# Foray 365 · Essai 01 — One project. Three pictures.

Executed 18 October 2026. Status: staged local construction and reader instrument. No human or AI reading-performance study has been executed. Native foray End and Ways remain unchanged.

## Question and setting

At a fictional project handover, an incoming manager is asked to authorise two fixed commissioning bookings from an outgoing lead’s record. Does a change of representational form preserve the information needed to decide, and which forms merit a later reader comparison?

Six invented cases share one formal structure: A Harbour sensors, B Museum climate, C Glasshouse pumps, D Archive scanners, E Hilltop relay, F Coastal beacon. They are structurally matched, not empirically equated for difficulty. No case is derived from a client, private Vault record, personal task, or confidential project. Foray 365 supplied the question and standards; the cases, rules and interface are original constructions for this essai.

## Model and answer oracle

Ten statements per case:

- F01: authorise both bookings if and only if all preparations are complete, consent is approved, and crew capacity is respected. These are the only gates in the fictional world.
- F02–F05: four preparation states.
- F06: consent approved, refused, or not recorded; not recorded means unknown, not refused.
- F07–F08: two bookings, each with start, finish and named crew.
- F09: capacity one per crew; distinct names denote distinct crews; touching intervals do not overlap.
- F10: no substitution, overtime or schedule change within this decision.

The evaluator returns **yes** if every gate holds, **no** if a gate is known to fail, and **unresolved** otherwise. It is an entailment judgement about this record, not an instruction to proceed under uncertainty. Where a collision and unknown consent coexist, the known collision settles no. The current six cases have a unique listed reason; the model’s general reason field selects a priority explanation, not every possible failed gate.

`verify.js` implements an independent oracle: complete the unknown consent as either approved or refused, then scan every minute of the day for over-capacity crews. All completions true means yes, all false means no, mixed means unresolved. It checks all 192 combinations of four binary preparations, three consent states, two crew assignments, and two interval relationships. Every evaluator result agrees. These finite counts describe the constructed state space, not prevalence in projects.

## Completed findings

| Case | Full-record answer | Decisive reason | Answer warranted after removing F06 | Possible full-record answers compatible with the remaining statements |
|---|---|---|---|---|
| A · Harbour sensors | Unresolved | Consent not recorded | Unresolved | Yes, no or unresolved |
| B · Museum climate | Yes | All gates met | Unresolved | Yes, no or unresolved |
| C · Glasshouse pumps | No | Same crew, overlapping bookings | No | No only |
| D · Archive scanners | Yes | Same crew, touching boundaries only | Unresolved | Yes, no or unresolved |
| E · Hilltop relay | No | Consent refused | Unresolved | Yes, no or unresolved |
| F · Coastal beacon | Unresolved | Consent not recorded | Unresolved | Yes, no or unresolved |

All six show the same **4/4 preparations complete** summary. All six also have the same network topology once its statement labels and attributes are stripped. Yet all three warranted decisions occur. Therefore neither compressed summary admits a decision function that is correct on all six records. This is a counterexample to those summaries’ sufficiency for this question. It is not a measured attention or comprehension failure.

Removing F06 prevents recovery of the original full-record answer in five cases: different possible full records agree on the remaining nine statements but have different answers. That does **not** make three responses equally warranted by the shortened record. The shortened record itself warrants **unresolved** in A/B/D/E/F and **no** in C, where the known capacity failure is decisive. In A and F, the original and shortened records both warrant unresolved. The possible full-record answers describe what the removed statement might have said, not alternative correct judgements of one incomplete record.

`evaluateWithoutConsent` masks the removed value as unknown before evaluating. Explicit checks confirm the shortened-record answer in every case, agreement with the independent oracle, and invariance to all three possible values of the removed F06 statement. Losing a fact need not change every answer, and an unanswered gate need not make a known failure uncertain.

The interactive counterfactual panel varies consent, crew assignment and interval relationship while retaining the progress summary. It deliberately changes facts to expose the failure of the summary. It is labelled separately from the form-only comparison.

## Presentation contract

The complete table, labelled network, layered board and progress-first control each render exactly the same ten canonical statement strings. `verify-browser.js` inspects rendered DOM fact IDs and text in all 24 complete case/presentation combinations, plus six nine-statement omission controls. Every statement has a visible box and is not display-hidden or transparent. The shared data file alone is not treated as proof of access.

| Presentation | Intended change | Extra encoding and limits |
|---|---|---|
| Table | Sequential baseline | Categories and stable fact identifiers |
| Network | Spatial grouping and labelled connections | Connections repeat the gate and booking/capacity relations; links do not denote chronology or completion |
| Layers | Decision / evidence / capacity grouping | Time bars redundantly encode the same stated intervals |
| Progress first | Large green completion area; other facts lower in display | All ten statements remain readable; a salience hypothesis, not an established misleading effect |
| Consent omitted | F06 removed | Deliberately unequal information; excluded from the form-only challenge |

Spatial groupings, redundant encodings and reading order differ. They may affect effort even though the propositions are recoverable. There is no equal-visibility or equal-comprehensibility claim. Layout also changes viewport and scrolling demands. On small screens, the network remains horizontally scrollable; a linear text route is explicitly available. Its use is recorded and must not be analysed as untouched network exposure. Non-colour labels and ordinary text preserve a keyboard/screen-reader route to every fact; this is an accessibility implementation check, not a full assistive-technology certification.

## Optional reader instrument

Each session sees six distinct cases, one presentation per case, twice each in table, network and layers. No same-case view switching and no answer feedback until all six are locked. Six permutations of view order cross six rotations of `[A,C,E,B,D,F]`, giving 36 possible arms. Across the complete allocation grid each case/position/view combination occurs twice. An individual or small random sample need not be balanced. There are repeated view sequences and structural practice effects; this design does not remove every carryover effect.

Recorded fields: case/version/hash, assignment, position, viewport at start, human/AI declaration, schedule familiarity, prior case exposure, public-brief exposure, decision, reason, collision answer, confidence in the decision, optional explanation, visible-tab time, wall time, and network-to-text route use. Timing begins on rendering and ends on submission; hidden-tab intervals are deducted from visible time. It includes scrolling and response entry, and is not pure reading time. Reload ends an attempt; only submitted answers survive. Prior exposure is retained even after an interrupted attempt or deletion of response sessions.

The UI reports correct decisions and wrong decisions at confidence ≥80%. Reason and collision answers remain separate, correlated checks. Two cases per view cannot rank the views. Confidence is confidence in the selected three-valued answer; “unresolved” may therefore be correctly chosen with high confidence.

This public instrument is not a blinded or controlled study: the introduction primes the mechanism, users self-select, answers/source are inspectable, exposure is self-reported as well as locally tracked, and there is no authenticated identity. Nothing is uploaded. JSON export requires an explicit click. Responses stay in this browser when storage works. Human and AI records remain distinguished; an AI test would need to disclose screenshot/text/source access and must not establish a human effect.

## Verification and reproducibility

Run from this folder:

```text
python3 generate_data.py
python3 build.py
node verify.js
node verify-browser.js
```

`verify-browser.js` requires Playwright and an installed Chrome; set `PLAYWRIGHT_MODULE` and `CHROME_PATH` if their defaults do not match your environment. Everything served to the visitor is embedded in `index.html`; it needs no remote scripts, fonts or runtime data fetches. Browser tests use an ephemeral isolated profile. Their intentionally synthetic response entries test scoring and are never exported as participant observations.

Outputs: `cases.json`, `results.json`, `browser-verification.json`, and inspection screenshots. The browser checks cover exact rendered statements, all three answer states, omission-world branches, responsive layouts at 375 px, the linear route, delayed answer reveal, scoring of a deliberate confident error, and prior-exposure persistence. Screenshots with “synthetic” in the name are QA evidence only.

## What would change the conclusion next?

For a prospective comparison, freeze cases and scoring, use fresh readers with a neutral brief, collect sufficient assignments to separate case and order from view, and retain route/device conditions. Test discovery, false inferences, confidence and effort before asserting preference or speed as reasoning benefit. Investigate the complete progress-first layout separately against a complete neutral control. A failure to improve correctness, an increase in confident errors, or equivalent outcomes with greater effort would count against that layout for this task. None of those empirical outcomes is established here.