# Grouping versus geometry — frozen before responses

18 November 2026. Primary outcome: exact identification of every conflicting task pair from a screenshot. Secondary: missing/spurious pairs and stated confidence. No elapsed time, human performance, or unobserved internal effort claim.

Six new fictional cases contain24 bookings across four crews (the earlier trial contained six bookings). Three forms contain identical task, crew, start and end strings and the same half-open interval rule: an ID-ordered table; a crew/start-ordered table; that same grouped table with aligned time bars. Each case is presented once per form to different fresh model sessions; each session sees each case once, two cases per form. assignments.json fixes the order. Readers receive images only, no canonical data, outcomes, feedback or other readers' responses. Record all answers without rerunning mistakes. This is a bounded pilot with related cases and sessions from the same available model configuration, not independent populations.

Pairwise overlap and an independently implemented minute-occupancy oracle agree on all six answer sets. freeze.json fixes case data, assignment and18 original PNGs before dispatch. Images are generated with Pillow, not browser screenshots, and are the exact visible trial stimuli. Task IDs are shuffled independently of crew/time. All visible facts are present in each condition; bars add a redundant spatial encoding. Grouping and sorting change together in the grouped condition; their separate effects are not identified. Aligned minus grouped isolates the added bar encoding within this construction.

No universal representation ranking follows. An all-correct result is a null for observed accuracy uplift, not proof of equivalence or of no effort benefit. The reader will show actual responses and exact stimuli; no invented participant records.
