An end-to-end CNN search for ultra-short-period planets in TESS: from the blend ceiling to the independent validation of TOI-6662 b

We trained a convolutional neural network to detect ultra-short-period planets in TESS photometry, following the procedure now standard in the field: synthetic transits as positives, eclipsing binaries as negatives. Measured that way it reaches an area under the ROC curve of 0.97. Keeping the same model and the same planets, but replacing the negatives with false positives that the TESS Follow-up Observing Program has actually rejected, the area falls to 0.50 -- the value of a random classifier. The conventional benchmark had been measuring a problem that does not exist. The reason is physical, and it bounds what is possible. Of 145 such false positives examined at pixel level, 42% are blends: eclipses belonging to a neighbouring star, diluted into the target aperture. Their light curves are indistinguishable from shallow transits, capping the attainable area at 1 - f/2 = 0.79. Retraining on real false positives reaches 0.80 +/- 0.06, confirming the ceiling independently. No classifier reading a single-band light curve can do better. A detector alone is the wrong instrument. We built CASCADE-USP, a pipeline chaining the network with difference-image centroid analysis, scene modelling with high-resolution imaging and the Gaia astrometric solution, supplying in sequence what the light curve lacks. Its output is a prioritisation, not a verdict: it records what evidence each candidate still needs. Only the classifier is domain-specific; the other stages apply to any transit candidate. Of 435 unresolved TESS candidates it isolates 20; 19 prove false positives, traced to causes invisible in TESS photometry. The twentieth is TOI-6662 b, an Earth-sized planet (0.94 +/- 0.03 Earth radii) orbiting an M dwarf every 7.05 hours. Our validation, independent of a concurrent one, agrees with it in planetary parameters and classification, closing the CASCADE procedure with an external check.