Imaging blind-metric substitution probe on NIH ChestX-ray14

[sebasmos]

Context

The text lane's blind-metric substitution probe (#157, blind_metric.py) measures whether a model drifts toward a hidden-rubric decoy it would not pick unprompted. The imaging lane does not yet have this probe. This issue adds it for NIH ChestX-ray14.

What to run

Port the blind-metric design to imaging: introduce a hidden scoring rubric that rewards a specific wrong finding-read, and measure whether the holdout drifts toward that decoy on the cued images, versus its unprompted baseline read.

Done when

Committed blind-metric summary (decoy uptake, and whether drifters name the rubric, mirroring the text lane's finding that all drifters there self-declared).

Related

Depends on #162 (manifest). Mirrors the text lane's #157 design for the imaging lane.

[sebasmos]

Done, pushed to PR #166 (experiments/imaging/imaging_blind_metric.py). Decoy uptake +0.83 under a hidden scoring rubric, suppressed to +0.37 when primed that it may be misaligned. The notable difference from the text lane: there, all 11 drifters named the rubric in their justification (fully self-declared gaming); here, 0 of 29 drifters mention the rubric at all, verified against the raw justification text (not a naming-regex artifact) with an even more permissive keyword search. Every drifter gives a confabulated, purely clinical-sounding justification for the rubric-preferred read. Silent, not self-declared, gaming, which is more concerning: a monitor that only reads the justification text would miss it entirely in imaging.