An observed model pilot
Aligned time.
Same answers.
Putting bookings on shared crew and time coordinates should make a particular comparison visible. In this pilot, it produced no observed accuracy advantage over a table.
Four fresh model sessions judged sixteen fictional commissioning records from screenshots. Every case was read once in each form, by different sessions. All 32 decisions were correct.
correct authorisation decisions
0 confident errors · 16 supported explanations
correct authorisation decisions
0 confident errors · 16 supported explanations
What was tested
Show the operation,
then test the judgement.
The first essai established that complete facts and a useful picture are different things. This trial moved from auditing information to observing readers.
Each record has six bookings and three crews. Authorisation requires that no crew have two bookings in progress together. A table leaves the reader to compare names and numeric intervals. The alternative groups each crew’s work against one shared time axis.
The cases include exact handovers, parallel work on different crews, nested bookings, five-minute overlaps and multiple clashes. Every record is complete; missing information is not the mechanism being tested.
The result constrains the claim
Both forms reached a ceiling. All explanations identified the relevant real relationships: accepted plans accounted for all three crews, and rejected plans named every actual clash.
Every response reported 100% confidence. There were no errors here to reveal overconfidence; this does not establish general confidence calibration.
Four sessions used the same inherited configuration, whose exact model identifier and reasoning setting were not exposed. These are actual model responses, not human observations or independent samples of different models. No speed or effort advantage was measured.
Inspect the evidence
Same case. Two original images.
Switch between the exact screenshots and read the two actual responses. This comparison is for inspection; trial readers saw only their assigned form and received no answer feedback.
Oracle and evidence audit for this case
All outcomes, including the ceiling
The complete response matrix
Each row is one case read in two different fresh sessions. The 32 responses are nested within four sessions and sixteen cases; they are not 32 independent trials.
| Case / mechanism | Warranted | Table response | Aligned response | Inspect |
|---|
What survives
A useful boundary
for the next question.
The result supports a narrow statement: this model configuration could recover the required judgement from either representation across these records, including the negative controls and short overlaps.
It does not establish that the bars changed how the model reasoned. Exact numeric statements remained visible in both forms, so a model could solve every case from text. Grouping, order and bar geometry also changed together.
The next discriminating test should separate those factors and create genuine search demand: for example, a frozen comparison of booking-order text, crew-grouped text and crew-grouped bars across larger schedules. That is a prospective hypothesis, not a retrofit to this ceiling result.
How the comparison was bounded
Frozen before reading.
- Sixteen cases, scoring rules, instructions, allocation and all 32 PNGs were frozen before dispatch. Checksums remain unchanged.
- Each session received eight distinct cases: four per form, with two feasible and two blocked plans in each. Paired sessions reversed case order and form, balancing form at every serial position.
- Readers gave a decision, confidence and free text. No diagnostic answer menu, source access, prior essay context, feedback, retries or replacements.
- An independent minute-occupancy oracle checked the decisions. A later semantic audit read all explanations and checked their concrete crew, booking and time claims against the frozen records.
Practice between related cases, stochastic session variation and ceiling effects remain. We did not test people, timed performance, harder schedules or a general ranking of representations.
Inspect and reproduce
Actual answers, preserved.
The download contains the original screenshots, frozen records and protocol, all four raw response records, their metadata, the manual evidence codes and the reproducible scorer. The response text is unchanged; JSON whitespace was normalized when saved.
Download the complete pilotThe conceptual starting point continues the first essai’s source-use record. Cases, visual design, predictions and measurements here are original to this construction. No source paper’s empirical result is claimed for this interface.