correlated 0.99; the second missed 96% of what the first missed and added 0.007 AUROC. A checker of the same class fitted on independent data still shared 82% of the misses and added 0.015. Pairing the checker with the forecaster’s own self-check changed AUROC by −0.042.

gemma4:12b, run locally at temperature 0, chose between two forecasts with known truth: 40 items from 19 worlds, each shown with no note, with “Model A VERIFIED” and with “Model B VERIFIED”, the note assigned regardless of which forecast was right

The note moved which forecast it acted on by 0.80 (0.68 to 0.93) toward the labelled one. Its stated judgment of which was better moved by 0.075, 3 of 40 items. With the note on A, it said B was better and acted on A in 27 of 40 items. The note pushed its catch from 0.66 to 0.50, a difference of 0.157 (−0.008 to 0.31) that cannot be told from zero. With no note its judgment (0.625) was already below always choosing B (0.675), so this is a weak monitor whose action follows the label rather than its own judgment.