Fixations of the Arcana

GLASSGAZE-DEEP is a simulated gaze model. Imagine it pre-trained on a corpus of real Rider–Waite plates and human free-viewing scanpaths hundreds of saccades deep; here that training is emulated by the generative process below, not loaded from weights. Nothing was actually measured — but the pipeline is the one a real model would run, and no stage hand-places an area of interest.

① Feature extraction (simulated CNN backbone)

The plate is decomposed into channels a vision backbone would produce: low-level contrast (multi-octave value noise standing in for luminance/edge energy) plus learned object channels — face, figure, hand, symbol, celestial, creature, scenery — as continuous confidence fields, not points.

② Top-down semantic re-weighting (the querent's task-set)

A Yarbus-style task vector — “a querent seeking meaning” — reweights the channels: faces, hands, and symbolic objects are amplified over raw contrast. The Semantic focus slider interpolates from pure bottom-up salience (0) to full meaning-seeking (1).

③ Priority map

P(x,y) = Σ αᵢ·channelᵢ · centralBias · borderPenalty — the model's pre-attentive belief about where signal lives, before a single saccade.

④ Oculomotor scanpath — hundreds of fixations

A generative eye traverses P under real dynamics:

1. Target selection. Softmax winner-take-all over the live priority map.
2. Inhibition of return. Each fixation deposits a decaying suppression bump, so the eye explores — and revisits once suppression fades (that's what buys hundreds of saccades of genuine study).
3. Saccade amplitude prior. A log-normal over jump distance biases toward short-to-mid saccades, like human oculomotor statistics.
4. Saccadic momentum. A directional-persistence bonus keeps sweeps coherent.
5. Dwell / lingering. Fixation duration grows with local priority and surprise; long dwells fragment into micro-fixation tremor, forming tight clusters.

⑤ Emergent areas of interest

Dwell-weighted fixations are accumulated into a KDE density surface. Non-maximum suppression then discovers peaks — the AOIs — and labels each by the nearest detected segment. Their positions and dwell are outputs of the walk, so they can and do diverge from raw detector confidence.

⑥ Gaze → glass

The emergent density surface (not any authored field) drives the tessellation: seed density ∝ density^γ (finer facets where the eye lingered), jewel saturation ∝ local dwell, came thickness ∝ density gradient.

This is an art object wearing a lab coat that has, this time, several more pockets. The architecture is real practice; the trained weights, the corpus, and every scanpath are generative fiction.