The Iron Citadel

One dungeon level brief, The Iron Citadel Infiltration, built three ways in Unity 6 with the same Synty POLYGON dungeon art: through our level framework, and by two AI models working cold from one prompt in an empty project. The same scorer measured all three.

The web builds need a desktop browser. Click the page to capture the mouse. Then use WASD to walk, the mouse to look, Shift to run, E to open a door, M for the map and L to switch the headlight off or on. The headlight starts on, and it is only in the web builds: the browser's light limits would leave many rooms dark.

The builds

Ours: rooms

Eleven themed-room prefabs on the level's 45 m grid, each room with its own colour mood. Conformance 10 pass, 0 fail, 1 by eye*. Rebuilt and re-filmed on 13 October 2026.

Play in the browser

Claude Opus 5.5

One unattended run, 1:13:33. Conformance 10 pass, 0 fail, 1 by eye.

Play in the browser

Claude Fable 5.1

One unattended run, 1:28:48. Conformance 8 pass, 2 fail, 1 by eye: the east wing is blocked, and the gallery stairs snag.

Play in the browser

Ours: patterns

The same level assembled from pattern tiles on 100 m slots, now furnished with set pieces. It walks 8 of 8 routes, from the entry to the warlord and out, and all 8 of its pattern gates hold.

Play in the browser

Round two: brief v2

A harder brief with no data file or grid, built cold by three Claude models from one prompt and by our framework, then re-themed as Hollowmere Manor. One scorer, kit v2 1.1. Films only; the prompts, runs, scores and Unity sources are under brief-v2/ in the repository.

Claude Opus 5.5

Kit 11 pass, 0 fail. Cold, 1:23:49, $27.34 at list price.

Claude Fable 5.1

Kit 10 pass, 1 fail: the guards are seen from the entry hall. Cold, 1:39:24, $36.55.

Claude Haiku 5.5

Kit 9 pass, 2 fail after a reimport; as delivered no door opens. Cold, 1:03:47, $5.02.

Ours: PCGTools and the room library

Kit 10 pass, 1 fail: the flank door. About 25 s a seed, no model call.

Hollowmere Manor: Opus

The same brief in Horror Mansion art. Kit 11 pass, 0 fail.

Hollowmere Manor: ours

The same brief in Horror Mansion art. Kit 10 pass, 1 fail.

Side by side

Ours, roomsOpusFable
Conformance (pass / fail / by eye)10 / 0 / 1*10 / 0 / 18 / 2 / 1
Routes walked6 of 66 of 64 of 6
Cost at API list price—$29.57$45.41
Prefab instances placed in the scene564,9292,555
Unique Synty pieces used441291202
Renderers built from code084611

* Check 7 asks that the whole great-hall floor be visible from the gallery rail. Since 13 October 2026 the scorer counts that as at least 75% of the floor, down from the 90% we first chose. Ours sees 87%, so under the old target it would score 9 / 1 / 1. Opus (96%) and Fable (98%) pass under either target, and no other check changed.

Every check, the scorer's maps and the full measurements are in the comparison report.

The prompts, run summaries, score reports and Unity sources are in the repository.