{ "model": "gpt-6-astra", "date": "2026-09-07", "status": "integrated", "expectedRuns": 16, "completedRuns": 16, "method": "One fresh Codex subagent per benchmark; fork_turns none; model gpt-6-astra selected explicitly. Inputs: the original benchmark specification path, a dedicated output path, and execution/validation instructions. No reference code or other model artifacts supplied.", "isolation": "Instruction-based separation in a shared filesystem. Agents were instructed not to inspect existing implementations or site source. This is not an access-control sandbox or independent file-access attestation.", "feedback": "In-run tool feedback and repairs were allowed. Benchmark 09 received coordinator review of the agent's own plan and construction order, as required by that specification; no reference code was provided.", "evaluation": "Tool-assisted implementation with in-run browser checks and repairs. Registry validation means an entry file was ingested; it is not a full benchmark pass or an aggregate score. See each validation report for actual checks, limitations, and substitutions.", "runs": [ { "benchmark": "low-poly-world", "agent": "/root/bench01", "prompt": "benchmarks/prompts/01-low-poly-world.md", "promptSha256": "9882b596bb047c95360574b0bd475457c32135d293126fca7d5cc5374ecca1e9", "artifactSha256": "dc18c2a153a6fc467061ed02992428f63b36c2cfae3adeee2c995f1132595726", "files": [ { "path": "index.html", "sha256": "dc18c2a153a6fc467061ed02992428f63b36c2cfae3adeee2c995f1132595726" } ], "entry": "/benchmarks/low-poly-world/gpt-6-astra/index.html", "manifest": "/benchmarks/low-poly-world/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/low-poly-world/gpt-6-astra/VALIDATION.md", "validation": "# Ember Isle — validation\n\nOutput: `index.html`, a standalone HTML document with inline CSS and JavaScript. No build step. Original implementation based only on `benchmarks/prompts/01-low-poly-world.md`; no benchmark implementations or site source were inspected.\n\n## Runtime external assets\n\n- https://cdn.jsdelivr.net/npm/three@0.160.0/build/three.module.js\n\nThis is the only runtime external request. No fonts, textures, models, images, or audio are fetched. The parent integrator should include this URL in `manifest.json.externalAssets`.\n\n## Checks performed\n\nTested using installed Google Chrome via Playwright and the local static server, at 1280 × 800.\n\n- Page opens directly into the WebGL world. No page errors or browser console errors in the verification run.\n- Initial scene: approximately 31,545 rendered triangles and 179 draw calls. The on-screen one-second FPS counter reported 60 on this host.\n- Actual keyboard movement: walked from the spawn to the nearest tree, pressed E, observed timber increase from 0 to 4 and live trees decrease from 40 to 39.\n- Actual keyboard movement: continued to the workbench, pressed C, clicked Trail stool, then pressed P. Timber decreased from 4 to 1, pending item cleared, placed-object count increased to 1.\n- Opening crafting away from the bench was rejected.\n- Separate instrumented browser response (test-only relocation helper, not written into the deliverable): harvested five trees using E, yielding 20 timber; used the actual crafting buttons and P to make and place all three items. Stool cost 3, lantern cost 5, planter cost 7; final timber 5, placed count 3, pending item null.\n- Dragging from the canvas center changes the compass from N / 000° to NE / 025°. The center reticle does not intercept pointer events.\n- Flight toggle and Space ascent work. Vertical flight clamps at 27 m. Returning to walking mode puts the camera exactly 1.7 m above the interpolated terrain.\n- Reduced-motion emulation returns `reducedMotion: true`; the implementation uses 10 particles instead of 70 and omits particle drift. There is no camera bob in either mode.\n- 390 × 844 compact layout: no horizontal document overflow. Desktop and compact screenshots captured and visually inspected during implementation.\n- Inspected the rendered desktop world: faceted terrain, layered cone trees, directional shadows, stone path, workbench, HUD, and sun are visible. Shader color output corrected following the first screenshot check.\n\n## Implementation details and limitations\n\n- Terrain collision uses barycentric interpolation over the same triangles rendered by the heightmap. Dry-ground movement and circular island bounds constrain walking; flight has matching horizontal bounds and floor clearance.\n- Water uses animated geometric waves and a changing normal-based directional sun highlight. All foliage, rocks, furniture, water, and sky are procedural geometry or shaders.\n- Crafting produces a carried item; P places it at the player's ground position. Recipes and placement instructions are shown in the crafting panel.\n- Not tested on a physical 2020 integrated-graphics laptop. The measured 60 FPS applies only to the available host and does not certify the target hardware.\n- Keyboard and pointer are required for world controls. Compact HUD rendering was checked, but touch-only movement controls are outside this benchmark implementation.\n- No saved-world persistence. Refresh resets the island.\n- Runtime requires WebGL and network access to the listed pinned Three.js CDN URL.\n", "externalAssets": [ { "kind": "script", "url": "https://cdn.jsdelivr.net/npm/three@0.160.0/build/three.module.js" } ] }, { "benchmark": "office-life", "agent": "/root/bench02", "prompt": "benchmarks/prompts/02-office-life.md", "promptSha256": "0997eac5495038f9d994e4d3a47c59f4e8fe4b36af1ffbeefaf6af2ce7e92b1f", "artifactSha256": "e170f4b2c1baf85cb9fab4a96fe63921513c4ac58d139637807bc892fcb8cec1", "files": [ { "path": "index.html", "sha256": "e170f4b2c1baf85cb9fab4a96fe63921513c4ac58d139637807bc892fcb8cec1" } ], "entry": "/benchmarks/office-life/gpt-6-astra/index.html", "manifest": "/benchmarks/office-life/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/office-life/gpt-6-astra/VALIDATION.md", "validation": "# Validation — The 9 to 5 / GPT-6 Astra\n\nImplemented independently from `benchmarks/prompts/02-office-life.md`. No reference implementation, other benchmark output, catalog, registry, or site source was inspected. The deliverable is one self-contained `index.html` with inline SVG artwork, CSS, and JavaScript.\n\n## Actual checks\n\nUsed Playwright with installed Google Chrome (`chromium.launch({ headless: true, channel: 'chrome' })`) against the parent's static server at `http://127.0.0.1:4346/benchmarks/office-life/gpt-6-astra/`.\n\n- Normal workday reached 5:00 PM after approximately 15 active real seconds. All 22 tasks completed, all five queues emptied, projector and room setup prerequisites completed, and the random incident fired exactly once. HUD measured 60 FPS in this browser session.\n- Observed the completed-day state and verified automatic restart after the eight-second intermission.\n- Space paused simulated time; a second Space resumed it. R reset the day. Separately verified Space works when an event button has keyboard focus.\n- Triggered Emergency Meeting during active work. Verified all five characters reached discussion state, simulated time and task queues stayed frozen, the discussion remained active after one second, and completion counts and remaining queues were preserved when coworkers returned.\n- Triggered Lunch Break during morning work. Sampled resources 50 times at 100 ms intervals: maximum kitchen occupancy was three, coffee queues contained no duplicate IDs, and no recorded capacity violations occurred.\n- Ran a complete day with lunch triggered around 11:16 AM. It ended with 21 completed tasks and Jordan actively finishing the final paperwork task, with no blocked coworker or stale coffee/printer queue. This exercised coffee reacquisition after lunch consumes a previously delivered cup.\n- Tested `prefers-reduced-motion: reduce`: characters teleported to their initial destinations and computed animation names were `none`.\n- Captured and visually inspected desktop and 390 px mobile screenshots. Verified the 390 px page has no horizontal document overflow. HUD and event controls use sticky positioning.\n- Browser instrumentation recorded no uncaught page errors and no external network requests. Only the local HTML navigation was requested.\n\nTemporary browser scripts and screenshots were kept in `/tmp`, outside the deliverable directory.\n\n## Runtime external assets\n\nNone. No CDN URLs, network fonts, external images, JavaScript libraries, or runtime fetches. Furniture and all five characters are inline SVG. System Arial and Georgia fonts are used. The parent agent supplies the integration manifest; its CDN/external asset list should be empty.\n\n## Limitations\n\n- Performance was measured in headless Chrome on this machine, not across all browsers or device classes.\n- The complete fixed floor plan scales down on phones; its small internal room labels are most legible at desktop sizes. Character names and current tasks remain available in the separate roster.\n- Navigation uses room-door routes rather than full furniture collision avoidance. Coworkers can visually overlap when crossing a narrow doorway.\n- Emergency meetings freeze office time while gathering, discussing, and returning, so an interrupted day takes longer than 15 wall-clock seconds. Pausing likewise stops the clock.\n- Manual lunch interruptions can leave unfinished work at the fixed 5:00 PM cutoff; the stats report tasks actually completed. A fresh day resets unfinished queues.\n- The finite role-specific queues are followed by idle desk behavior. This is a stylized office simulation, not an open-ended planner or conversational AI.\n", "externalAssets": [] }, { "benchmark": "vice-city", "agent": "/root/bench03", "prompt": "benchmarks/prompts/03-vice-city.md", "promptSha256": "0029eeeb677b800b604656722ff53ec3b4f9ffb9531b4c3601ec2f7aa7f53ad7", "artifactSha256": "2cfa0831ef420d2b366c77d47ee091bd02744090879b9db07b68f0537d59277f", "files": [ { "path": "index.html", "sha256": "2cfa0831ef420d2b366c77d47ee091bd02744090879b9db07b68f0537d59277f" } ], "entry": "/benchmarks/vice-city/gpt-6-astra/index.html", "manifest": "/benchmarks/vice-city/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/vice-city/gpt-6-astra/VALIDATION.md", "validation": "# Vice City — GPT-6 Astra validation\n\nImplemented independently from `benchmarks/prompts/03-vice-city.md`. No existing benchmark implementations, reference implementation, other model output, registry, or site source were inspected. Output is a standalone HTML document with inline styles, inline JavaScript, inline SVG control icons, and procedural Canvas 2D rendering of fixed-angle 3D geometry.\n\n## Actual checks\n\n- Tested in installed Chrome through Playwright at 1440×1000, 1440×900, and 390×844, using the parent static server.\n- No uncaught page errors during initial load, regeneration, day/night, traffic, earthquake, zoom, pan, responsive resizing, or reduced-motion changes.\n- Regeneration changed seed, generation number, building count, and road coordinates. Initial generation measured 9.1 ms; subsequent regeneration measured 1.7 ms in the first run. These timings include data generation and procedural drawing of the completed city layers, but exclude the intentionally staggered presentation.\n- Traffic was enabled initially. Four successive one-second HUD samples displayed 60 FPS. Car positions changed during a 300 ms observation.\n- Traffic toggle set the traffic state to false. Re-enabling with reduced motion active left every car position unchanged across 600 ms.\n- Day and night screenshots visually inspected: warm facades, directional shadows and trees by day; dark blue sky, moon, individual amber windows and street lamps at night.\n- Earthquake badge was visible immediately after activation and absent after 3.2 seconds.\n- Zoom-in changed zoom to 1.15; a pointer drag changed pan to x=40, y=20.\n- At 390×844, all seven buttons fit within the viewport and document horizontal overflow was false. Final desktop and mobile screenshots visually inspected.\n- Reduced-motion media changes were detected live. Code disables building stagger, building sway, camera shake, and vehicle motion in that mode.\n- Captured runtime requests contained only the HTML page itself.\n\n## Runtime external assets\n\nNone. No libraries, fonts, textures, images, or remote data are fetched. The UI uses local system fonts; all city materials and geometry are drawn procedurally.\n\n## Limitations\n\n- No physical 2020 laptop was available. The recorded generation time and frame rate are measurements from the available Chrome environment, not a guarantee for every device.\n- Uses a Canvas 2D projection of extruded 3D blocks rather than WebGL. Buildings and vegetation use painter ordering, with moving cars composited beneath the building layer.\n- Traffic follows straight grid lanes, pauses at intersections, and wraps at city boundaries; it does not model destination planning or collision avoidance.\n- Mobile moves the sidebar beneath the canvas to preserve usable controls. The FPS counter remains at the top right.\n- No cross-browser, screen-reader, or formal contrast audit was performed.\n", "externalAssets": [] }, { "benchmark": "jabberwock", "agent": "/root/bench04", "prompt": "benchmarks/prompts/04-jabberwock.md", "promptSha256": "5a0f218ba47224f7d9b4b830f714ae4dec476147e75c2167ca00d4504ef50455", "artifactSha256": "88380adcf46ac075ff4fdfcbee52207686c0f415997fa39565ada8fb5d81ba35", "files": [ { "path": "index.html", "sha256": "88380adcf46ac075ff4fdfcbee52207686c0f415997fa39565ada8fb5d81ba35" } ], "entry": "/benchmarks/jabberwock/gpt-6-astra/index.html", "manifest": "/benchmarks/jabberwock/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/jabberwock/gpt-6-astra/VALIDATION.md", "validation": "# Jabberwock validation\n\nOriginal implementation of benchmark prompt 04. No reference implementations, other benchmark outputs, or site source were inspected. Single `index.html`; no build step, local supplementary assets, runtime textures, audio files, or image requests.\n\n## Actual checks\n\nTested with installed Chrome through Playwright against `http://127.0.0.1:4346/benchmarks/jabberwock/gpt-6-astra/`.\n\n- Desktop viewport 1440 × 1000: loaded WebGL scene and custom fonts; no JavaScript page errors. FPS counter reported 60 FPS during the observed starting scene. Counter samples real elapsed time once per second.\n- Captured and visually inspected the introductory world, encounter scene, and final poem screen.\n- Completed the story using real keyboard events: clicked Enter the wabe, held W to reach the resting circle, pressed E, waited for the pre-encounter verse, approached the Jabberwock, used Space attacks and Shift dodges, reduced boss health to 0%, then held S to carry the head home. Verified the final ending overlay was visible and chapter read VII — O FRABJOUS DAY.\n- Separate failure run: entered the fight, took damage without attacking until the death overlay appeared, clicked the retry button, verified encounter chapter, all six courage pips, and restored 100% boss health.\n- Mobile viewport 390 × 844 with touch emulation and `prefers-reduced-motion: reduce`: loaded without JavaScript page errors; visually checked intro and active game. Corrected overlapping touch controls and rechecked the resulting screenshot.\n- Verified pause and resume buttons show and dismiss the pause overlay in the mobile/reduced-motion run. Verified touch controls become visible after starting.\n- Reduced-motion implementation skips ambient particle points and burble particles, removes camera shake, and stabilizes eye-flame flicker while preserving controllable movement.\n- Network requests were observed: Three.js plus Google Fonts stylesheet/font files below. No external textures were requested.\n\n## Creative implementation\n\nBrillig is a honey-colored late afternoon. The wabe is a pale meadow with a shallow pool and circling corkscrew-tailed toves. Borogoves are delicate, long-legged birds; mome raths pulse and bob nearby. The tulgey wood becomes cooler and darker as the player enters. Its hollow, amber-hearted Tumtum tree marks the contemplative pause. The Jabberwock is a tall antlered winged creature with flaming eyes, articulated claws, teeth, a trailing tail, and visible burble bubbles. A circling Jubjub bird and prowling, damaging Bandersnatch inhabit the wood. The vorpal sword produces a broad luminous sweep and punctuating snicker/snack text; the killing strike detaches a carried head. The player galumphs faster back to the elder and concluding verse.\n\n## Limitations\n\n- Performance was observed in the available Chrome environment, not measured on a physical 2020 laptop. No claim of target-hardware benchmarking is made.\n- WebGL and an internet connection for the CDN module and fonts are required. There is an error message if the world fails to load; no offline library bundle is included.\n- Mobile layout, reduced-motion loading, and pause/resume were checked; a complete touch-only playthrough was not performed.\n- Combat is intentionally forgiving, with proximity-based sword hits. Trees are scenic rather than physical collision obstacles.\n- Sound is omitted; burbling/outgrabe are interpreted visually.\n\n## Runtime external assets\n\nExplicit sources:\n\n1. `https://cdn.jsdelivr.net/npm/three@0.170.0/build/three.module.js`\n2. `https://fonts.googleapis.com/css2?family=Cormorant+Garamond:ital,wght@0,400;0,500;0,600;1,400;1,500&family=DM+Sans:wght@400;500;600&display=swap`\n\nObserved font resources supplied by that stylesheet:\n\n3. `https://fonts.gstatic.com/s/cormorantgaramond/v21/co3bmX5slCNuHLi8bLeY9MK7whWMhyjYqXtKky2F7g.woff2`\n4. `https://fonts.gstatic.com/s/dmsans/v17/rP2Yp2ywxg089UriI5-g4vlH9VoD8Cmcqbu0-K6z9mXg.woff2`\n5. `https://fonts.gstatic.com/s/cormorantgaramond/v21/co3ZmX5slCNuHLi8bLeY9MK7whWMhyjYrEtImSqn7B6D.woff2`\n\nGoogle Fonts may select additional subsets or revisions depending on browser and served stylesheet; the only font source families are Cormorant Garamond and DM Sans. Include these assets in the integrating manifest's `externalAssets`.\n", "externalAssets": [ { "kind": "stylesheet", "url": "https://fonts.googleapis.com/css2?family=Cormorant+Garamond:ital,wght@0,400;0,500;0,600;1,400;1,500&family=DM+Sans:wght@400;500;600&display=swap" }, { "kind": "script", "url": "https://cdn.jsdelivr.net/npm/three@0.170.0/build/three.module.js" }, { "kind": "font", "url": "https://fonts.gstatic.com/s/cormorantgaramond/v21/co3bmX5slCNuHLi8bLeY9MK7whWMhyjYqXtKky2F7g.woff2" }, { "kind": "font", "url": "https://fonts.gstatic.com/s/dmsans/v17/rP2Yp2ywxg089UriI5-g4vlH9VoD8Cmcqbu0-K6z9mXg.woff2" }, { "kind": "font", "url": "https://fonts.gstatic.com/s/cormorantgaramond/v21/co3ZmX5slCNuHLi8bLeY9MK7whWMhyjYrEtImSqn7B6D.woff2" } ] }, { "benchmark": "petri-dish", "agent": "/root/bench05", "prompt": "benchmarks/prompts/05-petri-dish.md", "promptSha256": "01d9ea990902695b59469264e5ef84a8ca229d982ebd0a7a83600e83141527dd", "artifactSha256": "a8a3ec6015048f65bcfc25dba20977a52a2805b9485bc76b64ce0188ee8208cf", "files": [ { "path": "index.html", "sha256": "a8a3ec6015048f65bcfc25dba20977a52a2805b9485bc76b64ce0188ee8208cf" } ], "entry": "/benchmarks/petri-dish/gpt-6-astra/index.html", "manifest": "/benchmarks/petri-dish/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/petri-dish/gpt-6-astra/VALIDATION.md", "validation": "# Validation\n\nOriginal implementation written from `benchmarks/prompts/05-petri-dish.md`. No reference implementation, existing benchmark, or site source inspected.\n\n## Browser checks performed\n\nTested using Playwright with installed Chrome in headless mode, served at `http://127.0.0.1:4346/benchmarks/petri-dish/gpt-6-astra/`.\n\n- Reset while paused: exactly 30 prey and 5 predators.\n- Pause: serialized agent state, simulation time, and history remained identical across a 500ms wait.\n- Add controls: one click each changed the paused population to 40 prey and 8 predators.\n- Speed: 4× advanced simulation by 4.00s in a 1s wait; 0.25× advanced it by 0.249975s.\n- Reduced motion: browser media emulation selected 0.50×, disabled the slider to preserve that preference, and advanced 0.9999 simulated seconds during a two-second wait. Code gates chart rendering to every 500ms in this mode.\n- History: at 80 simulated seconds, the oldest retained point was 19.8667s, retaining 60.1333s. Chart uses a fixed rolling 60-second viewport, connected interpolated samples, and a per-frame live endpoint.\n- Containment: checked every agent after each of 6,000 simulation steps (100 simulated seconds); zero agents reached/exceeded normalized radius 1; maximum observed center radius 0.986277.\n- Controlled energy checks verified prey reproduction, predator reproduction, and predator starvation.\n- Performance: with 175 prey and 25 predators initially, 120 simulation-and-canvas frames averaged 0.196ms CPU time, p95 0.300ms. A separate live animation run initialized with 200 agents reported 60 FPS after three seconds (199 agents remained).\n- No browser JavaScript errors recorded.\n- Viewed screenshots at 1440×1000 desktop, 1280×800 laptop, and 390×844 mobile. Mobile had no horizontal document overflow. Controls stack below the dish on mobile.\n\n## Runtime external assets\n\nNone. No libraries, remote fonts, images, requests, or build step. All simulation and chart graphics use Canvas 2D, with dish texture generated once per resize on an offscreen canvas. Typeface uses system Arial/Georgia/monospace fallbacks.\n\n## Limitations\n\n- Performance results describe this machine and headless Chrome, not all hardware or browsers.\n- Population dynamics are stochastic; extinction is possible and intentional. Reproduction is capped at 650 prey and 120 predators to bound ordinary simulation growth; explicit add buttons remain available.\n- Prey also expend energy and may starve. Food is capped at 450 particles.\n- Mobile moves the control sidebar below the simulation to keep both readable. The header remains sticky and the mobile population HUD is fixed to keep both visible.\n- No manual Safari/Firefox or screen-reader testing performed.\n", "externalAssets": [] }, { "benchmark": "helms-deep", "agent": "/root/bench06", "prompt": "benchmarks/prompts/06-helms-deep.md", "promptSha256": "0a7c8c01558998c09342ecd7bea55fa12b227f84a5fd3da87c17e93f53745e8c", "artifactSha256": "0be188bfdd2e887ea0fbe11811e3e73343f95d1e0fbdd3ccae4fe6b5cf5c6f51", "files": [ { "path": "index.html", "sha256": "0be188bfdd2e887ea0fbe11811e3e73343f95d1e0fbdd3ccae4fe6b5cf5c6f51" } ], "entry": "/benchmarks/helms-deep/gpt-6-astra/index.html", "manifest": "/benchmarks/helms-deep/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/helms-deep/gpt-6-astra/VALIDATION.md", "validation": "# Validation — Helm’s Deep / GPT-6 Astra\n\nImplemented independently from `benchmarks/prompts/06-helms-deep.md`. No existing implementations, references, other model output, registry, or shared site source were inspected.\n\n## Actual checks\n\n- Served the single HTML through the parent static server at `http://127.0.0.1:4346/benchmarks/helms-deep/gpt-6-astra/`.\n- Loaded and rendered using Playwright and headless Google Chrome, 1440 × 900. Verified a live WebGL context and 756 scene children. No page JavaScript errors were captured.\n- Visually inspected the landing screenshot and an interior first-person screenshot; refined the procedural cliff geometry after inspection and inspected the updated landing screenshot.\n- Clicked Enter and verified pointer lock activates. Held W for 1.5 seconds and observed the camera advance from z=82 to approximately z=74.5 while following the causeway elevation.\n- Started near the gate at (-39, 9.1, 27), held W, and verified passage to courtyard z≈16 with eye height 10.1 (courtyard surface 8.35 + eye 1.75).\n- Pressed Space in the courtyard; observed eye height rise to approximately 10.84 and return to 10.1 after landing.\n- Tested the collision resolver with a position within a main curtain wall; it moved the player out to z=22.92.\n- Verified ground evaluator values: causeway center 4.175, courtyard 8.35, top of interior ramp 13.3.\n- Verified Escape exits pointer lock and restores the pause overlay. An initial synthetic Escape issue was corrected and rechecked.\n- Resized to 390 × 844; canvas resized to 390 pixels and document scroll width remained 390. Captured a screenshot at that size.\n- Recorded network requests: the only HTTPS runtime request was the Three.js module below.\n\n## Runtime external assets\n\n- `https://cdn.jsdelivr.net/npm/three@0.169.0/build/three.module.js` — Three.js, loaded through an inline import map.\n- No external fonts, images, textures, models, audio, CSS, or API calls. Stone/mud maps are generated with inline canvas code; all meshes and rain are procedural.\n\n## Limitations\n\n- This is an original stylized interpretation of the movie geography, not a surveyed or exact film-set reconstruction.\n- Desktop mouse/keyboard and browser pointer-lock support are required for exploration; there are no touch movement controls.\n- Collision uses oriented boxes, simplified tower bounds, and analytical walkable surfaces rather than mesh-level physics. Decorative props and distant rocks do not all have individual collision hulls.\n- The great hall is exterior geometry, and the Glittering Caves are represented by their entrance; neither has an interior environment.\n- The full extent of every parapet segment was not traversed during automated validation. Ground heights and representative collision/gate/jump behavior were checked.\n- CDN availability and WebGL are required. Performance was not profiled on low-end devices or browsers other than Chrome.\n", "externalAssets": [ { "kind": "script", "url": "https://cdn.jsdelivr.net/npm/three@0.169.0/build/three.module.js" } ] }, { "benchmark": "hogwarts-broom-flight-simulator", "agent": "/root/bench07", "prompt": "benchmarks/prompts/07-hogwarts-broom-flight-simulator.md", "promptSha256": "d8a7287c80fd5cf9f1f5b85e414076cfd839e5e38f2defb64f46d1e9a423562a", "artifactSha256": "f608c4878d421d230dcaaa6f32ee7d03596274f0ab5b0781fc3fc9713f0bbc95", "files": [ { "path": "index.html", "sha256": "f608c4878d421d230dcaaa6f32ee7d03596274f0ab5b0781fc3fc9713f0bbc95" } ], "entry": "/benchmarks/hogwarts-broom-flight-simulator/gpt-6-astra/index.html", "manifest": "/benchmarks/hogwarts-broom-flight-simulator/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/hogwarts-broom-flight-simulator/gpt-6-astra/VALIDATION.md", "validation": "# Benchmark 07 — Hogwarts broom flight simulator\n\n## Deliverable\n\n`index.html` is a self-contained original implementation. It embeds Three.js 0.160.1, including its upstream MIT license header, and generates all geometry, materials, lighting, water, terrain, trees, rider, and interface procedurally. No build step is needed.\n\n## Actual checks\n\n- Opened the served page in headless Chrome using Playwright at 1280 × 800. No JavaScript page errors occurred.\n- Observed automatic movement immediately after loading. The HUD reported approximately 58–60 FPS on the available host, including 60 FPS after final clearance changes.\n- Captured and visually inspected desktop screenshots of arrival, castle courtyard, stadium hoop passage, lake, forest corridor, and the final wide reveal.\n- Space paused progress exactly across a timed sample, and resumed movement. M changed camera mode; F changed speed state. Right arrow reached the next chapter and left arrow returned to the previous chapter.\n- Confirmed hint opacity reaches zero after its six-second delay and fade.\n- Checked the closed curve endpoints: positional separation was zero, and tangent dot product was 0.9999949439.\n- Cast 12,000 short rays along the route centerline and corrected detected stadium support and entrance intersections. The resulting centerline test returned no intersections.\n- On the final path, cast 18,000 additional short rays across six offset trajectories: ±1.5 vertical units, ±1 horizontal unit, and ±2 depth units. These sampled rider/broom clearances returned no intersections with scene mesh geometry after correcting the bridge descent and stadium approach/departure.\n- Emulated `prefers-reduced-motion: reduce`, reloaded, and confirmed automatic progress remained functional. Reduced mode scales down banking and flight speed, lowers spark count from 160 to 40, and disables water animation.\n- Checked a 390 × 844 viewport and visually inspected the screenshot. Canvas matched the viewport and controls remained on screen.\n- Recorded browser requests: no runtime external requests were made.\n\n## Runtime external assets\n\nNone. `manifest.json.externalAssets` can be `[]`.\n\nEmbedded library provenance: Three.js 0.160.1, downloaded during authoring from https://cdn.jsdelivr.net/npm/three@0.160.1/build/three.module.js. This address is not fetched at runtime. All visual assets are procedural.\n\n## Limitations\n\n- Performance was measured on the available test host, not a physical 2020 laptop.\n- Clearance checks are sampled rays, not an exhaustive swept-volume proof. The camera uses lagged smoothing and has no separate collision solver.\n- The lake uses animated geometry and stylized shimmering highlights; it does not implement planar castle reflections.\n- The forest uses low-polygon evergreen silhouettes, branch forms, distance fog, shadows, and glowing eyes; there is no volumetric moonbeam renderer.\n- Route controls move along the continuous flight curve at an accelerated rate to the selected chapter rather than cutting instantly.\n", "externalAssets": [] }, { "benchmark": "universe-simulator", "agent": "/root/bench08", "prompt": "benchmarks/prompts/08-universe-simulator.md", "promptSha256": "fa184d620229e1d068216d5430e940ad77153be2d83a078b166e0dec93459099", "artifactSha256": "f4969fa888f39a63315d2a1d93c4df37affc10d030b63db92d3b560b8e41616c", "files": [ { "path": "index.html", "sha256": "f4969fa888f39a63315d2a1d93c4df37affc10d030b63db92d3b560b8e41616c" }, { "path": "three.module.js", "sha256": "0a3368c165eea773490aec7b77c22de70e3eac288503409256fdbf4d12578416" } ], "entry": "/benchmarks/universe-simulator/gpt-6-astra/index.html", "manifest": "/benchmarks/universe-simulator/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/universe-simulator/gpt-6-astra/VALIDATION.md", "validation": "# Validation\n\nImplemented independently from `benchmarks/prompts/08-universe-simulator.md`. No reference implementation or other benchmark output was inspected.\n\n## Actual checks\n\n- Loaded the page through the parent HTTP server in headless Chrome using Playwright.\n- Confirmed Three.js/WebGL rendering initialized and the loading overlay disappeared.\n- Captured and visually inspected screenshots at 1440 × 1000 and 390 × 844. Adjusted mobile profile placement to prevent title overlap, then repeated checks and screenshots.\n- No JavaScript page errors were reported.\n- Verified pause freezes the displayed simulation date.\n- Selected Jupiter and verified its profile updates.\n- Activated planet focus and verified the following state.\n- Exercised orbit, label, and star toggles; verified hidden labels.\n- Set the time slider to its maximum and verified 10,000 days / sec.\n- Reset simulation time and verified September 7, 2026.\n- Exercised top view, zoom in, zoom out, and camera reset buttons.\n- Opened the field guide and dismissed it with Escape.\n- Confirmed no horizontal document overflow at the mobile viewport.\n\n## Limitations\n\n- This is an educational solar-system visualization, not a cosmological model or accurate ephemeris. Initial planetary phases are illustrative.\n- Orbital periods and eccentricities inform Kepler-equation motion; distances and planet sizes are deliberately compressed. Orbits share one plane; perturbations and gravitational interactions are not modeled.\n- Procedurally generated planet surfaces are stylized, not photographic maps. No moons are simulated.\n- Requires WebGL. Browser checks were limited to Chrome; other browsers and real touch hardware were not tested.\n- Planet picking and drag/wheel handlers are implemented, but automated interaction checks focused on the HUD controls.\n\n## Runtime external assets\n\n- Google Fonts stylesheet: `https://fonts.googleapis.com/css2?family=DM+Sans:wght@400;450;500;600;700&family=Space+Grotesk:wght@400;500;600;700&display=swap`\n- Font binaries served by `https://fonts.gstatic.com/` through that stylesheet. System sans-serif fallbacks are provided.\n- Three.js 0.169.0 is vendored locally as `three.module.js`, obtained from `https://cdn.jsdelivr.net/npm/three@0.169.0/build/three.module.js`; it does not require runtime CDN access. Its MIT license header is retained.\n- All planets, star fields, asteroid positions, and glow assets are generated locally in the browser. No remote image assets.\n", "externalAssets": [ { "kind": "stylesheet", "url": "https://fonts.googleapis.com/css2?family=DM+Sans:wght@400;450;500;600;700&family=Space+Grotesk:wght@400;500;600;700&display=swap" } ] }, { "benchmark": "yingzao-fashi-assembly", "agent": "/root/bench09", "prompt": "benchmarks/prompts/09-yingzao-fashi-assembly.md", "promptSha256": "5ffad45769267f358cd1f21333d8b426af224cb53ee11897b92a92ce993db29e", "artifactSha256": "9a0b5fa02f4fa6d456bafa7e96514df08d90a1eaaca5323cc5d04613453c1a5e", "files": [ { "path": "index.html", "sha256": "9a0b5fa02f4fa6d456bafa7e96514df08d90a1eaaca5323cc5d04613453c1a5e" } ], "entry": "/benchmarks/yingzao-fashi-assembly/gpt-6-astra/index.html", "manifest": "/benchmarks/yingzao-fashi-assembly/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/yingzao-fashi-assembly/gpt-6-astra/VALIDATION.md", "validation": "# Benchmark 09 validation\n\n## Delivery\n\n- Original, standalone `index.html`; embedded JavaScript/CSS, procedural Three.js geometry and canvas-generated wood texture. No build step, imported model, photographs, or local media assets.\n- Model: 南禅寺大殿 / Nanchan Temple Main Hall, Wutai, Shanxi; Tang dynasty, 782 CE.\n- Plan was sent to the parent agent and explicitly reviewed/approved before the HTML was written. Validated order: columns → fang → beams/purlins → dougong → rafters/roof; reverse undoes this order.\n\n## Actual browser checks\n\nTested September 7, 2026 with Playwright and headless installed Google Chrome, opening the HTML directly through `file://`. No local server was required. Three.js was fetched over the network.\n\n- Canvas and building name, Chinese name, location, and period rendered successfully.\n- Clicked Play from the completed model. Captured stage state at successive intervals: only columns; columns + fang; columns + fang + beams/purlins; these plus dougong; all five groups including rafters/roof. Playback reached exactly 5 and stopped automatically.\n- Clicked Reverse, then Play. Playback reached exactly 0 and stopped; all five construction groups were hidden, leaving the foundation.\n- Clicked a stage selector to restore the finished model. All six floating labels were present, with Chinese, accented pinyin, and function sentences.\n- Moved age from 0 to 100 and captured screenshots at both endpoints. Personally reviewed clear transition from red/brown painted wood to very dark, desaturated weathered wood. Material roughness changed from approximately 0.38 to 0.98. Source review confirms generated pigment patches/fading, grain and erosion marks, interpolated live updates, and subtle member rotation/settlement.\n- Clicked Structural X-ray. Roof and wall material opacity became 0.085; 52 amber connection indicators became visible. Screenshot confirmed the timber skeleton was exposed.\n- Moved the browser pointer over projected column and dougong geometry; actual raycast tooltip text matched `柱 · Zhù` and `斗拱 · Dǒugǒng` respectively.\n- Tested 1440 × 1000 desktop and 390 × 844 mobile viewports. Mobile document scroll width was 390 (no horizontal overflow). Resizing updated the rendering camera and canvas.\n- Keyboard-focused the labels button and pressed Enter: labels hid and `aria-pressed` changed to false; Enter restored them. Step advanced from foundation to stage 1. Play/Pause stopped at a stable intermediate time.\n- Reviewed HTML source and event handlers: no TODOs, placeholder comments, missing handlers, or unimplemented functions found.\n- Browser recorded zero uncaught JavaScript errors throughout the five required checks, including the final complete rerun after roof geometry fixes.\n- Screenshot inspection identified protruding upper rafters/purlins; those members were shortened/lowered and a fresh complete test confirmed the repaired closed hip-and-gable roof. Mobile camera framing was expanded to include the full model.\n\nBrowser test scripts and screenshots were kept in `/tmp/bench09-*`, outside the deliverable. Parent agent independently reviewed desktop/mobile screenshots and found labels readable, the order clear, and controls reachable.\n\n## Architectural basis\n\n- [ChinaCulture: Main Hall of Nanchan Temple](https://en.chinaculture.org/library/2008-02/15/content_34559.htm): date 782 and three bays in both directions.\n- [University of Hong Kong: Nanchan temple](https://learning.hku.hk/ccch9051/group-13/items/show/41): approximately 11.75 m frontage and 10 m depth, twelve perimeter columns, no interior columns, wider center bay, hip-and-gable roof and bracket support.\n- The model follows these broad characteristics. The UI explicitly explains that this Tang building predates the 1103 Song treatise and uses its component vocabulary as an educational framework.\n\n## Runtime external assets\n\nExactly one required external runtime resource:\n\n- `https://cdn.jsdelivr.net/npm/three@0.160.1/build/three.min.js` — Three.js 0.160.1.\n\nThe ChinaCulture reference is a user-activated hyperlink, not a fetched runtime asset. Fonts use installed system/serif fonts; no font requests, image requests, imported controls, model downloads, or other dependencies.\n\n## Limitations\n\n- Educational reconstruction, not a measured conservation survey or structural engineering simulation. Exact joinery profiles, five-puzuo detail, camelhump braces, statues, inscriptions, historical pigment placement, and all later restoration differences are not reproduced. Columns are consistently round in this simplified model.\n- Bracket sets are simplified stacked bearing blocks and orthogonal arms. Joinery is represented by markers rather than cutaway mortise geometry. Major groups move as construction stages, not an individual craftsman's complete assembly procedure.\n- Erosion is an artistic continuous material/transform effect; it is not a dating or physical decay model.\n- Requires WebGL and an internet connection for the permitted Three.js CDN resource. CDN failure displays an in-app reload message. Fully offline loading was not claimed or tested.\n- Hover inspection is designed for a pointer; persistent floating labels keep equivalent component information available on touch devices. The narrow-screen composition places settings below the viewer.\n- Automated browser checks used Chrome only; other browser engines were not tested.\n", "externalAssets": [ { "kind": "script", "url": "https://cdn.jsdelivr.net/npm/three@0.160.1/build/three.min.js" } ] }, { "benchmark": "mechanical-watch-simulator", "agent": "/root/bench10", "prompt": "benchmarks/prompts/10-mechanical-watch-simulator.md", "promptSha256": "07c00af9ad23d44166b32b09ca0298d1c127b023b3b7f389274e5297050d112e", "artifactSha256": "895058786fd68a7c45c4af402dd74acc8713177aabd2292cc862cae70ba70be1", "files": [ { "path": "index.html", "sha256": "895058786fd68a7c45c4af402dd74acc8713177aabd2292cc862cae70ba70be1" } ], "entry": "/benchmarks/mechanical-watch-simulator/gpt-6-astra/index.html", "manifest": "/benchmarks/mechanical-watch-simulator/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/mechanical-watch-simulator/gpt-6-astra/VALIDATION.md", "validation": "# Mechanical watch simulator validation\n\nImplemented independently from `benchmarks/prompts/10-mechanical-watch-simulator.md`. No existing implementation, reference implementation, model output, or site source was inspected.\n\n## Browser checks\n\nTested in headless Google Chrome through Playwright, served over HTTP. Desktop viewport: 1280 × 800; mobile viewport: 390 × 844. Browser state measurements use the exposed `window.watch` inspection API to read the same object rotations rendered by Three.js.\n\n1. **Immediate load:** movement runs on arrival, canvas renders, and the browser reports no page or console errors.\n2. **Gear ratios:** at 60 simulated seconds the fourth wheel is at −6.283185307 radians; at 3,600 simulated seconds the center wheel is at −6.283185307 radians. Minute/hour motion uses periods of 3,600/43,200 seconds.\n3. **Beat timing:** timed approximately ten real seconds at 1× for each caliber. ETA produces 80 beats, Omega 70, Zenith 100. Spring Drive produces zero beats. The mechanical frequencies are 4, 3.5 and 5 Hz respectively.\n4. **Slow escapement:** entered inspection at 1/50×, let it run, and inspected close-up screenshots. Also sampled the first ETA beat at simulated times 0, 0.075, 0.0875, 0.1125 and 0.125 seconds. Escape rotation stays fixed through lock, moves during impulse/release, then ends at −π/15: exactly half a 15-tooth pitch. Separate lock, impulse and release screenshots were personally inspected. The pallet alternates banking positions and the balance/hairspring change visibly.\n5. **Spring Drive:** no balance or pallet is constructed. Continuous glide samples at 0.01-second intervals advance by −0.502654825 radians; running seconds advances by −0.001047198 radians per sample, without beat quantization. The electromagnetic coil assemblies and moving glide marker are visible.\n6. **Views:** X-ray and exploded modes were activated while running. The five layers separate to Z positions approximately 0, 0.96, 1.91, 2.77 and 3.67; time and wheel rotations continue. Screenshots inspected. Inspection hides the rotor and temporarily collapses the exploded stack so the regulator remains readable; returning restores the prior exploded state.\n7. **Caliber switching:** all four loaded without a page reload, showed their corresponding names/frequencies/regulators/reserves, and preserved camera azimuth, elevation, distance and target. Keyboard movement selection also checked.\n8. **Chronograph:** start rotates the column wheel to π/4 and brings the lateral clutch from X=2.1 to X=1.65. Stop disengages the clutch and preserves elapsed time. Reset zeros the elapsed time/chronograph hand and indexes the column wheel again. Buttons and T/R keys exercised.\n\nAdditional checks: Space pause, X, E, C, G, +/−, arrow selection, number selection, pointer drag, wheel zoom, and reduced-motion startup/play were exercised. Reduced motion starts paused with an obvious Play button. HUD and controls were visually checked at both viewports. Observed render HUD was approximately 60 FPS on the validation host.\n\nThe complete eight-check script was rerun after correcting a stale initial beat count and refining inspection visibility. Test scripts and screenshots were kept in `/tmp`, outside the deliverable.\n\n## Runtime external assets\n\n- `https://cdn.jsdelivr.net/npm/three@0.160.0/build/three.module.js` — Three.js ES module. Add this URL to the integration manifest's `externalAssets` array.\n\nNo external textures, fonts, images, or other runtime assets are fetched. All movement geometry is generated in the single HTML file.\n\n## Limitations\n\n- These are original, explanatory caliber assemblies, not manufacturer CAD replicas. Layouts, tooth geometry and the co-axial pallet mechanism are simplified; contact forces, tolerances, torque and collision-driven gear meshing are not physically solved. Specified rotational periods and beat timing are explicit kinematic functions.\n- 30 FPS on an actual 2020 laptop was not hardware-tested; approximately 60 FPS was observed on the available host.\n- Touch pinch has an implementation but was not tested on a physical touchscreen. No Safari, Firefox or screen-reader audit was performed.\n- The CDN module requires network access on first load. No local library fallback is bundled.\n", "externalAssets": [ { "kind": "script", "url": "https://cdn.jsdelivr.net/npm/three@0.160.0/build/three.module.js" } ] }, { "benchmark": "stormwind-trebuchet-simulator", "agent": "/root/bench11", "prompt": "benchmarks/prompts/11-stormwind-trebuchet-simulator.md", "promptSha256": "9faf5823a18c8cad050066b98c01e3301357784c7fd54005f984753e626a8a76", "artifactSha256": "d0edba93d7632a748e45e61845944872302db195a47c422a5d390fbd6a177757", "files": [ { "path": "index.html", "sha256": "d0edba93d7632a748e45e61845944872302db195a47c422a5d390fbd6a177757" } ], "entry": "/benchmarks/stormwind-trebuchet-simulator/gpt-6-astra/index.html", "manifest": "/benchmarks/stormwind-trebuchet-simulator/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/stormwind-trebuchet-simulator/gpt-6-astra/VALIDATION.md", "validation": "# Validation — Stormwind trebuchet simulator\n\nImplemented independently from `benchmarks/prompts/11-stormwind-trebuchet-simulator.md`. No reference implementation or other benchmark output was inspected. Single HTML file; no build step.\n\n## Browser checks actually performed\n\nChrome headless through Playwright on the available host, 1280 × 800, and a 390 × 844 mobile viewport. Screenshots of loaded, launching, chase, impact, and mobile states were captured and personally inspected. The full regression script was rerun after keyboard and mobile layout fixes.\n\n1. Opened directly in loaded state, with a timber counterweight engine and built white-stone / blue-roof / gold-trim city. No page JavaScript errors.\n2. Fired through the visible button. Inspected launch and chase screenshots; counterweight lowered, arm rotated, sling and stone moved. Simulation automatically entered chase and reached impact.\n3. Stepped through release at 1/240 second. Release-record position and velocity exactly matched the live projectile at the transition; preceding substep displacement was 0.0991 m. Release is an interpolated angular crossing, not a timer.\n4. Zero-wind ground shot at 40° traverse: predicted and actual range were both 170.7613 m (0 m discrepancy).\n5. With the same ground-shot configuration, increasing counterweight from 5,000 to 10,000 kg increased release speed from 41.2849 to 49.3542 m/s and range from 170.7613 to 238.4226 m.\n6. Default release time was 1.32740 s; pin 60° changed it to 1.40367 s, and sling length 9 m changed it to 1.34887 s. Predicted arcs changed with the mechanism.\n7. Wind 30 m/s at 90° produced more than 11 m lateral drift. Prediction and live collision coordinates matched exactly.\n8. A synthetic swept segment from x=100 to x=250 at y=5, z=25 hit the wall at x=162.5. This calls the same collision function as live flight.\n9. Fired actual configurations for Ground (pin 80°, aim 0°), Gate (76°, 0°), Wall (76°, 8°), and Tower (74°, 14°). Verified the reported collision kind and localized point. Gate impact screenshot showed the localized mark and fragments. Tower collision uses the rendered 12-sided tower footprint, not just its enclosing box.\n10. During flight, C cycled wide → trebuchet → chase. R started replay; the stored trajectory remained byte-identical during replay. Replay reads recorded frames.\n11. Space paused at exactly 1.96667 s over a 300 ms wait, then resumed. T changed trajectory visibility. Enter returned to loaded state. Repeated reset/fire completed. Tab advanced counterweight → projectile, Shift+Tab reversed it. Mobile drawer button opened its scrollable tuning panel.\n12. A browser context with reduced motion enabled defaulted to wide view and completed the shot in wide view. No camera shake is implemented in either mode; reduced motion lowers the spring strength and disables idle orbit.\n13. Two resets followed by fire and identical `step(30)` calls produced byte-identical telemetry.\n14. Separate real-time frame measurements on this host: launch 58.05 fps, chase 58.88 fps, impact 60.00 fps. Final regression also measured approximately 60 fps. These are host-specific measurements, not a certification of a 2020 laptop.\n\n## Physics / implementation notes\n\nThe coupled two-coordinate mass matrix includes counterweight and arm inertia, arm–sling coupling, gravity, configurable axle friction, sling damping, and aerodynamic force. Mechanism integration uses RK4 at 240 Hz. Counterweight remains vertical below its hinge. Release copies the computed pouch state into flight. Flight uses gravity, quadratic drag, stone-density-derived frontal area, relative air velocity, and swept collisions. Prediction calls the same advance function as live simulation. Geometry, materials, terrain details, and debris are procedural; scenery randomization uses a fixed seed.\n\n## Limitations\n\n- Performance on an actual 2020 laptop could not be certified with the available host.\n- The vertical counterweight is constrained to hang vertically; it is not a third freely swinging rigid-body coordinate. The sling is an ideal taut pendulum before release. Rope elasticity, pouch-ground contact at extreme settings, and sling/frame contacts are not modeled.\n- The arm is represented by its configurable inertia; distributed arm gravitational loading is not modeled separately. Energy HUD exposes component values, not a complete conservation residual.\n- Collision targets are the outer walls, gate, and polygonal tower bodies. Decorative roofs, rear skyline, and the practice shield props do not have individual collision solids.\n- Replay samples stored state at 30 Hz. Its path is unchanged, but it does not interpolate every stored mechanism frame.\n- Impact marks and fragments are localized; there is no volumetric dust simulation or large-scale structural destruction.\n- WebGL and the library CDN must be available at startup. No offline library fallback is bundled.\n\n## Runtime external assets\n\nOnly this JavaScript library is fetched at runtime (must be listed by integration in `manifest.json.externalAssets`):\n\n- `https://cdn.jsdelivr.net/npm/three@0.169.0/build/three.module.js`\n\nNo runtime external art, textures, models, audio, fonts, or data are fetched.\n", "externalAssets": [ { "kind": "script", "url": "https://cdn.jsdelivr.net/npm/three@0.169.0/build/three.module.js" } ] }, { "benchmark": "starfall-arena", "agent": "/root/bench12", "prompt": "benchmarks/prompts/12-starfall-arena.md", "promptSha256": "50d1e4dd79f17c0b5ff7256a04bca2747493ce3636fa430db0237ff194c2eb36", "artifactSha256": "a9e6b4512cbc2b085cb09d8af661c26fac7930608871c00926f2a92e12b097a8", "files": [ { "path": "index.html", "sha256": "a9e6b4512cbc2b085cb09d8af661c26fac7930608871c00926f2a92e12b097a8" } ], "entry": "/benchmarks/starfall-arena/gpt-6-astra/index.html", "manifest": "/benchmarks/starfall-arena/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/starfall-arena/gpt-6-astra/VALIDATION.md", "validation": "# Starfall Arena — validation\n\nImplementation: `index.html`, standalone raw Canvas 2D with HTML HUD and accessible buttons. No build step. No external runtime assets, fonts, libraries, images, audio, analytics, or data requests. `manifest.json.externalAssets` should be `[]`.\n\nBrowser checks performed September 7, 2026 with installed Google Chrome through Playwright, headless, at 1280 × 800. Initial loading and keyboard checks used the actual `file://` document. Controlled combat checks used the exact implementation in a browser with a temporary, test-only extension exposing its existing closure functions and state. This instrumentation was not included in the delivered file. Browser screenshots of start, late combat, and 390 × 844 layout were inspected.\n\n1. **Start:** exact goal and controls visible; Space starts immediately. No browser script errors across the checks.\n2. **Movement and firing:** WASD moves in all four directions. Boundary clamping verified at (30, 125) and (1250, 715). With targets placed at 700 and 900 on the player's horizontal axis, the real targeting function fired toward the nearer target.\n3. **Combat and XP:** real projectile updates killed two enemies, left a gem, and contact collection raised XP to 6 with pickup particles.\n4. **Level-up:** grantXP produced exactly three button cards. Keyboard 1 increased projectile count from 1 to 2 in the recorded run; clicking another card installed armor. All nine enhancement definitions were separately applied twice through the real selection function, with measurable state changes and stack counts of two.\n5. **HUD:** inspected in screenshots; wave, timer, numeric HP, XP, level, kills, goal, and milestone are displayed at the edges. HP and XP values tracked the damage and collection checks. Gameplay timer advanced; it froze during pause and upgrade selection.\n6. **Enemy types:** naturally spawned all four types after jumping to wave 10 and advancing real updates. A controlled behavior check showed direct horizontal chaser pursuit, off-axis weaver drift, two minis from splitter death, and elite HP 240 versus chaser HP 32 at wave 1.\n7. **Defeat/restart:** real damage reduced HP from 100 to 80, then to zero. Defeat displayed five completed waves after a jump to wave 6, level 1, one kill, and elapsed time. Clicking Restart reset wave, level, HP, kills, and time without navigation.\n8. **Boss:** hook jump to wave 15 displayed its HP bar with 18,000 HP. Phase 1 emitted 16 radial bullets; crossing 67% HP cleared bullets and set a 1.98-second transition telegraph. Phase 2 emitted 21 bullets (radial plus aimed volley). Crossing 33% HP entered phase 3; its attack spawned two chasers. All calls used actual damage/update functions.\n9. **Victory:** actual boss damage and delayed explosion completion produced victory with 15 waves survived and matching run stats.\n10. **Pause/reduced motion:** keyboard P froze time, Esc resumed. Browser reduced-motion emulation produced `reduced: true`, zero shake, and normal damage (HP 90).\n11. **Best run:** victory persisted `{ \"wave\": 15, \"victory\": true }` in localStorage. Return to Start displayed “BEST RUN / STARFALL DEFEATED.”\n12. **Performance:** approximately 59.2 fps measured using requestAnimationFrame intervals during wave 12 combat with 223 live enemies, 683 particles, and 62 shots at 1280 × 800. A separate run measured approximately 60 fps. Reduced motion was disabled for the stress sample.\n\nAdditional check: 390 × 844 layout had no horizontal overflow.\n\nLimitations: no full real-time human 6–10 minute playthrough or first-time-player difficulty study was performed. Fourteen 27-second waves place the boss at 6:18 active play time; upgrade decisions and boss combat extend the run. End states and boss thresholds were exercised with controlled browser instrumentation rather than an unassisted boss kill. Performance is a local Chrome sample, not a guarantee on every device. No cross-browser or physical touchscreen testing was performed.\n", "externalAssets": [] }, { "benchmark": "neon-drift", "agent": "/root/bench13", "prompt": "benchmarks/prompts/13-neon-drift.md", "promptSha256": "52425362117e2baaf69eb73b55b68223eeae9b485d9a8f2c0ef2ba2e708e7738", "artifactSha256": "42f8dcf730d5c3eb974b5648fa1183e9c119dd06dc04b4993ebb103e9222725c", "files": [ { "path": "index.html", "sha256": "42f8dcf730d5c3eb974b5648fa1183e9c119dd06dc04b4993ebb103e9222725c" } ], "entry": "/benchmarks/neon-drift/gpt-6-astra/index.html", "manifest": "/benchmarks/neon-drift/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/neon-drift/gpt-6-astra/VALIDATION.md", "validation": "# Neon Drift — GPT-6 Astra validation\n\nImplemented from `benchmarks/prompts/13-neon-drift.md` without inspecting other benchmark implementations, reference code, or site source. The output is a self-contained `index.html` using native Canvas 2D, DOM controls, and system fonts.\n\n## Browser checks performed\n\nTested September 7, 2026 in installed Chrome through Playwright at 1280×800. Inspected menu, active drift with ghost, and champion screenshots. Browser page-error listeners reported no errors. Test scripts were kept outside the deliverable in `/tmp/neon13-check.mjs`, `/tmp/neon13-extra.mjs`, and `/tmp/neon13-fps.mjs`.\n\nThe deterministic race tests use the public input hook and the same production `step(1/120)` function in the browser; they do not change car positions, completion times, unlocks, or checkpoint state. A pursuit controller supplied steering/throttle/brake inputs. A separate real-time requestAnimationFrame sample was used for performance, without Playwright's virtual clock.\n\n1. **Initial screen:** reset progress; verified exact goal sentence, three track buttons, two visibly locked tracks, locked selection rejected, and race startup.\n2. **Driving inputs:** throttle increased speed above 100; brake reduced speed; reverse drove backward. Steering changed heading. Handbrake produced a nonzero slip angle, the DRIFTING indicator, and visible twin light trails.\n3. **Surfaces, bounds, checkpoints:** drove into runoff and perimeter collision; speed fell below 90 world units/s. Checked world bounds on 600 successive simulation steps. Reversing across the start and driving off-road did not award a lap or bypass the next ordered checkpoint.\n4. **Lap timing:** drove all three circuits to completion. Sunset completed three laps; sum of recorded lap durations matched race total within 0.001 seconds. Countdown did not contribute to timing.\n5. **Delta:** waiting at the start produced positive/behind delta; a successful drive produced negative/ahead delta. HUD uses pink/red for positive delta and mint green for negative delta.\n6. **Unlocking and retry:** Sunset completed in approximately 32.69 seconds against 39 seconds gold, unlocked Canyon, and updated the visible ladder. A separate 71.59-second Sunset run showed the gap and Retry while keeping unlocks at one.\n7. **Ghost:** completed lap generated over 200 actual position/heading/time samples. Restart replayed a sample from that recording verbatim at the appropriate lap time. Screenshot inspection confirmed a translucent mint ghost alongside the pink drifting player.\n8. **Keyboard:** P paused, time remained unchanged during a one-second clock advance, Enter resumed, and R reset the countdown and timing. Escape → track ladder → arrow navigation → Enter started a race. Fixed a focus retention bug and reran both complete check scripts after the fix.\n9. **Persistence:** reload preserved unlocks and best times. `resetProgress()` followed by reload restored the fresh locked ladder. File URL loading also returned `window.__drift.ready === true`.\n10. **Champion and Free Run:** Canyon finished in 57.45 seconds against 65 gold; Grid City finished in 70.37 seconds against 76 gold. Champion screen displayed all three career best times. Clicking Free Run and driving beyond three laps kept the game in racing state.\n11. **Reduced motion:** repeated identical input/step sequences with media preference set to normal and reduce. Car results were exactly identical; normal camera scale target was approximately 0.970, reduced target was fixed at 1.\n12. **Performance:** separate real-time six-second Chrome run at 1280×800 measured **60.00 fps**, 362 sampled frames, 193 frames in the drifting state, and an active 206-sample ghost recording. No virtual clock was installed for this measurement.\n\n## Runtime external assets\n\nNone. No libraries, fonts, art, textures, audio, data, analytics, or other runtime resources are fetched. Request inspection observed only the document navigation/reload requests. `manifest.json.externalAssets` should be `[]` when integrated by the parent task.\n\n## Limitations\n\n- Performance was measured on the available host, not a separately provisioned 2020 laptop. The stated hardware target cannot be certified from this environment.\n- Gold times were calibrated with deterministic scripted drives; subjective drift feel and difficulty have not been tested with a panel of human players.\n- Ghost playback uses recorded samples without interpolation, preserving the recorded states at approximately 20 Hz.\n- Validation covers Chrome and a 1280×800 desktop viewport; other browser engines and touch devices were not tested. Touch controls are not implemented.\n- File URL loading works, but persistence under file URLs remains subject to individual browser storage policies; unavailable storage degrades to session-only progress.\n", "externalAssets": [] }, { "benchmark": "low-poly-tower-defense", "agent": "/root/bench14", "prompt": "benchmarks/prompts/14-low-poly-tower-defense.md", "promptSha256": "3afadb6cc349189cfca6570e4e64352c242881cf87b15315245338fc64cf9e45", "artifactSha256": "7255dc0110f406a0e4686e6198382d59bcf6d474285191aa68f024d3d254b2ad", "files": [ { "path": "index.html", "sha256": "7255dc0110f406a0e4686e6198382d59bcf6d474285191aa68f024d3d254b2ad" } ], "entry": "/benchmarks/low-poly-tower-defense/gpt-6-astra/index.html", "manifest": "/benchmarks/low-poly-tower-defense/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/low-poly-tower-defense/gpt-6-astra/VALIDATION.md", "validation": "# Heartwood — validation\n\nValidated in headless Google Chrome through Playwright at 1280×800, with an additional 390×844 reduced-motion viewport. Tests ran against the actual HTML served from the parent static server and also opened the HTML directly using `file://`.\n\n## Method\n\nThe required `window.__defense` hook calls the same placement, upgrade, sale, wave, speed, and restart functions as the UI. Long runs were accelerated by invoking the live game's `step(1/60)` function inside the browser; no separate simulation or alternate economy was used. Input tests used actual keyboard events. Targeted leak and combat boundary checks positioned enemies or supplied fixture gold in the live browser state. Performance and the speed-ratio check used real requestAnimationFrame elapsed time.\n\n## Checks performed\n\n1. **First screen and persistence:** The exact goal, “Stop 10 waves. Don't lose 20 lives.”, controls, and start button appeared. Setting the stored best to 4 and reloading showed “YOUR BEST · WAVE 4 / 10”. Starting produced no JavaScript errors.\n2. **Scene readability:** Inspected desktop screenshots of the intro, build phase, boss wave, victory, and defeat. Procedural island cliffs, pines, clouds, winding path, labeled rift and island heart, and 12 marked build plots were visible.\n3. **Placement and economy:** Starting gold 130; two 60g arrow towers left 10g. Invalid ID returned false and displayed build-plot feedback. Occupied plot was rejected without changing gold. Keyboard Tab/Enter built arrows on plots 4 and 8.\n4. **Composition and automatic combat:** Wave 1 spawned exactly its six previewed soldiers. Path progress increased and projectiles fired automatically. All ten waves were completed in order; cumulative composition was 84 soldiers, 91 runners, 39 brutes, and one boss.\n5. **Bounties and leaks:** Instant bounty deltas were runner 9g, soldier 12g, brute 28g, boss 350g. Boundary leaks deducted 1, 3, and 10 lives for soldier, brute, and boss. Wave 1 yielded 72g from six soldier kills.\n6. **Splash and frost:** One cannon projectile damaged two nearby soldiers from 55 HP to 23 HP each; splash-hit diagnostic counted 2. Frost-only fixture retained full HP, applied cyan `#91dfe5`, and reduced one second of movement to approximately 1.02 units compared with 1.45 normal. Frost base slow was 30%; upgraded values were 45% and 60%.\n7. **Upgrades and refunds:** Tested both upgrades for all three types. Arrow damage 15→25→43, range 3.6→4.2→4.8, interval .55→.43→.32 seconds; cannon damage 32→57→90; frost range 3.2→3.8→4.5. Tower head height increased .95→1.18→1.41 and accent components increased 4→5→6. Arrow total investment 255g refunded floor(255×.75) = 191g.\n8. **Controls:** Enter started the run; Tab/Enter selected and built two plots; 1 selected arrow; Space launched the wave; the keyboard-established defense completed wave 1 with 20 lives. P froze enemy snapshots. F toggled to 2×; real-time movement measured approximately .725 units at 1× and 1.45 at 2× over equal half-second intervals. Code routes 2/3 to cannon/frost and Esc to the same pause function.\n9. **Endings and restart:** A complete economy-funded ten-wave run killed the boss and reached victory with 20 lives, 215 kills, 3,269 earned gold, and ten built towers. A one-life leak fixture reached defeat at zero. Both ending screens displayed run statistics. Clicking their restart buttons reset to build phase, 130g, 20 lives, wave 0, speed 1, zero counters, no towers, and no enemies without a reload.\n10. **HUD:** Desktop screenshots confirmed gold, lives, wave /10, and Start Wave remained visible and matched hook values. The boss has an additional large health bar. Inspected the final mobile viewport after a responsive-layout correction; controls and full island fit the screen.\n11. **Reduced motion:** A reduced-motion browser completed wave 1 with six kills and 20 lives. Crystal bobbing/rotation and bounty drift are suppressed; combat stays active.\n12. **Performance:** A real three-second requestAnimationFrame sample during wave 10 with the boss present measured **60.32 fps** at 1280×800. No JavaScript page errors were recorded. Direct `file://` loading exposed a ready verification hook.\n\nThe twelve-check run was repeated after the final layout changes. Additional boundary tests verified all bounty values, all tower upgrade geometry, occupied tiles, frost-only health, and the real-time speed ratio.\n\n## Runtime external assets\n\n- `https://cdn.jsdelivr.net/npm/three@0.160.1/build/three.min.js` — Three.js rendering library.\n\nAll geometry, materials, effects, layout, and game data are authored in `index.html`. No external models, textures, fonts, audio, analytics, or data services are used.\n\n## Limitations\n\n- Opening the single HTML file requires network access for the listed Three.js CDN library, or an existing browser cache. No offline copy is embedded.\n- Performance was measured on this validation machine in Chrome; other GPUs and browsers may differ. Safari and Firefox were not tested.\n- Long-run validation accelerated the actual browser game loop rather than waiting through every wave in real time. Defeat and exact leak boundaries used live-state fixtures; victory used the normal starting economy and earned gold.\n- Small screens fit the entire island at a smaller scale; desktop is the primary play surface.\n- No audio is included.\n", "externalAssets": [ { "kind": "script", "url": "https://cdn.jsdelivr.net/npm/three@0.160.1/build/three.min.js" } ] }, { "benchmark": "ember-glider", "agent": "/root/bench15", "prompt": "benchmarks/prompts/15-ember-glider.md", "promptSha256": "4c8d400bfae941324ed383aa454844adc1a90a8f786411ea2a50bf5e1c5e89e9", "artifactSha256": "e87da01d5279e3cf4474cbcab058b7f0aa2413e68e63866d0a374d2a4c3b9ebf", "files": [ { "path": "index.html", "sha256": "e87da01d5279e3cf4474cbcab058b7f0aa2413e68e63866d0a374d2a4c3b9ebf" } ], "entry": "/benchmarks/ember-glider/gpt-6-astra/index.html", "manifest": "/benchmarks/ember-glider/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/ember-glider/gpt-6-astra/VALIDATION.md", "validation": "# Ember Glider validation\n\nImplemented independently from `benchmarks/prompts/15-ember-glider.md`. No reference implementation or other benchmark implementation was inspected.\n\n## Browser checks performed\n\nTested in headless installed Google Chrome using Playwright, at 1280 × 800, over the parent HTTP server and separately over `file://`. No page JavaScript errors occurred. Browser screenshots were captured to `/tmp/ember.png` and `/tmp/ember-flight.png`; the opening composition was visually inspected.\n\n1. Opening screen contains the exact goal sentence, keyboard controls, 3:00 par, and stored best after reload. Enter launches the game.\n2. The actual flight update function was exercised with scripted pitch/bank inputs. Diving increased speed to 33.19 m/s; subsequent climbing reduced it to 26.34 m/s. A forced low-speed condition activated stall recovery. Bank interpolated to 0.195 radians after a 0.1-second update rather than snapping to its target.\n3. Still-air altitude decreased. An island thermal increased altitude. A rendered HUD check at an injected 110 m position displayed 110.\n4. Teleporting near ring 3 and advancing the same game update increased progress to 3 and stored checkpoint (0, 82, -280). The collection code triggers the expanding ring burst and checkpoint notice.\n5. Injecting a below-water position caused the crashed phase, one 10-second penalty, and subsequent respawn near the saved checkpoint. Three collected rings remained collected.\n6. Off-screen target tests on opposite sides produced `← NEXT RING 319 m` and `→ NEXT RING 319 m`.\n7. The public final-ring teleport hook followed by a gate-position fixture ended the run at The Beacon, displaying total time, par, 12 rings, and crash count.\n8. Under-par finish displayed Ember Medal and persisted its best across reload. An elapsed-time fixture of 181 seconds displayed 3:01 and Medal missed.\n9. The Retry button restarted the run. Keyboard P paused with elapsed time unchanged across a 200 ms browser wait. Free Flight stopped the clock and a water collision added no penalty.\n10. An island-surface fixture landed safely. The shared relaunch routine cleared landing and restored flight speed.\n11. Normal and emulated reduced-motion runs with identical update steps and inputs returned identical glider state. Reduced-motion camera up was exactly [0, 1, 0], removing roll; the stall shudder is also disabled.\n12. Measured approximately 60.3 requestAnimationFrame updates per second for two seconds while positioned inside a thermal, with all 900 animated ember particles rendering at 1280 × 800.\n\nAfter the final visual/thermal changes, the core browser suite was rerun, followed by the directional, HUD, file URL, and reduced-motion checks.\n\nAdditionally, a scripted feedback pilot completed the actual continuous flight update from launch through all 12 rings and the beacon in 51.82 seconds with zero crashes. It used only bounded pitch/bank commands, without teleporting or modifying position during that route. This establishes a flyable route with substantial margin under the 180-second par.\n\n## Runtime external assets\n\n- `https://cdn.jsdelivr.net/npm/three@0.160.1/build/three.min.js` — Three.js runtime library.\n\nNo external fonts, textures, models, audio, or runtime data are used. All geometry, sky shading, water glitter, trees, waterfall surfaces, particles, glider, and rings are generated in code.\n\n## Limitations\n\n- The performance measurement is from this environment's Chrome, not a physical 2020 laptop; that exact hardware was unavailable.\n- Physics edge cases were tested using browser fixtures invoking the same update functions, rather than a complete manual keyboard playthrough. The continuous full-route test used scripted input.\n- Three.js requires network access on first load, including when opening the HTML by file URL. Best-time persistence depends on browser storage availability and file-origin policy.\n- Waterfalls are stylized translucent geometry. Water has procedural glitter rather than a fluid simulation. No audio is included.\n", "externalAssets": [ { "kind": "script", "url": "https://cdn.jsdelivr.net/npm/three@0.160.1/build/three.min.js" } ] }, { "benchmark": "city-scroll-journey", "agent": "/root/bench16", "prompt": "benchmarks/prompts/16-city-scroll-journey.md", "promptSha256": "287577ab9b81ef7fbe717323d3263e282e59fc0672e837b025128ad0f4ad24e6", "artifactSha256": "47bf86a5532d80bb38c0c847f35b53942716e6a3ce3fe343319df238a732b556", "files": [ { "path": "GENERATION.md", "sha256": "1d2f4f51541e13d1fd8f6111eb7927e02b8f3e0e121305a2cff3e65799f6a546" }, { "path": "assets/bifana.svg", "sha256": "0c0eb6d07c507705f33d865c985111a31222ed6472c78b7ec39f3ad068aea88c" }, { "path": "assets/nata.svg", "sha256": "4dd86eed99bb8d5c86c8d89ff247192fbc4faea262b64511cf4638d25cfa821b" }, { "path": "assets/place-0.png", "sha256": "6df9244a3fb5a5a71f4eda8cae831bc09f9981c935d15d29412f2fbc9cce28c4" }, { "path": "assets/place-1.png", "sha256": "f0112e027c2477c24a0fede588871c95d5fd63eabf1e8cc240a09158f5560d9f" }, { "path": "assets/place-2.png", "sha256": "afbcaee1708904c7245060057056ac6b523888719057806e100bef32ca0540e6" }, { "path": "assets/sardines.svg", "sha256": "6866add03d2715d72269c305266eec53856fb3de68ce7e6c994c6892b1534865" }, { "path": "assets/scene-0.png", "sha256": "f05b54ee130836d2ac6b1614507e06264724bfeae47ccbc395a11f5e9a77322f" }, { "path": "assets/scene-1.png", "sha256": "a8165b33766c3881206aa6cb3d167c8223c5d001f11a953df9eecdfb8fd6c9b5" }, { "path": "assets/scene-2.png", "sha256": "d117060aa3c1d0a3ab39a49753906753bf2c378ef83504b84a1c7c65ece0ca04" }, { "path": "assets/scene-3.png", "sha256": "650652d21726cd469741224629d07c91ad7b86828a11e945ad6d44b301c7199c" }, { "path": "index.html", "sha256": "47bf86a5532d80bb38c0c847f35b53942716e6a3ce3fe343319df238a732b556" }, { "path": "journey.js", "sha256": "3b64545c648c6accee9bac5e966797292668027278cd53950ae4114e9f10dba2" }, { "path": "style.css", "sha256": "0c523b204e53b1f4eb318612f974c9d052d2192b596126b51316638b255bbb85" }, { "path": "validation-results.json", "sha256": "a7a5399afd382a3e649a8b87f5874a1a37588eefb695a155f5b6002974d78522" } ], "entry": "/benchmarks/city-scroll-journey/gpt-6-astra/index.html", "manifest": "/benchmarks/city-scroll-journey/gpt-6-astra/manifest.json", "validationReport": "/benchmarks/city-scroll-journey/gpt-6-astra/VALIDATION.md", "validation": "# Validation — Little Lisbon\n\nVerified in headless installed Chrome through Playwright, opening both `file://` and the operator's local HTTP server. Desktop viewport 1280×800; mobile viewport 390×844, resized to 390×740. Final automated result snapshots are in `validation-results.json`. Desktop, all four scene dwell views, mobile, reduced-motion, destination, food, poster-only, and dark-mode screenshots were captured and personally inspected during the build.\n\n## Required checks\n\n1. **Initial poster:** passed. The local first PNG loaded with JavaScript disabled. Every image loaded with nonzero natural width. Normal opening paints the same original city canvas immediately. No video matching requirement applies.\n2. **Bidirectional native scroll:** passed. Native scroll positions 0, 600, 1800, 3000, 4200, 4800, then back to 0 produced the expected active scenes and release state. Canvas pixels remained identical across a 700ms idle interval; the camera does not autoplay.\n3. **Scene distinction:** inspected all four dwell screenshots. Rooftops/church, yellow tram street, arch/statue square, and tower/river are distinct locations in one miniature palette.\n4. **Seams:** the camera remains in one world; no media swaps occur. Samples immediately before/after progress .25, .5, and .75 contained zero black sampled pixels and similar image brightness. Full-journey scrubbing completed with no loading or console errors.\n5. **Scene copy:** all four native dwell positions produced opacity 1 and the corresponding title, eyebrow, body and accent. Boundary samples receded to opacity .03. Camera easing has zero derivative at each dwell.\n6. **Final CTAs:** clicked both actual links. Primary reached `#destinations`; secondary reached `#food`.\n7. **Reduced motion:** emulated. Pinned journey is `display:none`; four static scene articles are visible in normal flow with relative positioning and full copy/CTAs. Ambient animation and reveal transitions are disabled.\n8. **Mobile:** 390px wide loaded and scrubbed. Local mobile measurement reported 55.2 fps. Height-only resize retained a 5908px journey height, scrollY 2000, and the same progress .3949447 with no horizontal overflow. The route height uses a width-sensitive stable viewport measurement.\n9. **Supporting sections:** three destination cards and three food items loaded local visuals. Destination reveal opacity reached 1 after native scroll; food section revealed and was visually inspected.\n10. **Nav and release:** transparent overlay initially. After release, computed nav position is `sticky` and its bounding rect top is 0. Supporting sections participate in normal page flow.\n11. **Hook:** scene metadata and state agree with the four ordered scenes. `setProgress(.5)` returns scene index 2 with the square's copy (at its incoming transition, opacity .03). `setProgress` calls the same update/render paths as native scroll.\n12. **Performance:** full 181-step programmatic scrub driven by requestAnimationFrame reported 60 fps at 1280×800 in local Chrome. The hook measures actual render cadence. Exact target hardware was not available.\n13. **Offline:** `file://` loaded with zero non-file requests. HTTP opening had no console errors or failed requests. Runtime external assets are empty.\n14. **Provenance:** all PNG and SVG media were created during this build, inside the artifact folder. Prompts and named-tool substitutions are recorded in `GENERATION.md`.\n\nAdditional checks: first keyboard Tab reaches the visible-on-focus skip link; no horizontal overflow at desktop/mobile; dark color-scheme emulation applies the alternate supporting-surface palette; all image files complete successfully.\n\n## Limitations\n\n- Higgsfield and Mobbin were not callable. This is a local-rendering fallback, not generated cinematic video or a Mobbin-informed build.\n- The camera is an eased continuous orthographic traversal of a stylized miniature model. Lisbon geography and monuments are deliberately simplified rather than geographically accurate.\n- Performance numbers are from the available local headless Chrome environment. Physical 2020 laptop performance and actual mobile browser address-bar behavior were not tested; viewport resize was simulated.\n- Static reduced-motion posters are desktop compositions cropped responsively. They retain the city story but do not reproduce the default mobile camera framing exactly.\n- No assistive technology screen-reader session was performed; semantic controls, labels, focus and native keyboard scrolling were checked.\n\n## Runtime external assets\n\n`[]` — no external assets, libraries, fonts, APIs, media, or tracking.\n", "externalAssets": [] } ] }