An ocean,
contained.
A procedural Three.js ship-in-a-bottle study, developed live with screenshot checkpoints and a guided, labeled camera tour.
The performance verdict.
The selected measured configuration is gpu-opt-05 / optimization level 14. Mean frame time changed from 85.63 ms to 51.55 ms (39.8% reduction). CPU submission: 16.22 → 8.45 ms. GPU elapsed: 75.84 → 45.81 ms. Draw calls: 488 → 489.
The measured optimization series uses actual WebGPU on an Apple M4 Pro with a 20-core GPU. All authored shader effects are Three.js Shading Language (TSL) graphs and NodeMaterials. The same scene also runs on the WebGL2 backend of WebGPURenderer. Two accessible tabs switch backends with settings preserved; unavailable WebGPU is disabled and fallback is explicitly labeled. Three.js is pinned to 0.186.1 / r186, the latest npm release rechecked on 1 October 2026 JST.
| Run | Change | Mean ms | FPS | p95 ms | CPU ms | GPU ms | Gain vs best | Plateau streak | 9-view fidelity |
|---|---|---|---|---|---|---|---|---|---|
| gpu-opt-00 | Cinematic TSL baseline | 85.63 | 11.7 | 120.90 | 16.22 | 75.84 | +0.00% | 0 | PASS |
| gpu-opt-01 | Skip zero-contribution dielectric SSR rays; allocation-free normals | 62.59 | 16.0 | 96.10 | 16.19 | 53.70 | +26.91% | 0 | PASS |
| gpu-opt-02 | Shared GPU cloud optical-depth light cache | 56.41 | 17.7 | 89.50 | 17.63 | 47.22 | +9.87% | 0 | PASS |
| gpu-opt-03 | 16-bit wave display; 32-bit FFT retained | 55.51 | 18.1 | 93.40 | 8.66 | 48.58 | +1.60% | 1 | PASS |
| gpu-opt-04 | Low-resolution 2×2 cloud-light cache — rejected by RMS audit | 51.79 | 19.4 | 87.30 | 8.90 | 45.72 | +6.70% | 0 | REJECT / incomplete |
| gpu-opt-05 | Refined optical-depth cache: 256×64×128, 2×2 amortization | 51.55 | 19.4 | 84.60 | 8.45 | 45.81 | +7.13% | 0 | PASS |
| gpu-opt-06 | Typed, phase-cached spectral buoyancy | 52.00 | 19.2 | 87.40 | 8.62 | 45.94 | -0.87% | 1 | PASS |
| gpu-opt-07 | Freeze nonmoving world transforms | 58.55 | 17.1 | 92.00 | 8.51 | 51.16 | -13.57% | 2 | PASS |
| gpu-opt-08 | Reuse FFT quad and camera across butterfly stages | 55.74 | 18.0 | 91.80 | 8.61 | 48.25 | -8.12% | 3 | PASS |
| gpu-opt-09 | Cache shadow-light traversal | 57.44 | 17.4 | 87.80 | 8.48 | 49.64 | -11.42% | 4 | PASS |
| gpu-opt-10 | Pre-bit-reverse spectrum and conjugate inputs | 55.31 | 18.1 | 87.90 | 9.51 | 47.82 | -7.30% | 5 | PASS |
Stopping rule: MET
5 consecutive distinct optimization candidates produced less than 5% improvement versus the fastest prior quality-accepted incumbent. The last five observations: gpu-opt-06: -0.87%, gpu-opt-07: -13.57%, gpu-opt-08: -8.12%, gpu-opt-09: -11.42%, gpu-opt-10: -7.30%.
Fixed 1280×720 viewport, DPR 1, Cinematic quality, 100% tempest, 150 warmup and 360 sampled frames. Simulation advances at a fixed 1/60 second for reproducibility; this is a scene benchmark, not a claim about every interactive camera. The queue is capped at two frames and FPS is derived from actual GPU-completion fences, not JavaScript submission speed. Whole-frame timestamp markers span all captures and post-processing, rather than adding overlapping pass timers. GPU elapsed can include contention and submission gaps; CPU and GPU durations overlap and must not be added. Timing differences are observational, not statistically isolated causal estimates. Earlier WebGL experiments, invalid summed GPU timers and submission-only FPS readings are excluded from the formal acceptance series. Measurements are immutable locked snapshots; later hot reloads cannot replace them.
Detail preserved, measured.
Nine deterministic views compare the selected candidate with the corrected baseline at the same simulation time. Both the whole scene and detail-region tests must stay below 1% local-luminance SSIM loss and 1% mean absolute RGB error. The worst accepted SSIM loss is 0.2736%. Only the benchmark panel is masked.
No reduction of viewport, pixel ratio, wave-field size, cloud ray steps, cloth lattice, texture-map detail or geometry is used to pass these tests. Cinematic retains 80 primary cloud steps, eight self-shadow samples, 4096² shadows, full-resolution SSR/GTAO, four-sample MSAA and the authored geometry. World-space light caching retains primary density and silhouette sampling. The low-resolution temporal cache is rejected after its lighting RMS error exceeded 1%; the refined cache must pass this extra gate. It is not screen-space temporal reprojection. Screenshot agreement is not proof of photographic realism or a universal perceptual guarantee.
What was built.
Sea, sky and optics
- 128×128 directional wind-wave spectrum and GPU inverse FFT; finite-depth gravity dispersion, horizontal choppiness, wave normals and persistent breaking foam. FFT accumulation remains 32-bit; a separately audited 16-bit display texture reduces repeated sampling bandwidth without reducing wave modes.
- Planar reflections, refraction/depth capture, Fresnel response, absorption and underwater fog, with full-resolution SSR, GTAO and depth-aware denoising in Cinematic. These supplement environment reflections; they are not hardware ray tracing or full path tracing.
- Bounded 3D storm clouds, self-shadowing, rain, lightning and room-derived bottle reflections. The WebGPU acceleration caches light in a world-space field, updating a 2×2 phase per frame and refreshing completely on pause/quality/weather transitions. Camera motion does not reproject screen history; this avoids screen-space trail accumulation.
- Hull-scale buoyancy probes retain over 97% of spectral energy; their smoothing is a real-time approximation, not a full pressure-integrated fluid solution.
Collision and clearance
- Aft shrouds and outboard backstays clear the helm; the test suite checks every standing-rigging segment against its clearance volume.
- Cloth responds to capsule contacts and complete segment/triangle intersections, including thin rigging, mast fittings and yards. Conservative broadphase rejection avoids only impossible contacts.
- Walking checks solid ship geometry; rain is masked out of dry compartments. The water surface excludes the moving hull footprint.
Ship, whale and historical detail
- Three masts, nine constrained cloth sails, stays/ratlines, cannon battery, deck fittings, a working wheel and companion hatch.
- Captain's cabin, gun deck and cargo hold: beams, furniture, chart desk, cot, hammocks, cannon carriages, provisions and worn surfaces.
- Metal-nib dip pen and gimballed 1847-style marine chronometer, with wear, fine geometry and moving hands.
- Circling blue whale: mottled skin, integrated throat-fold relief, small recessed eyes and orbital folds, tapered hydrofoil flukes with span-wise flex, surfacing and moist exhalation.
- 23 labeled cinematic stops, including a tracking shot alongside the whale and close-ups of its tail, eye, the pen, clock and rigging.
- Original Call Me Sensei SVG lockup, following Toonlab's byline treatment while retaining Tempest's mark.
Where the time goes.
Feature ablations — diagnostic only
| Run | Frame ms | Completed FPS | CPU ms | GPU elapsed ms | Graphics valid |
|---|---|---|---|---|---|
| diag-full | 87.82 | 11.4 | 8.71 | 76.22 | True |
| diag-no-water | 52.68 | 19.0 | 6.86 | 45.00 | True |
| diag-no-shadows | 77.92 | 12.7 | 14.46 | 70.18 | True |
| diag-no-lighting | 63.24 | 15.8 | 10.68 | 55.49 | True |
| diag-no-clouds | 73.05 | 13.7 | 11.45 | 64.49 | True |
| diag-no-ssr | 52.83 | 19.1 | 8.71 | 46.95 | True |
| diag-no-ao | 92.01 | 10.9 | 10.56 | 81.13 | True |
Backend and quality measurements
| Run | Frame ms | Completed FPS | CPU ms | GPU elapsed ms | Graphics valid |
|---|---|---|---|---|---|
| verify-gpu-cine | 24.01 | 41.7 | 11.65 | 20.46 | True |
| verify-gpu-high | 18.37 | 54.5 | 9.76 | 15.43 | True |
| verify-gl-cine | 30.74 | 32.6 | 16.09 | 54.63 | True |
| verify-gl-high | 21.89 | 45.7 | 14.13 | 38.34 | True |
| verify-final | 24.84 | 40.2 | 12.81 | 21.33 | True |
| verify-lifecycle | 23.80 | 42.0 | 11.76 | 20.69 | True |
Ablations used 60 warmup / 120 samples at the same initial configuration. They remove whole features to identify costs, not to pass the quality-preserving optimization gate. Water removal includes FFT, foam and optical captures; CPU buoyancy remains. Lighting removal concerns scene lights/shadows and environment, not every manually shaded water/cloud term. Differences overlap and are not additive; host variation obscures the AO result.
Final live cloud-bearing probe audit: 0.653% RMS, 0.236% mean absolute error, 1627 active probes from 4096 scattered positions. Maximum local probe error is 12.571%; the sub-1% image/RMS tests are aggregate instruments, not a guarantee for every pixel.
The labeled camera study.
Representative settled shots and an exhalation-peak capture. The original failed framing shots are retained in the evidence bundle; these corrected views are not a guarantee against every possible transient occlusion.
Verification.
- Automated tests: All 17 numerical, rigging-clearance, surfacing-clearance and layout tests passed.
- TypeScript and production build: Passed; bundle warning retained (1.13 MB JS / 326 KB gzip). No compile errors.
- Latest Three.js: npm registry rechecked: 0.186.1, on 1 October 2026 JST.
- Optimization convergence: Five consecutive distinct, image/numerical-accepted candidates below 5% improvement vs fastest prior accepted incumbent.
- Temporal cloud lighting: Four moving phases evaluated against full shadow integration: maximum phase RMS 0.653%, mean error 0.228%. Low-resolution candidate rejected.
- Wave precision: FFT spot error 1.066e-7; final RGBA16F display RMS 0.0415% versus RGBA32F FFT.
- TSL backend image comparison — high: 7/9 cross-backend views pass the strict <1% gate; worst SSIM loss 1.307%. Both backends render and FFT audits pass; pixel-identical cross-backend parity is not claimed.
- TSL backend image comparison — cine: 5/9 cross-backend views pass the strict <1% gate; worst SSIM loss 1.411%. Both backends render and FFT audits pass; pixel-identical cross-backend parity is not claimed.
- Guided camera tour: All 23 labeled stops reached and captured after settling on WebGPU Cinematic, with no renderer error flag. Water, rigging, whale/tail/eye/breath, pen, clock and compartments are included. Finite captures do not certify zero clipping at every possible animation state.
- Tour correction: Whale-surfacing shot re-framed from the outboard side, breath triggered after camera approach, keel-overlap surfacing suppressed, visible exhalation peak captured. Pen close-up moved clear of nearby furniture.
- Interactive controls: Boarding and all four compartments, pause/resume (wave time held at 187.010), Cinematic/High/Balanced changes and both backend tabs exercised. Weather numeric value survived switching; its restored caption was corrected.
- Post-correction regression: All nine selected-incumbent image comparisons passed after tour/surfacing fixes. Worst SSIM loss 0.1940%. FFT and four-phase cloud audits pass. Final repeat: 36.6 completed FPS Cinematic; earlier repeat 41.7, High 54.5.
- Responsive viewport: 393×852 mobile viewport captured; engine tabs and controls remain available. Override reset to the natural desktop viewport. No renderer error flag after resizing.
- WebGPU resize lifecycle repair: Responsive test exposed stale destroyed transmission textures in r186. Refraction and post-processing targets are now recreated after queue draining. Repeated 393×852, 850×700 and natural desktop transitions passed with no graphics-error flag.
- Mobile header repair: Brand, both engine tabs, labeled Guided Tour and pause controls fit within 393px; status/accounting no longer overlap. Browser viewport override reset.
- Final resize-fix regression: All nine views pass after target lifecycle repair; worst SSIM loss 0.2015%. Full Cinematic repeat 42.0 completed FPS at 1280×720 / DPR1; numerics and graphics valid.
- HTML report and timelapse: HTML report opened and visually checked in the browser; Stopping rule: MET visible. 37 original checkpoints compiled into a 92.5-second captioned MP4. Both report and live scene kept open.
Automated numerical and rigging tests, TypeScript checking and production build diagnostics are retained with the project. The GPU FFT spot audit is recorded inside each benchmark JSON and is compared with independent CPU Fourier evaluation.
The development timelapse.
Original progress screenshots are preserved. The video overlays elapsed time and the latest logged cumulative token counts for each saved screenshot. A checkpoint with no saved image is omitted rather than fabricated.
Checkpoint accounting uses the screenshot file's save timestamp and the last usage event at or before it. Earlier on-page timer bugs do not override this timeline. Counter snapshots can lag the capture; the exact usage-as-of timestamp is retained.
00:24:03 001-bottle-and-ship-blockout.jpg
Saved 2026-09-30T05:59:01.434108+00:00. Usage as of 2026-09-30T05:58:33.676Z.
Input 2,029,489 · cached 1,888,384 · output 24,036 · total 2,053,525.
00:27:56 Wave field, volumetric clouds and reflective glass
Saved 2026-09-30T06:02:54.790772+00:00. Usage as of 2026-09-30T06:02:43.561Z.
Input 3,228,937 · cached 3,065,472 · output 28,315 · total 3,257,252.
00:33:41 Detailed three-masted frigate, rigging and Verlet cloth sails
Saved 2026-09-30T06:08:39.725162+00:00. Usage as of 2026-09-30T06:08:22.377Z.
Input 3,548,768 · cached 3,383,040 · output 35,127 · total 3,583,895.
00:41:23 Cartographer’s study, carved walnut stand, cork, hanging compass and lanterns
Saved 2026-09-30T06:16:21.188695+00:00. Usage as of 2026-09-30T06:16:04.927Z.
Input 4,061,237 · cached 3,879,296 · output 44,312 · total 4,105,549.
00:44:29 Corrected transparent depth layering for visible contained clouds and reduced glass glare
Saved 2026-09-30T06:19:27.607265+00:00. Usage as of 2026-09-30T06:17:16.522Z.
Input 4,419,078 · cached 4,057,472 · output 45,347 · total 4,464,425.
00:52:23 Finite-depth directional spectrum, Gerstner choppiness, planar reflections and coherent self-shadowed clouds
Saved 2026-09-30T06:27:21.124242+00:00. Usage as of 2026-09-30T06:27:09.470Z.
Input 5,171,311 · cached 4,791,040 · output 56,895 · total 5,228,206.
01:15:30 First-person steering and three genuine interior compartments with furnishings
Saved 2026-09-30T06:50:27.937188+00:00. Usage as of 2026-09-30T06:50:09.657Z.
Input 6,208,048 · cached 5,804,544 · output 85,000 · total 6,293,048.
01:31:04 Breathing blue whale, depth-based water refraction, aged dip pen, ticking chronometer and cabin wear
Saved 2026-09-30T07:06:02.683584+00:00. Usage as of 2026-09-30T07:05:31.114Z.
Input 7,894,541 · cached 7,455,232 · output 102,658 · total 7,997,199.
01:56:29 Twenty-one-stop cinematic guided tour with explanatory captions and world-space detail pointers
Saved 2026-09-30T07:31:26.938221+00:00. Usage as of 2026-09-30T07:27:48.283Z.
Input 10,322,974 · cached 9,749,248 · output 113,754 · total 10,436,728.
02:14:30 Laid three-strand hemp with fiber relief, whale tracking shots, and performance baseline preparation
Saved 2026-09-30T07:49:28.044088+00:00. Usage as of 2026-09-30T07:48:52.661Z.
Input 12,416,773 · cached 11,807,360 · output 130,338 · total 12,547,111.
03:04:34 Controlled high-detail benchmark with 360 measured frames, CPU submission timing and GPU timer queries
Saved 2026-09-30T08:39:32.012009+00:00. Usage as of 2026-09-30T08:35:47.966Z.
Input 19,260,261 · cached 18,564,096 · output 169,128 · total 19,429,389.
03:09:15 Storm foam now forms only at breaking compression and persists as sparse advected streaks; benchmark camera isolation fixed
Saved 2026-09-30T08:44:13.316739+00:00. Usage as of 2026-09-30T08:43:28.536Z.
Input 20,383,972 · cached 19,674,496 · output 177,625 · total 20,561,597.
03:54:34 Profiling pass reuse · preserved high-detail geometry and optics
Saved 2026-09-30T09:29:32.060465+00:00. Usage as of 2026-09-30T09:28:45.210Z.
Input 26,398,041 · cached 25,535,616 · output 214,802 · total 26,612,843.
03:59:00 Storm reference correction · 55 m swell, wind ripples, darker water and filament foam
Saved 2026-09-30T09:33:58.443066+00:00. Usage as of 2026-09-30T09:33:35.378Z.
Input 26,779,017 · cached 25,905,408 · output 220,362 · total 26,999,379.
04:12:17 Replacing rounded wave geometry · full 128² GPU wind spectrum and inverse FFT
Saved 2026-09-30T09:47:15.741739+00:00. Usage as of 2026-09-30T09:46:52.166Z.
Input 28,412,070 · cached 27,508,352 · output 235,223 · total 28,647,293.
04:17:08 Inspecting spectral crests · correcting excessive foam coverage
Saved 2026-09-30T09:52:06.759406+00:00. Usage as of 2026-09-30T09:51:04.013Z.
Input 29,278,487 · cached 28,358,400 · output 239,636 · total 29,518,123.
05:10:42 Branding installed · inspecting helm clearances and the refined whale at high detail
Saved 2026-09-30T10:45:40.124337+00:00. Usage as of 2026-09-30T10:45:00.791Z.
Input 36,817,619 · cached 35,809,920 · output 297,285 · total 37,114,904.
05:10:43 Collision-aware ship · capsule-constrained cloth, dry-cabin rain mask and solid-geometry walking
Saved 2026-09-30T10:45:41.288577+00:00. Usage as of 2026-09-30T10:45:00.791Z.
Input 36,817,619 · cached 35,809,920 · output 297,285 · total 37,114,904.
05:23:41 Whale refinement · tapered flexible foil flukes and recessed eyes with soft eyelid folds
Saved 2026-09-30T10:58:39.255520+00:00. Usage as of 2026-09-30T10:57:40.958Z.
Input 38,698,540 · cached 37,672,320 · output 309,751 · total 39,008,291.
05:58:21 Integrated whale folds · subtle geometric recesses and shared cool underwater shading
Saved 2026-09-30T11:33:19.686156+00:00. Usage as of 2026-09-30T11:29:35.454Z.
Input 44,284,912 · cached 43,218,432 · output 339,749 · total 44,624,661.
07:03:18 TSL migration: shared FFT, ocean, clouds, rain, spray and post-processing implemented; testing both engines
Saved 2026-09-30T12:38:16.502766+00:00. Usage as of 2026-09-30T12:38:02.133Z.
Input 54,107,945 · cached 52,627,328 · output 399,618 · total 54,507,563.
07:35:16 Shared TSL migration works on both backends; tab toggle and full-optics Cinematic mode under test
Saved 2026-09-30T13:10:14.333705+00:00. Usage as of 2026-09-30T13:09:52.516Z.
Input 63,447,651 · cached 61,878,272 · output 425,676 · total 63,873,327.
07:59:16 WebGPU/WebGL2 tabs verified; Cinematic reflections and contact shadows working; optimizing completed-frame performance
Saved 2026-09-30T13:34:14.451885+00:00. Usage as of 2026-09-30T13:33:56.227Z.
Input 67,977,020 · cached 66,369,152 · output 442,015 · total 68,419,035.
09:07:07 Cinematic GPU ablation complete; starting quality-gated optimization iterations
Saved 2026-09-30T14:42:05.048757+00:00. Usage as of 2026-09-30T14:40:18.192Z.
Input 83,423,380 · cached 81,733,120 · output 481,411 · total 83,904,791.
09:23:55 Full-quality GPU optimization: SSR early rejection; cloud and water precision candidates
Saved 2026-09-30T14:58:53.800737+00:00. Usage as of 2026-09-30T14:58:26.740Z.
Input 85,071,848 · cached 83,306,240 · output 490,024 · total 85,561,872.
09:25:53 028-cloud-light-cache.jpg
Saved 2026-09-30T15:00:51.651411+00:00. Usage as of 2026-09-30T15:00:35.786Z.
Input 85,667,706 · cached 83,887,232 · output 490,972 · total 86,158,678.
09:30:09 Water display precision verified numerically; testing temporal cloud-light reuse
Saved 2026-09-30T15:05:07.337862+00:00. Usage as of 2026-09-30T15:04:45.351Z.
Input 86,355,534 · cached 84,556,672 · output 494,244 · total 86,849,778.
09:35:59 World-space cloud-light amortization: temporal fidelity audited
Saved 2026-09-30T15:10:57.770529+00:00. Usage as of 2026-09-30T15:09:45.812Z.
Input 87,248,248 · cached 85,439,616 · output 498,501 · total 87,746,749.
09:40:12 Refined cache passes: 0.643% lighting RMS; phase-cached buoyancy test
Saved 2026-09-30T15:15:10.144553+00:00. Usage as of 2026-09-30T15:14:26.527Z.
Input 87,932,445 · cached 86,115,200 · output 502,234 · total 88,434,679.
09:43:54 Cloud audit passed; CPU and transform optimizations under test
Saved 2026-09-30T15:18:52.474749+00:00. Usage as of 2026-09-30T15:17:57.854Z.
Input 88,657,438 · cached 86,830,720 · output 504,689 · total 89,162,127.
09:47:58 GPU and CPU candidates: visual checks remain below 1%
Saved 2026-09-30T15:22:56.862465+00:00. Usage as of 2026-09-30T15:22:06.586Z.
Input 89,569,026 · cached 87,735,552 · output 507,753 · total 90,076,779.
09:56:26 Four consecutive sub-5% candidates; prepacked FFT input is next
Saved 2026-09-30T15:31:24.164875+00:00. Usage as of 2026-09-30T15:30:59.063Z.
Input 92,214,492 · cached 90,177,152 · output 514,315 · total 92,728,807.
10:02:09 Optimization plateau reached; final WebGPU/WebGL2, quality and tour validation
Saved 2026-09-30T15:37:07.459838+00:00. Usage as of 2026-09-30T15:36:17.593Z.
Input 93,680,982 · cached 91,635,200 · output 518,087 · total 94,199,069.
10:10:16 Final mode checks: WebGPU 41.7 FPS Cinematic / 54.5 FPS High; testing WebGL2
Saved 2026-09-30T15:45:14.232517+00:00. Usage as of 2026-09-30T15:44:00.746Z.
Input 95,417,213 · cached 93,359,872 · output 524,021 · total 95,941,234.
10:22:54 Both TSL backends render; checking the 23-shot guided tour and interactive controls
Saved 2026-09-30T15:57:51.960994+00:00. Usage as of 2026-09-30T15:57:03.188Z.
Input 97,886,889 · cached 95,786,368 · output 532,166 · total 98,419,055.
10:37:38 Tour framing repaired; whale waits for keel clearance before surfacing
Saved 2026-09-30T16:12:36.753703+00:00. Usage as of 2026-09-30T16:11:44.682Z.
Input 99,682,971 · cached 97,556,608 · output 544,097 · total 100,227,068.
11:06:14 Final delivery: responsive engine tabs, report, screenshots and accounting
Saved 2026-09-30T16:41:12.572688+00:00. Usage as of 2026-09-30T16:39:54.695Z.
Input 107,162,542 · cached 104,988,544 · output 565,109 · total 107,727,651.
Tokens, time and hypothetical API cost.
Wall time starts at 2026-09-30T05:34:57.889Z. Report accounting was generated at 2026-09-30T17:09:39.810572+00:00, using usage reported through 2026-09-30T17:05:18.100Z. This includes interactive feedback and benchmarking, not only typing or compiler time.
| Token category | Count | Reconstructed cost |
|---|---|---|
| Uncached input | 2,205,683 | $5.53642 |
| Cached input | 108,187,776 | $16.27050 |
| Cache writes | 0 | $0.00000 |
| Output, including reasoning | 583,520 | $7.07538 |
| Model-token total | 110,976,979 | $28.88231 |
| Reasoning subset — not billed again | 354,954 | Included above |
Official OpenAI API pricing, checked against current official documentation (30 September UTC / 1 October JST 2026): Standard short-context rates per million are $2 uncached input, $0.10 cached input, $2.50 cache writes and $10 output. Above 272K input in a request, rates are $4, $0.20, $5 and $15. Rates were applied per unique usage increment, not blindly to an aggregate prompt. 167 of 486 reconstructed requests were long-context; the largest reported input was 424,636 tokens.
This is a hypothetical model-token API equivalent, not a Codex subscription charge or invoice. Cumulative input re-counts context on every request; cached input is included in input totals. Fast mode, regional uplifts and taxes are not assumed. External tool, hosting, hardware and electricity charges are excluded. Unreported in-flight usage and subsequent messages, including the final response, are outside this snapshot. Session metadata identifies gpt-6.1-sol with max effort; service_tier is null, so Standard is an explicit pricing assumption.
References & reproducibility.
- Conor Dart's ship-in-a-bottle post — compositional inspiration.
- Three.js Water Pro and its documentation — inspected visual reference, not purchased or copied source.
- NOAA: Blue whale — anatomy/behavior context.
- Science Museum Group: marine chronometer, V&A: historical writing implements, and National Museum of the Royal Navy — period-informed interpretation.
- Dan Greenheck: cloud amortization — adaptation inspiration; our cache is world-space optical depth, not the post’s screen-space reprojection.
- Dan Greenheck: water texture precision — tested mixed precision; the creator’s speedup is not a claim about this app.
- Official TSL reference and r186 SSR source — portable node graph and zero-contribution ray rejection.
- Toonlab and original Call Me Sensei SVG — requested brand lockup.
npm install · npm run dev -- --port 5173 · npm test · npm run typecheck · npm run build
Selected benchmark: http://127.0.0.1:5173/?renderer=webgpu&quality=cinematic&bench=1&opt=14&iteration=reproduce. Per-screenshot elapsed-time/token ledger · Raw benchmark and quality evidence. Raw measurements and per-view quality JSON live in artifacts/benchmarks.