# Aligned time. Same answers.

Foray 365 · Essai 02 · executed 19 October 2026.

## Observed result

Four fresh Codex model sessions completed eight screenshot-only judgements each. The booking-order table produced 16/16 correct decisions and the aligned crew/time view produced 16/16. All 32 responses reported 100% confidence. No confident errors occurred. No session was retried or replaced; no access deviation was reported. Every case received the correct answer in both forms.

This is a ceiling in a small formative pilot, with no observed accuracy advantage. It does not establish equivalence between forms, an effect on reasoning process, improved speed or effort, confidence calibration, or a human benefit. Four sessions of the same inherited configuration are not four independent model populations. The exact configured model identifier and reasoning setting were not exposed and are not invented here. [Run metadata](responses/run-metadata.json) records these limits.

## Cases and representation

Sixteen complete fictional commissioning records instantiate eight mechanisms with two variants each. Every record contains six bookings, three distinct named crews, complete preparations and approved consent. The sole variable gate is one-booking-at-a-time crew capacity. All times are same-day and explicit; exact endpoint handovers do not overlap. Half the records are feasible, half blocked. The bank includes different-crew parallel work, touching chains, five-minute gaps, nested bookings, five-minute overlaps, overlap chains and conflicts in two crews.

The table keeps shuffled booking-ID order. The aligned view groups bookings by crew and start time and draws durations against shared coordinates. Both contain the same four rule statements and six complete booking statements, word for word. Exact times remain readable in both, so success need not depend on the bars. Grouping, order and redundant bar geometry change together; this trial does not isolate those components. Thirty-two rendered-DOM audits verified statement identity and visible boxes. Representative final PNGs were visually inspected before freeze. Images are 1400 by 1100 pixels.

## Freeze and assignment

The [protocol](PROTOCOL.md), instructions, cases, rubric, assignments and 32 PNG stimuli were frozen at **2026-10-19 11:47:25.251291 UTC**, before any reader dispatch. The original files and assigned packet copies were marked read-only. `freeze.json` contains their SHA-256 hashes. Scoring reverified all 74 frozen source files; none changed.

R1/R3 received K01–K08; R2/R4 received K09–K16. R3/R4 reversed case order and inverted form. Each reader saw four cases in each form, with two Yes and two No cases in each form. Every case was seen once per form across different readers; each serial position was balanced across forms within paired sessions. No reader saw the same case twice. Related structures can still cause practice. The original assignment file retains historical local packet paths; `portable-assignments.json` maps the identical PNGs to relative archive paths.

## Actual reader procedure

Each reader was a fresh `fork_turns=none` session with the inherited configuration. Dispatch supplied only neutral instructions, reader ID and eight assigned PNG paths. Readers were instructed to inspect those images with `view_image`, in order, without source files, OCR, programmatic calculation, network, earlier conversation or other readers. They selected Yes, No or Unresolved first, then supplied confidence, what needed checking and brief visible evidence. No diagnostic reason menu or answer feedback was supplied. The business rule necessarily described the capacity constraint being applied.

R1/R2 ran first, then R3/R4 under unchanged instructions. The orchestrator saved their actual final JSON content; only whitespace formatting was normalized. Raw records are [R1](responses/R1.json), [R2](responses/R2.json), [R3](responses/R3.json) and [R4](responses/R4.json). No fabricated or automated test response is included. Tool/model latency and hidden computation were not controlled or interpreted as effort.

## Scoring and evidence audit

The frozen rubric scores the exact authorisation decision. All records are complete and determinate, so Unresolved would be incorrect. A wrong decision at confidence >=80 is a confident error. The predeclared secondary review asks whether free explanations cite true conflicts, omit another actual conflict or invent a conflict; readers were not required to list every clash.

The case generator uses pairwise interval intersection. An independently implemented oracle scans each minute and identifies simultaneous bookings by crew. Both agree on all cases, including the exact conflicting pairs and intervals. `score.py` repeats an occupancy check, verifies the frozen hashes and assignments, scores all actual responses and writes `observed-results.json`.

After the runs, an agent read all 32 explanations and manually encoded their concrete semantic claims in `audit_claims.py`: actual crew names, booking IDs, quoted intervals/endpoints, touching boundaries, gaps and overlap intervals. Those explicit claims are then checked programmatically against the frozen records. This is not a claim that a keyword script understands arbitrary explanations. `evidence-audit.json` preserves every full evidence field beside the interpretation, allowing another reviewer to contest the coding.

All 32 explanations are supported. Every rejected-plan response names all actual clashes, including both clashes where two occur. Every accepted-plan response accounts for all three crews through exact handovers, disjoint intervals, gaps or a single booking. No invented conflict or wrong time assertion was found. These correlated checks strengthen the evidence that the decisions were supported; they are not additional independent trials.

## Interpretation and next discriminating question

The observed model configuration recovered the correct judgement from either representation across this bank. The aligned representation visibly externalises a potentially relevant operation, but that intended mechanism did not yield a measurable accuracy difference here. The table already sufficed. A model could have read the numeric statements in both conditions, and correct evidence does not identify its internal reasoning process.

A future frozen experiment could separate booking-order text, crew-grouped text and crew-grouped bars, with larger schedules to create real search demand. That proposal is a new hypothesis; neither cases nor scoring were tuned after the ceiling was observed. A human comparison remains a separate enquiry.

The first essai’s [source-use record](../SOURCES.md) supplied the informational/computational-equivalence distinction. The cases, form contrast and observations in this pilot are original constructions. No source paper's empirical finding is transferred to this interface.