# Foray 365 · Essai 02: does alignment externalise the judgement?

Frozen before any reader runs. This is a formative comparison of four fresh model sessions, not a human study or four independent samples from a population of models.

## Design and prediction

Sixteen complete fictional commissioning records each contain six bookings, three named crews, one rule and explicit boundary semantics. Four matched pairs are feasible; four matched pairs contain one or two capacity conflicts. The pairs use renamed crews, shifted times and shuffled booking identifiers. Negative controls include touching boundaries and overlapping work assigned to different crews. Positive cases include a nested booking, a five-minute overlap, an overlap chain and conflicts involving two crews.

The table preserves booking order. The aligned view groups bookings by crew and puts their bars against one shared time axis. Both show exactly the same four rule statements and six complete booking statements. Only grouping, ordering, redundant bars and image height differ. No conflict is highlighted or computed on a participant image. The intended mechanism is recognising same-crew interval overlap through shared coordinates. Exact numeric statements remain available in both forms; an effect could come from grouping rather than bar geometry. A future ablation would separate those components.

Prediction: aligned form may improve conflict discovery or reduce false conflicts. This is a hypothesis; a ceiling, reversal or no observed difference will be retained. No case, rubric or allocation changes after reader observations.

## Allocation and access

Four fresh sessions, eight distinct cases per session. R1/R3 receive K01–K08; R2/R4 receive K09–K16. R3/R4 reverse order and invert the form. Each session has four cases per form, including two Yes and two No per form. Each case is seen once in each form by different sessions. View and position are counterbalanced within each paired-reader cohort. There is no same-case within-reader switching or answer feedback. Related structures can still produce practice; between-session stochastic variation remains.

Readers receive only neutral instructions and eight PNG paths, in assigned order. They may inspect those images using `view_image`; they may not access source, OCR tools, files, network, rubric or other readers. Images include all business rules; instructions offer no diagnostic menu. They record their judgement first, then confidence and free text. No response or reader is generated by script. Their inherited model and reasoning settings are recorded by the orchestrator. A fresh context prevents prior essay exposure; it does not make sessions independent models.

## Frozen outcomes

Primary: correct authorisation decision against the independently checked oracle. Secondary: wrong decision at confidence >=80; concrete conflict evidence or invented conflict in free text, coded with exact excerpts. Unresolved is a selectable answer but all cases are complete and determinate. Asking to verify facts is recorded separately, not automatically counted as an error. No speed claim: model/tool latency is not a measure of reasoning effort. No private reasoning transcript is requested or published.

The unit reported is the response, nested within four sessions and sixteen cases. Describe paired case outcomes and per-session outcomes; do not treat 32 responses as independent trials or claim population significance. Preserve all raw final responses. Report tool failures, unreadable images and deviations rather than silently excluding them. If rendering fails, correct and re-freeze before dispatch; no correction after any responses unless the affected run is explicitly invalidated and reported.

## Verification and freeze

`python3 generate.py`; `python3 oracle.py`; `node render.cjs` (Playwright and installed Chrome). The independent oracle scans each minute for concurrent bookings rather than using pairwise interval intersection. DOM checks verify exact visible statement equality in all 32 renders. Manually inspect representative table and aligned PNGs, including the shortest overlap case, before freezing hashes. `freeze.json` records SHA-256 values for cases, rubric, assignments, protocol, instructions, render audit and all PNGs. Assigned PNGs must remain unchanged after dispatch.

This study makes no human-performance claim. A correct answer can be reached by reading the numeric text in either form. A perfect result would support recoverability under these model/task conditions, while failing to establish a benefit from alignment. The foray's wider End remains open.