# hospitalSIM verification · v0.4 · October29–30,2026

Working AI-assisted prototype. Production hosting is a deployment status, not a claim of commercial release quality. No studio interest, clinician/educator involvement, learner evaluation or learning effect is claimed.

## Executed software checks

The original Canvas checks pass. Diorama checks cover room capacity, reachability, blocked doors/routes, construction over people, occupied-room removal, explicit staffing, undo cutoff, finite budget, bounded fatigue, deterministic replay, save/load equivalence, end-of-shift termination and embedded GLB dependencies. Release regression checks classify every elapsed visit tick exactly once, verify interior route endpoints/platform clearance, and migrate actual v0.3 saves without lost IDs or counts.

The expanded sweep runs100 full shifts: five scenarios, ten seeds, baseline/plan pairs. Every tick checks accounting, unique reservations and navigation positions; full import validation runs periodically and at termination. Maximum observed blocked-navigation stall:5ticks. The tested consultation-plus-care-upgrade plan increases completions in43pairs and ties in7, with no losses in this limited sweep. This does not establish a universally dominant strategy or real-world capacity effect. Default seed431:27arrivals; baseline10completed/9departed/8unfinished; guided plan12completed/8departed/7unfinished.

The portfolio production build compiles and passes149tests in16files, including two new integration checks. The production dependency audit reports zero known vulnerabilities;16 findings remain in the existing development-tool tree (13high). No forced dependency upgrade was applied to unrelated tooling.

## Browser and asset checks

Wide1440×960 and narrow390×844 desktop viewports exercise construction, non-spatial placement, invalid keyboard placement, undo, room/patient lists, pause/speed/reset, reduced motion and the REPLISOME workstation. The workstation pauses hospital simulation and rendering; closing removes its iframe and leaves the hospital paused for deliberate resumption. Existing REPLISOME files are preserved byte-for-byte. Narrow controls remain in document flow and all scenario/policy controls remain available.

Original equipment and revised humans were actually generated/exported with Blender4.0.2. Editable .blend files and Python sources are retained. Each character has39weighted bones and9clips, one primitive, and a tiny embedded roughness palette. High-detail scrub/patient12,346triangles; coat13,498. LOD4,194/4,586. A separate1,744-triangle occupied-care drape was modeled/exported with Blender and shown only after transfer; it is static cloth with a brief visibility blend, not a cloth simulation. Equipment uses bevels and differentiated PBR materials; runtime platforms now have beveled edges. The browser has no CDN dependency. Character shader/rig export and round-trip evidence is in the source evidence directory.

## Visual assessment and limits

The supplied reference remains stronger in human anatomy, facial variation, cloth folds, hand/equipment contact and studio-light nuance. hospitalSIM has more circulation floor, repeated procedural figures, approximate interior task poses and grid-based crowd behavior. Faces, garment seams, surface roughness, doorway motion and heading changes have improved; they do not establish parity with the reference or a commercial studio game. The side-by-side presents actual browser capture beside the user-supplied reference, for assessment only.

Actual iPhone Safari, thermal/battery performance, long-session device behavior, VoiceOver user testing and a full accessibility audit remain pending. The in-app test browser keeps inspected pages visible, so an actual hidden-tab event is not verified here; background pause is implemented. A device-local iPhone check/export is provided. Clinical assumptions, usefulness, enjoyment, learning transfer and educational effectiveness require external review and consenting participant evidence. The assessment protocol remains an unvalidated draft.

The final release JSON, recording log and performance file supply the measured publication/evidence details. Earlier v0.3 evidence is retained as history, not silently reused as a v0.4 benchmark.

The revised characters passed exported-GLB round-trip checks: transfer/Rest endpoint difference below4.8e-7, authored walk stance drift3.06mm, maximum static mattress overlap8.25mm. These are source checks, not motion-capture validation or exact browser-contact guarantees.

Measured desktop wide gameplay:1,800frames,16.7ms mean/17.6ms p95, up to207,996 visible triangles in the inspected session. Representative steady draw calls80–83; static shadow initialization/checkpoint restoration peaked166, exceeding the provisional150 target briefly. Renderer totals include additional shadow-pass triangles. Initial selected asset files total8,258,953bytes raw and2,240,523bytes under per-file gzip9 (compression estimate, not measured network transfer). A390×844 paused-hospital panel sample measured901frames,16.7ms mean/17.1ms p95. Browser-reported GPU: ANGLE Apple M2 Metal. Real iPhone results remain pending.