Foray 365 · Essai 04 · Recorded image-reading trial

Three displays.
Three confident errors.

At higher search burden, the table, grouped table and aligned bars each produced five fully correct answers out of six. The errors show why a correct overall verdict is not enough for the next person's action.

Separate the grouping from the picture

The earlier trial found 16/16 correct decisions in both a numeric table and a crew/time diagram. That ceiling did not tell us whether grouping or geometry helped. This trial uses 24 bookings per case instead of six and asks for every conflicting task pair.

Six new cases were each drawn three ways: task-ID order; crew/start order with the same numeric columns; the same grouped rows with time bars added. Every condition shows the same 24 task, crew, start and end facts and the same boundary rule. Three fresh model sessions each saw six PNGs, two per form, and never saw a case twice. All three received the two list images first. The list comparison is confounded with position, even though each case appeared in every form. Answers, errors and confidence were retained without corrective reruns.

These are related constructed cases and three sessions from the same available model configuration. The exact underlying model identifier and reasoning setting were not exposed by the reader interface. This is neither a human study nor a general ranking of diagram types.

Same responses, different standard of success

Visible formCorrectMissing pairsSpurious pairs

No observed accuracy uplift for grouping or added bars on the predeclared exact-pair outcome. The equal scores do not establish equivalence: the cases and errors differ, and the sample is small. No speed, effort or human outcome was measured.

Inspect a captured answer against its original image

Begin with an error, or choose any of the 18 case/form observations. The original image remains available at full size. This is an explanatory reader, not a new blinded participant trial.

Open the exact full-size PNG

Original trial image with24 bookings and the same-crew overlap rule

Actual response

Independent answer and evidence audit

Selections are temporary. Nothing is submitted or saved; reset and refresh restore the starting example.

A correct verdict can still point at the wrong work

In K 3's task-ID table, the reader described the two true overlap intervals, yet attached both to incorrect task IDs. In K 6's aligned display, a reader substituted T 19 for T 06. In K 4's grouped table, a reader invented an extra overlap at an exact permitted handover. Their confidence was 99%,100% and 99% respectively.

Those three answers still implied the correct simple verdict: some clash exists. A manager acting on the named tasks would receive a different quality of help. The explanatory audit also finds a five-minute error in the stated interval of K 4's otherwise correct pair. Fluency and a correct headline did not guarantee correct grounds.

This supplies an empirical instance of the previous essai's conceptual distinction: preserve the information needed for the next question. It also refines the design target. A view should help a reader keep the relationship bound to the right object identities, and should expose the relevant source rows when a finding is used. The observed mistakes motivate that design hypothesis; they do not prove its cause or that a new interface would fix it.

Use disagreement to choose what to inspect

The list missed two pairs and added two, the grouped table missed none and added one, and the bars missed one and added one. Pair-by-pair majority across the three recorded readings recovers every conflicting pair in all six cases. The readers disagree exactly in K3, K4 and K6. This is a retrospective cross-check. Changing both the reader and the display leaves their individual effects unresolved.

Earlier pilot · 16 cases and 32 responses

The simpler trial asked for a crew or a clear schedule. Each form scored 16/16. Inspect its original images, responses and protocol separately from the larger pair-identification trial above.

Open the pilot and its complete response matrix

Open the evidence panel in its own window

What we can—and cannot—carry forward

The higher burden broke the earlier ceiling without separating the forms' primary accuracy. The result therefore earns a null on the proposed representation advantage, alongside a specific failure mode worth addressing. We should not claim that drawing bars solved reference binding: one such error occurred in the aligned form.

The task increased both record count and answer granularity, so this trial does not identify a difficulty threshold. Grouping and sorting change together. The aligned-versus-grouped comparison adds only the redundant bar encoding, but each observation comes from a different session. Fixed assignments balance two views per reader and one view per case/reader; they do not remove every order, case or session effect. Confidence on a handful of related answers is not a calibration estimate.

Source, method and reproduction

Larkin and Simon(1987), especially pp 98–99, distinguishes information from the computational processes its indexing supports, including grouping and matching symbolic labels. This motivates the three forms and the post-trial error interpretation. Their cognitive-efficiency findings are not transferred to these readers.

Prospective protocol · All actual responses and exact scoring · Root's supplementary semantic audit · Frozen materials and reproducible scoring. The archive contains the 18 exact PNGs, case data, assignment, frozen hashes and captured responses. Run python 3 score.py --verify with Python 3.9 or later; it calls no models. The original image generator needs Pillow and the stated font, but is not needed to verify the completed trial.

Pairwise interval intersection and an independent minute-occupancy oracle agree. The frozen input/image hashes are checked before scoring. A response counts as exact only when its whole pair set matches; all raw answer wording remains available. The binary question is a derived secondary score of these same answers, not a separately administered trial.