Crouton Alpha creative review: Blender, knowledge videos, SVG and animation
Crouton Alpha creative review with Blender, knowledge-video, SVG and animation artifacts.
Updated: 2026-11-04
Model and source
Crouton Alpha is an anonymous model listed by OpenRouter. This review ran through OpenCode v2.2.5 using the exact route openrouter/stealth/crouton-alpha. The underlying vendor and fixed checkpoint remain undisclosed.
OpenRouter describes it as a lightweight model for coding, agentic subtasks, tool use, document analysis and multimodal work, with a 1,048,576-token context and zero listed prompt/completion prices; these are provider-listed metadata, not tested results. OpenRouter lists a November 3, 2026 release date. Source: OpenRouter model page
No official benchmark images for this model were found, so none are reproduced here.
Test scope and findings
The suite covers three Blender tasks, four knowledge videos, three static SVG tasks and one SVG animation, using the same frozen topics for the day where the suite requires them. Details of tools and run conditions are recorded in the data.
Static SVGs and the loop animation follow explicit constraints consistently; the parametric bookshelf has correct dimensions but simple materials. The four corrected knowledge videos deliver complete captioned explanations, with remaining schematic and wording limits. Complex Blender tasks show clear failures: the reducer's physical carriers and shafts do not follow the motion and meshing is obscured; the interior has circulation, material and camera-readability problems. Delivery and quality acceptance are recorded separately; author selfchecks do not replace inspection.
3D modelling

Parametric bookshelf and parameter revision
Both variants have correct dimensions,18mm boards and equal openings, with a genuinely open back; wood material is simple.
More views and test details
6m 11s · Requests / complete responses: — / 42 · 1,040,229 / 11,708 input / output tokens
- A single parameter script generates both3- and4-compartment variants, with separate boards and props excluded from measurement.
- Three renders clearly compare width/opening changes and reveal the open back.
- Wood uses flat colour without grain or detailed surface character.
Itemized verification
- Dimensions and compartments · pass
- Reopened variants measure0.8/1.0×0.3×1.2m with18mm boards; three376mm and four277.5mm clear openings.
- Native projects and views · pass
- Both native projects reopen unchanged with no embedded Text or external dependencies; all three1920×1080 renders inspected.
- Material detail · partial
- Panels and edges read clearly, but wood is flat-coloured without grain and the presentation is basic.
Parametric bookshelf: 3- and 4-compartment variants, boards and open back compared.. Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
attempt-1 · default · model_finished · Requests / complete responses: — / 42 Run record
See run record for counts and raw tokens.
Source brief and prompt
参数化书架:3格与4格两版,板件与背面开放结构对比。 / Parametric bookshelf: 3- and 4-compartment variants, boards and open back compared.. 以下为协调者短摘要;完整冻结提示词SHA-256见promptSha256,原文不公开。 Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
Parametric bookshelf: 3- and 4-compartment variants, boards and open back compared.
SHA-256: 2806ab28c5e7dd5437d75e7acc0043239c9cefef44792ad7b01a4de0b74b9aac

24/18/60-tooth planetary reducer and axial explosion
The reducer project, four stills and six-second film are delivered. Sun-to-orbit-pivot ratio is3.5, but physical carrier discs and input/output shafts do not follow the motion.
More views and test details
Music and sound effects, no narration.
6-second turntable film
374m 16s · Requests / complete responses: — / 119 · 5,206,758 / 37,093 input / output tokens
- The same Crouton session preserved and repaired first-version errors: external tooth law reused for the ring, and a world-origin filter leaving planet profiles empty. Final local profiles have nonempty guards; teeth24/18/60,42mm centers and ratio3.5. Nine reopened phases and20 discrete2D phases show no outline crossings, without continuous-contact, load or manufacturing validation.
- A separate coordinator process reopened the unchanged blend with network and autoexec disabled and sampled world transforms at7times: sun540degrees, orbit pivot154.286degrees, planet world rotation-360degrees and ring fixed. Both physical carrier discs and input/output shafts remain at zero rotation; the author pivot-ratio test does not establish complete mechanical linkage.
- The0.08degree half-angle backlash is a nominal display choice, not a globally minimum proof. Tolerances, stresses and manufacturability are untested.
- The coordinator viewed all four actual stills and all three film contact sheets. Alt/clay expose a contact region, but the opaque front disc hides most internals in the main view and film; materials and lighting remain pale. Visible partial motion does not pass clear meshing presentation.
- The blend reopens with autoexec disabled, unchanged hash and no embedded text/external images. UV layers and LOD meshes exist; complete UV overlap/density and all-LOD visual acceptance were not performed.
- 1280x720,24fps,144frames,6s, full decode and objective audio checks pass; no listening. One coordinator-interrupted model message has no token record, so usage is a known subtotal, not a complete total.
Itemized verification
- Project and nominal kinematics · partial
- Sun/orbit-pivot ratio3.500000007,planet centersR42 and world-spin relation pass; complete linkage has separate failures.
- Physical carrier and shaft linkage · fail
- Seven reopened samples show physical front/back carrier discs and input/output shafts do not rotate; the animated empty pivot does not establish these bodies following motion.
- Materials and framing · partial
- Four stills delivered, with flat pale lighting/materials and main-view front-disc occlusion.
- Film meshing visibility · fail
- Playable6s film, but an opaque front disc obscures most meshing and prevents a clear full-mechanism demonstration.
- Video technical checks · pass
- 720p24,144frames,6s, full A/V decode.
- UV and LOD · partial
- UV layers and reduced LOD meshes are present; full overlap/density and all-LOD image quality are unverified.
- Manufacturing and listening · not-tested
- No manufacturing/load validation or listening performed.
24/18/60-tooth planetary reducer: assembly, four stills and a 6s turntable film.. Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
attempt-1 · default · model_finished · Requests / complete responses: — / 119 Run record
See run record for counts and raw tokens.
Source brief and prompt
24/18/60齿行星减速器:装配、四张静帧与6秒转台动画。 / 24/18/60-tooth planetary reducer: assembly, four stills and a 6s turntable film.. 以下为协调者短摘要;完整冻结提示词SHA-256见promptSha256,原文不公开。 Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
24/18/60-tooth planetary reducer: assembly, four stills and a 6s turntable film.
SHA-256: 66ed7b5df9169be945cfdc9ff33ce228e29f6c561c5218efb147a32aaf9f4dcb

12×9m two-bedroom interior, rooms and floor plan
The 12×9m two-bedroom project, six real stills and a 6s 144-frame film are delivered; the top view communicates the main rooms and furniture relationships. Circulation, materials, clay contrast and camera readability fall short: wardrobe circulation is not established, the clay close-up fails, and the master ending has no full reveal.
More views and test details
Music and sound effects, no narration.
6-second turntable film
438m 9s · Requests / complete responses: — / 164 · 8,955,166 / 73,304 input / output tokens
- Delivery is recorded separately from quality: the project, six real stills and a 6s film with procedural music/SFX are delivered, which does not certify quality.
- Wardrobe circulation FAIL: a nightstand was moved into the wardrobe zone; the 0.56m side strip ends near a vanity with about 50mm turning clearance.
- Clay close-up contrast FAIL; materials PARTIAL; bedroom framing is tight and obscures spatial relationships.
- Kitchen and master labels are occluded by geometry; the master ending is dominated by wall and bed crops with no full reveal — FAIL.
- Decode, frame-count and objective audio checks pass; they do not prove hearing, quality, or manufacturing/load validation. No listening review.
- Actual pipeline is offline EEVEE with procedural assets at frozen resolutions, NOT the original Cycles/OptiX 4K/HDRI reproduction.
Itemized verification
- Delivery · partial
- Project, six real stills and a 6s 144-frame film delivered; top view is legible.
- Wardrobe circulation · fail
- 0.56m side strip ends near the vanity with ~50mm turning clearance; nightstand moved into the wardrobe zone.
- Clay close-up contrast · fail
- Low contrast; contact relationships are not clearly presented.
- Materials · partial
- Materials and lighting remain coarse; bedroom framing is tight.
- Labels and ending reveal · fail
- Kitchen and master labels occluded; master ending has no full reveal.
- Technical and listening · not-tested
- Decode and objective audio checks pass; no listening review, which does not prove hearing or quality.
12×9m two-bedroom interior: rooms, floor plan and turntable film.. Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
attempt-1 · default · model_finished · Requests / complete responses: — / 164 Run record
See run record for counts and raw tokens.
Source brief and prompt
12×9米两居室:多房间装修、俯视平面与转台动画。 / 12×9m two-bedroom interior: rooms, floor plan and turntable film.. 以下为协调者短摘要;完整冻结提示词SHA-256见promptSha256,原文不公开。 Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
12×9m two-bedroom interior: rooms, floor plan and turntable film.
SHA-256: 953ab0d133ce8cd9b6b48e39545a4084d1ae00ddab85993e5330facc9cdce582
Knowledge videos
Why does water float after it freezes?
The sequence links molecular arrangement, volume and density; mass is conserved, freezing expands volume by about 9%, and ice remains mostly submerged.
Test details
22m 50s · Requests / complete responses: — / 50 · 1,561,038 / 20,154 input / output tokens
- Hydrogen bonds appear in both states, with mass and volume changes clearly distinguished.
- Captions are burned into the film, legible in Chinese and clear of mechanism labels.
- The on-screen “density = mass of equal volume” is a colloquial equal-volume comparison, not a rigorous mass/volume formula.
- Schematics are not exact proportions and should not be measured as real ratios.
Itemized verification
- Duration and decode · pass
- 69.3s, 1920×1080, full decode passes.
- Verified facts · pass
- Mass conserved, lower ice density, hydrogen bonds, ~9% expansion, mostly submerged.
- Burned captions · pass
- Burned Chinese captions legible, no tofu or overlap.
- Listening check · not-tested
- Listening not performed; pronunciation not certified.
Why ice floats: molecular arrangement, hydrogen bonds, volume and density.. Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
attempt-1 · default · model_finished · Requests / complete responses: — / 50 Run record
See run record for counts and raw tokens.
Source brief and prompt
水结冰为何上浮:分子排列、氢键、体积与密度因果关系。 / Why ice floats: molecular arrangement, hydrogen bonds, volume and density.. 以下为协调者短摘要;完整冻结提示词SHA-256见promptSha256,原文不公开。 Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
Why ice floats: molecular arrangement, hydrogen bonds, volume and density.
SHA-256: 2a44ec391079b06132503e2ac5d7a26a8e168483c99b2262a7b960784a651075
Why can a damaged QR code still be read?
Color separates data, error-correction information and damaged codewords: 50 of 200 total codewords means correctable damaged codewords, and percentages apply to total codewords, not image area. Capacity and size tradeoffs are separated by fixed conditions; the grids are schematic, not a real scanning test.
Test details
29m 12s · Requests / complete responses: — / 131 · 5,578,408 / 32,191 input / output tokens
- 50 correctable damaged codewords are correctly distinguished within the 200 total, and percentages are not rewritten as arbitrary covered image area.
- Arbitrary damage patterns and functional-pattern loss are explicitly not guaranteed recoverable; capacity diagrams are qualitative.
- Burned Chinese captions are readable and separate from mechanism labels.
Itemized verification
- Duration and decode · pass
- 85.013s, 1920×1080, 30fps, full decode passes.
- Error-correction facts · pass
- 200 total / 50 correctable-damaged distinction correct; percentages apply to total codewords, capacity conditions separated.
- Schematic limits · pass
- Grids are schematic, not a real scan; arbitrary damage and functional-pattern loss are not guaranteed recoverable.
- Burned captions · pass
- Burned Chinese captions readable, separate from mechanism labels, no tofu.
- Listening check · not-tested
- Listening not performed; pronunciation not certified.
Why damaged QR codes can still be read: data and error-correction codewords, L/M/Q/H levels and capacity cost.. Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
DENSO WAVE / QRcode.com error correction
attempt-1 · default · model_finished · Requests / complete responses: — / 131 Run record
See run record for counts and raw tokens.
Source brief and prompt
二维码为何能纠错恢复:数据与纠错码字、L/M/Q/H四档与容量代价。 / Why damaged QR codes can still be read: data and error-correction codewords, L/M/Q/H levels and capacity cost.. 以下为协调者短摘要;完整冻结提示词SHA-256见promptSha256,原文不公开。 Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
Why damaged QR codes can still be read: data and error-correction codewords, L/M/Q/H levels and capacity cost.
SHA-256: 8cd0bf84eda56320b44c501cdda2c4384a32d02b13dea8915ee626a5199ea5ce
Why does Webb use five thin sunshield layers?
Original animation shows five films, hot and cold sides, vacuum insulation and sideways radiation. The thick-plate comparison uses bottom-to-top heat flow without clearly declaring its changed orientation relative to other shots, weakening directional continuity.
Test details
369m 27s · Requests / complete responses: — / 67 · 2,137,482 / 18,525 input / output tokens
- 84.5s video stream; container 84.6s including audio tail; 1920×1080 at 30fps and full A/V decode pass.
- The coordinator viewed all three encoded contact sheets covering ten moments, head/tail, captions and core mechanisms. Materials and layer tradeoffs follow NASA; diagrams are qualitative, without thermal simulation.
- The thick-plate orientation change is insufficiently explained and some small inter-layer labels have weak contrast; author selfchecks do not establish universal visual correctness.
- Narration uses the shared Kokoro production tool, not native model speech. No listening review; pronunciation, sound quality and complete semantic sync are not human-certified.
Itemized verification
- Duration and decode · pass
- 84.6s container, 1080p30, full decode passes; tail samples use video-stream duration rather than the AAC tail.
- Core source facts · pass
- Separated films, lateral heat radiation, Kapton/aluminium/sun-facing silicon on the hottest two layers and 2-7-layer modelling agree with NASA.
- Direction and label clarity · partial
- Bottom-to-top flow in the plate panel lacks a declared orientation change from the other shots' hot-side-above convention; some small labels have weak contrast.
- Caption timing · pass
- All 13 SRT cues lie within the film without overlap; burned Chinese captions are visible in samples.
- Listening · not-tested
- Objective audio checks only; listening not performed.
Why Webb's five-layer sunshield beats a thick plate: separated films, hot/cold sides and lateral radiation.. Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
attempt-1 · default · model_finished · Requests / complete responses: — / 67 Run record
See run record for counts and raw tokens.
Source brief and prompt
韦布五层遮阳膜为何胜过厚板:分离薄膜、热侧冷侧与侧向辐射。 / Why Webb's five-layer sunshield beats a thick plate: separated films, hot/cold sides and lateral radiation.. 以下为协调者短摘要;完整冻结提示词SHA-256见promptSha256,原文不公开。 Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
Why Webb's five-layer sunshield beats a thick plate: separated films, hot/cold sides and lateral radiation.
SHA-256: afe56d66e496b52002528283fc818547d45737717c6928282ab6483b9434b81c
Why does noise cancellation work without music?
Passive isolation, the microphone-processor-speaker path and three waveform traces explain noise cancellation, retaining voice and seal limits; the illustration is a local-at-ear amplitude-mismatch schematic, not device performance measurement.
Test details
402m 2s · Requests / complete responses: — / 61 · 2,607,373 / 21,745 input / output tokens
- Passive isolation and the active signal path are separately explained; the actual signal sums are correct, and the residual shown is local-at-ear amplitude mismatch, not headphone device performance.
- All five crossfades and caption-timing/fade technical checks pass.
- The line near 70.7s, 'even in the idealized cancellation there is residual sound', is imprecise: the visible simulation is amplitude mismatch; do not claim perfectly equal-amplitude opposite waves retain residual.
- Fourteen captions follow real Kokoro synthesis chunks (not word-level); narration is disclosed 1.15x speedup from Kokoro zf_xiaobei, not native model voice. No listening review; pronunciation not certified.
Itemized verification
- Duration and decode · pass
- 79.59s, 1920×1080, 30fps, full decode passes.
- Cancellation mechanism · pass
- Passive and active paths separated; three-waveform sums correct as local-at-ear amplitude-mismatch illustration.
- Crossfades and caption timing · pass
- Five crossfades and 14-caption timing checks pass; captions follow synthesis chunks, not word-level.
- Wording precision · partial
- 'Idealized cancellation retains residual' is too loose; visible simulation is amplitude mismatch, not proof equal-amplitude opposite waves retain residual.
- Listening check · not-tested
- No listening review; Kokoro zf_xiaobei at 1.15x, not native model voice; pronunciation not certified.
Why noise-cancelling headphones work without music: passive isolation, active path and three waveforms.. Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
Bose how noise cancelling works
attempt-1 · default · model_finished · Requests / complete responses: — / 61 Run record
See run record for counts and raw tokens.
Source brief and prompt
降噪耳机为何无音乐也安静:被动隔音与主动降噪路径、三路波形。 / Why noise-cancelling headphones work without music: passive isolation, active path and three waveforms.. 以下为协调者短摘要;完整冻结提示词SHA-256见promptSha256,原文不公开。 Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
Why noise-cancelling headphones work without music: passive isolation, active path and three waveforms.
SHA-256: 617856550a5c1c23b6cfbe49c52b706987baa3c138aa6ece1d84a3d8889eeb12
SVG and animation

An analog clock at 9:41
Numerals 1-12 are complete and correctly placed; the short hour hand sits between 9 and 10, the long minute hand points just after 8 with clear separation, and the 09:41 label is below the dial.
Test details
1m 36s · Requests / complete responses: — / 12 · 220,073 / 4,534 input / output tokens
- All twelve numerals are complete, correctly placed and unobscured.
- Hand angles are approximately correct (hour ~290.5°, minute ~246°); the minute tip clears numeral 8.
Itemized verification
- Numerals · pass
- 1-12 complete, ordered and correctly positioned.
- Hand angles · pass
- Hour ~290.5° and minute ~246° match 9:41.
- Minute/8 clearance · pass
- Minute tip visibly clears numeral 8.
Analog clock at 9:41: static SVG with numerals and hour/minute hands.. Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
attempt-1 · default · model_finished · Requests / complete responses: — / 12 Run record
See run record for counts and raw tokens.
Source brief and prompt
9:41模拟钟表:静态SVG,含数字与时分针。 / Analog clock at 9:41: static SVG with numerals and hour/minute hands.. 以下为协调者短摘要;完整冻结提示词SHA-256见promptSha256,原文不公开。 Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
Analog clock at 9:41: static SVG with numerals and hour/minute hands.
SHA-256: fbf8775af76d1ad3fa691c4db26a4a1c35c0fca6768fbfbf860d6bb759a04f86

Data chart and proportional bars
All four bars share the zero baseline and scale correctly for 12, 24, 18 and 30. Values, quarter order, total 84 and best quarter Q4 are legible, with ample whitespace and no visible occlusion.
Test details
5m 18s · Requests / complete responses: — / 21 · 460,465 / 6,684 input / output tokens
- All four bars share the zero baseline and scale correctly for 12, 24, 18 and 30. Values, quarter order, total 84 and best quarter Q4 are legible, with ample whitespace and no visible occlusion.
Itemized verification
- Bar proportions · pass
- Shared zero baseline; heights in ratio 2:4:3:5.
- Labels and total · pass
- Q1-Q4 order, total 84 and best Q4 are legible.
Data chart: quarterly revenue bars with proportional heights and total label.. Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
attempt-1 · default · model_finished · Requests / complete responses: — / 21 Run record
See run record for counts and raw tokens.
Source brief and prompt
数据图表:季度营收柱状图,柱高比例与总计标注。 / Data chart: quarterly revenue bars with proportional heights and total label.. 以下为协调者短摘要;完整冻结提示词SHA-256见promptSha256,原文不公开。 Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
Data chart: quarterly revenue bars with proportional heights and total label.
SHA-256: 75c045108ea59d8967a50b16c6d3658f4f6aa13a3abdab8eeea6b0e2597988e8

Chinese poster and layout constraints
The pale green layout has clear text hierarchy. All six required strings are readable, the bicycle sits between the event details and button, and no missing glyphs, text overlap or clipping are visible.
Test details
5m 18s · Requests / complete responses: — / 17 · 375,874 / 5,654 input / output tokens
- The pale green layout has clear text hierarchy. All six required strings are readable, the bicycle sits between the event details and button, and no missing glyphs, text overlap or clipping are visible.
Itemized verification
- Six strings · pass
- All six required strings appear once and read clearly.
- Hierarchy · pass
- Clear hierarchy, no missing glyphs or overlap.
Chinese poster: cycling event details with six text lines and layout hierarchy.. Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
attempt-1 · default · model_finished · Requests / complete responses: — / 17 Run record
See run record for counts and raw tokens.
Source brief and prompt
中文海报:骑行活动信息,六段文字与版式层次。 / Chinese poster: cycling event details with six text lines and layout hierarchy.. 以下为协调者短摘要;完整冻结提示词SHA-256见promptSha256,原文不公开。 Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
Chinese poster: cycling event details with six text lines and layout hierarchy.
SHA-256: 1bbcb03701d9aba48994231b1a1d16582aec35f580572eae9b8b19aae6bb105f

Animated cycling pelican
The long bill and orange throat pouch are recognizable, with a connected two-wheel bicycle, saddle and handlebar. Spokes, crank and both legs move during the four-second loop; in sampled phases and unpaused playback the feet stay on the pedal region and the legs attach beneath the body.
Test details
5m 18s · Requests / complete responses: — / 31 · 776,692 / 10,053 input / output tokens
- The long bill and orange throat pouch are recognizable, with a connected two-wheel bicycle, saddle and handlebar. Spokes, crank and both legs move during the four-second loop; in sampled phases and unpaused playback the feet stay on the pedal region and the legs attach beneath the body. The bird, bill and tyres stay inside the canvas. The loop looks continuous, with a tiny disclosed nonzero boundary pixel difference.
Itemized verification
- Loop animation · pass
- 40 unique frames; 4s cycle moves spokes, crank and legs with a continuous loop.
- Riding pose · pass
- Feet stay near pedals, legs attach beneath body, nothing leaves canvas.
Cycling pelican animation: 4-second SMIL loop with spokes, crank and legs.. Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
attempt-1 · default · model_finished · Requests / complete responses: — / 31 Run record
See run record for counts and raw tokens.
Source brief and prompt
鹈鹕骑车动画:4秒循环SMIL,轮辐、曲柄与腿部运动。 / Cycling pelican animation: 4-second SMIL loop with spokes, crank and legs.. 以下为协调者短摘要;完整冻结提示词SHA-256见promptSha256,原文不公开。 Short coordinator summary; full frozen prompt hash preserved in promptSha256, original text not published.
Cycling pelican animation: 4-second SMIL loop with spokes, crank and legs.
SHA-256: 3e0f1b0386fa1155b0d0655e533fe37aaa49de781658ee1fa4f51fec0678b6fc
Download all creative results and source index
Run settings, usage and full notes
openrouter/stealth/crouton-alpha · 2026-11-04
Artworks, project sources and narration scripts were authored by the exact Crouton in native OpenCode v2.2.5 persistent per-case sessions via route openrouter/stealth/crouton-alpha. Codex coordinators inspected pixels and technical QA and sometimes executed the exact author-supplied foreground render commands in persistent controller jobs; coordinators made no artistic fixes and no other creative models were used. QA = author self-checks plus coordinator AI inspection and separate-process measurements, NOT blind or independent human scores, and NOT all inside one Crouton session. The environment is the harness default; only specified read-only inspection/capture tools deny networking, with no whole-offline claim. Long wall time includes productive, model, render, queue, idle/resume and coordinator-QA time, not model speed. Frozen scope 3 Blender + 4 knowledge videos + 3 static SVG + 1 SVG animation, date 2026-11-04, topic SHA from frozen batch; shared Pillow/NumPy/FFmpeg/browser SVG capabilities and the common Kokoro TTS production tool (scripts model-authored, voice NOT model-native; knowledge-4 speedup disclosed). Native completedResponses are assistant step_finish counts, NOT actual provider requests; the clock reuses the existing Build 12 steps, the successful preflight Plan and external publication work are excluded, and the local trial receipt is retained.
OpenRouter officially lists zero input/output prices; the actual invoice is unverified, and no cost is inferred from normalized counters.
Counts are native assistant step_finish events (NOT provider requests); outputs are visible output plus separately reported reasoning counted once; missing upstream counters are normalized to 0, not raw billing; the clock uses the Build 12-step scope with preflight Plan, publication and coordinator usage excluded and the local trial receipt preserved. Aggregate KNOWN subtotals (OpenCode normalized harness counters, not raw invoices): input 28,919,558, output 241,645 (visible output 238,506 + separately reported reasoning 3,139); complete totals UNKNOWN, actual provider request count UNKNOWN, invoice UNKNOWN. Some cases' wallMs includes long idle/resume/queue time and coordinator QA, never model inference speed.
Visual commentary by AI; dimensions and media checks use scripts.
Requests / complete responses: — / 715 · Errors / interruptions: 1 / —
— / — input / output tokens, including cache and thinking
- Anonymous vendor and fixed checkpoint undisclosed; no official benchmark image/score reproduced.
- Shared fixed Kokoro zf_xiaobei is a production tool, not native model speech; ANC 1.15x and segment-level captions disclosed; no listening or human score.
- Known normalized assistant-step counters are not raw provider requests or invoices; one interrupted reducer message has null usage, so full token totals unknown. Clock Build 12-step scope excludes original preflight Plan; publication/coordinator usage excluded.
- Frozen same-day topics/spec hashes retained; original pipeline adaptations and shared host resource contention disclosed; wall time not inference speed.
- Mechanical linkage, wardrobe route, film occlusion, gray contrast and scientific wording partial/fail findings retained; no manufacturing/load or universal UV/LOD certification.









