Kimi K3.1 Pro creative review: Blender modeling and knowledge videos

Kimi K3.1 Pro creates editable Blender scenes and knowledge videos, with inline playback, music by default, optional narration, source downloads, failures and usage evidence.

Updated: 2026-10-27

Read the model review: access, pricing, text, code and SVG tests

TOKRACE original creative evaluation · 2026-10-27

The delivered work, with its evidence

Ten tasks deliver six Blender scenes and four knowledge videos. The parametric bookshelf and reducer do well on core geometry/kinematics, while complex buildings, mechanical assets and explainers retain dimension, contact, UV or scientific-trajectory failures. Work, usage, failures and independent AI findings are reported per case; a playable file is not a content pass.

Creative tasks: 10
Requests dispatched: 282
Complete responses: 272
Request errors / interruptions: 8 / 2
Received input tokens, incl. cache: 9,044,253
Received output tokens, incl. thinking: 346,996

Token totals cover received final usage. Failed or interrupted requests can consume additional quota; their final usage is unknown. These tasks do not establish a general model ranking.

Model, workflow and billing

Kimi K3.1 Pro · k3.1-pro · Kimi Coding Plan

Run on 2026-10-27 through the native Kimi Coding Plan Messages endpoint; all 272 complete responses return k3.1-pro. Initial four: bookshelf/email/printer request max; Jupiter retains max failures at 600/900-second deadlines and a separate high recovery at 900 seconds. Six additions request high with 60 minutes/50 turns, 600 seconds per request and max_tokens=32768 including thinking, using isolated offline local tools. Four fixed-source BlenderWorld tasks use Blender 5.2.3 EEVEE/two threads and adapted output budgets, not official benchmark scores. Two Opus 5.5 video briefs are shortened into original Pillow/NumPy/FFmpeg films, without reusing Manim or creator outputs. Kimi writes/revises code, runs shell and inspects images; independent checks follow. The batches differ and share concurrent hardware, so they are not controlled effort or speed rankings.

Requests consume Kimi Coding subscription quota; no reconciled per-task monetary invoice is available, so actual cost is unknown. Usage covers complete responses with final counters only; extra failed/interrupted consumption is unknown. Public narration replacement is a separate publishing step, not part of the original Kimi score.

An autonomous agent session includes many model calls, tool runs and self-inspections. A high-effort recovery is labeled separately from failed max-effort attempts.

01 · Blender · Requested effort: high

24/18/60-tooth planetary reducer and axial explosion

Independent samples verify tooth counts, dimensions, involute profiles, coaxial explosion and analytic kinematics. Carrier/shaft UVs overlap and fit tolerances remain unverified; this is not a production- or engine-ready asset.

Download MP4 · Playback starts only when you press play.

Six-second analytic mechanism motion; not dynamics

Task wall time: 56m 0s
Requests / complete responses: 38 / 37
Received input tokens, incl. cache: 1,307,323
Received output tokens, incl. thinking: 38,117

AI assistant observations

Measured 24/3×18/60 teeth, module about 2mm, 20mm face width, approximately 20° involute profiles and a 42mm/120° planet layout.
Forty-two sampled BVH surface-pair checks over seven times find no intersections; 22 exploded components move only axially. This is not contact-stress or load certification.
All 144 frame matrices give sun/carrier ≈3.5, a fixed ring and planet/carrier absolute ratio ≈−7/3: analytic mechanism motion rather than a camera orbit or dynamics.
Evaluated LOD0/1/2 counts are 160930/22120/9218 triangles. Carrier/shaft UVs each have at least 25 positive-area overlaps. Fit tolerances are unverified, machining microtexture is absent and the alternate view crops the shaft top.
The public project only removes an embedded source-attachment text; geometry/materials/cameras/actions/code remain unchanged. Original/public hashes and reopened structural fingerprints are downloadable.

These are qualitative AI observations, not independent human scores. Automated checks are listed separately.

Original case and evaluation changes

Published task and specification

Below are TOKRACE execution additions and a source summary, not the full third-party prompt. Offline independent generation with disclosed protocol changes; no original output reuse or official benchmark score.

The linked creator or official example is a source. The output above is our independently generated run.

Itemized verification

Geometry and analytic motion· Pass: Independent native geometry, all-frame matrices, seven-time BVH and explosion checks.
UV and production readiness· Fail: UV intersections confirmed; manufacturing fits and engine performance unverified. LOD presence is not a blanket asset pass.

Source brief and public rerun specification

Source excerpt or summary; follow the link for the original

来源概要：按公开机械规格制作单级行星减速器、可检查的装配与同轴爆炸。下面给出本次适配规则；不转载原始附件长文。
Source summary: a specified single-stage planetary reducer with inspectable assembly and coaxial explosion; full source attachments are not republished.

Archived source prompt SHA-256: 4b2e0fcc930a262160e2804f8bfbeadc8dc98c1df6f80f8f2a8a64e5133104dd

Public specification, with third-party wording omitted

以下为TOKRACE新增执行条件，并非原benchmark协议：本轮离线，仅本题公开输入，不访问公网或外部模型/资产库；k3.1-pro high，60分钟/50次请求/150000输出token，2线程。使用本机Blender5.2.3、EEVEE。原题的Cycles、OptiX、HDRI下载、采样/分辨率要求改为本次统一协议，必须在README披露；不声称官方BlenderWorld分数或严格复现原渲染管线。完整保留几何结构、尺寸、数量、配合、材质区分及原始UV/LOD等资产需求，不能为填任务数量简化；缺项如实列出。
先分功能块落盘完整可运行源代码，然后实际后台工厂启动建模/渲染，不使用单次巨型响应。交付final.blend、完整源码/资源、1920×1080主图+两张1600×900不同机位+结构/中性灰模近景；另交6秒1280×720/24fps的turntable.mp4（144帧）。减速器/蒸汽机可用机构运动代转台，但必须明确是运动学，非受力仿真，保留静态检查。先真实预览、view_image检查，最多两轮有效视觉修正，程序报错另记；保留首个预览与失败。视频检查至少6时点与首尾。独立进程重开工程，交artifacts.json/README（原题差异、实际命令版本、未完成项）/真实尺寸与对象清单自检JSON。离线程序材料替代外部图片/HDRI时必须说明；不得用平面或贴图代关键实体。

本题新增可测规格：读取asset_budget.json，全部几何/齿数/名义配合/UV/LOD要求保留。两张不同机位至少一张为同轴爆炸图，灰模近景看清太阳—行星—内圈接触。未实际校核的制造公差只能称名义展示。若运动，保持理论运动学关系，非动力学。

公开说明：这是评测方新增约束及来源概要，原始完整输入本地保留并用SHA-256绑定。原输入/附件请从来源固定链接取回；不是完整实际输入逐字副本。原本地系统配音仅作评测证据，公开版由Kokoro替换旁白，未修模型画面。
Public note: these are evaluator additions and a source-summary rerun guide, not the full verbatim input. Full local inputs are hash-bound and source attachments must be retrieved separately. Public Kokoro narration replacement does not repair original imagery.

This is the evaluation adaptation and a rerun guide, not a verbatim transcript of the full original request. Exact original inputs remain archived locally.

Original evaluation input SHA-256: 6acb4fbd03f76f248c5ed82ba658b89305605037b2c35817f1e702d31bbfeba0

Attempts, failures and recovery · 1

attempt-01 · high · Model finished

Requests / complete responses: 38 / 37

Autonomous high session finished with 1 request error(s) and 1 runtime traceback record(s). Model edits, previews and re-renders are included; this was not a single call.

Attempt evidence

02 · Blender · Requested effort: high

Horizontal steam engine, flywheel and Watt governor

Stroke, rod, crank and 3:1 analytic values hold, but the piston leaves the cylinder and static lower governor links do not follow their balls. LODs and twelve 4K textures exist; UV quality is only partly verified. Whole-mechanism motion and asset quality do not fully pass.

Download MP4 · Playback starts only when you press play.

Six-second analytic mechanism motion; not dynamics

Task wall time: 50m 3s
Requests / complete responses: 30 / 29
Received input tokens, incl. cache: 880,453
Received output tokens, incl. thinking: 35,172

AI assistant observations

Independent checks over 144 frames confirm a 300mm stroke, 600mm rod, 150mm crank and 3:1 analytic transmission. This is kinematics, not steam, force or control-system simulation.
The piston extends up to 145.0mm beyond the cylinder barrel, or 102.5mm beyond the envelope including end flanges. In 33 of 144 frames the whole piston lies outside that envelope. Correct stroke values do not establish correct containment. Both lower governor links stay static; even their nearest endpoints miss the matching ball centres by 28.31–198.65mm.
Actual LOD triangle counts are 115974/29710/9765 and twelve 4096×4096 textures are packed. Image resolution alone does not validate UV or production material quality.
Median UV density is about 32768px/m, differing from the self-reported 512 with variation within faces. The source gives 512 only as an example, not a unique mandatory target; UV quality is only partly verified. Full-asset UV non-overlap is unverified. Bright clay and cropped views limit inspection; the video uses a fixed camera.
M12/M16-labelled unthreaded shanks measure about 9.600/11.520mm; the corresponding bolt fits are not verified.

These are qualitative AI observations, not independent human scores. Automated checks are listed separately.

Original case and evaluation changes

Published task and specification

The linked creator or official example is a source. The output above is our independently generated run.

Itemized verification

Core dimensions and analytic values· Pass: All-frame stroke/rod/crank/transmission values hold; this is not a whole-mechanism or dynamics pass.
Piston and governor connections· Fail: Piston extends 102.5mm beyond the flange-inclusive envelope and lies wholly outside in 33 frames; static lower governor-link endpoints miss ball centres by 28.31–198.65mm.
LOD, 4K and UV· Partial: LODs and 4K files exist; UV density varies, contradicts the self-report, and full-asset non-overlap is unverified.

Source brief and public rerun specification

Source excerpt or summary; follow the link for the original

来源概要：有完整机械规格的复古卧式蒸汽机，包含曲柄连杆、飞轮、Watt调速器、管路与资产预算。
Source summary: a specified horizontal steam engine with crank linkage, flywheel, Watt governor, piping and asset budgets.

Archived source prompt SHA-256: 46f733406de20668bf8ab71825ef43e931d42ca8d62946befaae89be3ae97ff4

Public specification, with third-party wording omitted

以下为TOKRACE新增执行条件，并非原benchmark协议：本轮离线，仅本题公开输入，不访问公网或外部模型/资产库；k3.1-pro high，60分钟/50次请求/150000输出token，2线程。使用本机Blender5.2.3、EEVEE。原题的Cycles、OptiX、HDRI下载、采样/分辨率要求改为本次统一协议，必须在README披露；不声称官方BlenderWorld分数或严格复现原渲染管线。完整保留几何结构、尺寸、数量、配合、材质区分及原始UV/LOD等资产需求，不能为填任务数量简化；缺项如实列出。
先分功能块落盘完整可运行源代码，然后实际后台工厂启动建模/渲染，不使用单次巨型响应。交付final.blend、完整源码/资源、1920×1080主图+两张1600×900不同机位+结构/中性灰模近景；另交6秒1280×720/24fps的turntable.mp4（144帧）。减速器/蒸汽机可用机构运动代转台，但必须明确是运动学，非受力仿真，保留静态检查。先真实预览、view_image检查，最多两轮有效视觉修正，程序报错另记；保留首个预览与失败。视频检查至少6时点与首尾。独立进程重开工程，交artifacts.json/README（原题差异、实际命令版本、未完成项）/真实尺寸与对象清单自检JSON。离线程序材料替代外部图片/HDRI时必须说明；不得用平面或贴图代关键实体。

本题新增可测规格：完整读取brief.md，底座、缸径行程、600mm连杆、1200mm飞轮的6条曲辐、离心调速器、滑阀机构和管路等原规格不缩水。4K材质贴图、UV、SM_命名、hero/LOD1/LOD2预算与运动副轴心需求保留；未完成或未烘焙须公开。不得只用彩色圆柱替代机械关系。增加背面、另一侧与中性灰模近景，观察汽缸—活塞杆—十字头—连杆—曲柄连接。视频可为转台或运动学机构；后者保持300mm行程、150mm曲柄、600mm杆长，不称蒸汽或受力仿真。

公开说明：这是评测方新增约束及来源概要，原始完整输入本地保留并用SHA-256绑定。原输入/附件请从来源固定链接取回；不是完整实际输入逐字副本。原本地系统配音仅作评测证据，公开版由Kokoro替换旁白，未修模型画面。
Public note: these are evaluator additions and a source-summary rerun guide, not the full verbatim input. Full local inputs are hash-bound and source attachments must be retrieved separately. Public Kokoro narration replacement does not repair original imagery.

This is the evaluation adaptation and a rerun guide, not a verbatim transcript of the full original request. Exact original inputs remain archived locally.

Original evaluation input SHA-256: d13d8511ebab01c1b79a9d657b3a611c41d62ca85bc611c2069469b7747e1064

Attempts, failures and recovery · 1

attempt-01 · high · Model finished

Requests / complete responses: 30 / 29

Autonomous high session finished with 1 request error(s) and 3 runtime traceback record(s). Model edits, previews and re-renders are included; this was not a single call.

Attempt evidence

03 · Blender · Requested effort: high

12×9m two-bedroom interior, rooms and floor plan

Five stills, a six-second video and an editable project are delivered, but wall bounds are 12.24×9.24m, the second bedroom is oversized, and chairs/bed float above the floor. Delivery is not a layout or scale pass.

Download MP4 · Playback starts only when you press play.

Six-second real camera orbit around the same scene

Task wall time: 29m 51s
Requests / complete responses: 25 / 25
Received input tokens, incl. cache: 688,950
Received output tokens, incl. thinking: 27,942

AI assistant observations

12×9m is the floor slab; actual external wall bounds are 12.24×9.24m, differing from the exterior requirement.
The second-bedroom/study footprint is about 19.89m² versus the source’s approximate 11m² target; approximate source areas are not treated as exact architectural truth.
Independent floor rays measure 25–49mm chair-leg gaps and 66–90mm bed-platform gaps.
The project reopens and contains five stills plus a 144-frame, six-second video. Bedroom framing is tight; clay/orbit views are dark. UV quality was not checked, and LOD/4K textures are not explicit requirements for this case.

These are qualitative AI observations, not independent human scores. Automated checks are listed separately.

Original case and evaluation changes

Published task and specification

The linked creator or official example is a source. The output above is our independently generated run.

Itemized verification

Bounds and room scale· Fail: Wall bounds are 12.24×9.24m and the second room is about 19.89m², differing from the brief.
Furniture contact and presentation· Fail: Measured chair/bed floor gaps; some views are dark or tightly cropped.

Source brief and public rerun specification

Source excerpt or summary; follow the link for the original

来源概要：带完整布局附件的现代两居室，要求实际墙体门窗、家具、分房间集合与材料。
Source summary: a modern two-bedroom interior with a layout attachment, actual walls/openings, furniture, room collections and materials.

Archived source prompt SHA-256: 3947abd2d7e39da19d44da55cc45bb9ceeba7220bd9d25b9ba44485f6f3c8579

Public specification, with third-party wording omitted

以下为TOKRACE新增执行条件，并非原benchmark协议：本轮离线，仅本题公开输入，不访问公网或外部模型/资产库；k3.1-pro high，60分钟/50次请求/150000输出token，2线程。使用本机Blender5.2.3、EEVEE。原题的Cycles、OptiX、HDRI下载、采样/分辨率要求改为本次统一协议，必须在README披露；不声称官方BlenderWorld分数或严格复现原渲染管线。完整保留几何结构、尺寸、数量、配合、材质区分及原始UV/LOD等资产需求，不能为填任务数量简化；缺项如实列出。
先分功能块落盘完整可运行源代码，然后实际后台工厂启动建模/渲染，不使用单次巨型响应。交付final.blend、完整源码/资源、1920×1080主图+两张1600×900不同机位+结构/中性灰模近景；另交6秒1280×720/24fps的turntable.mp4（144帧）。减速器/蒸汽机可用机构运动代转台，但必须明确是运动学，非受力仿真，保留静态检查。先真实预览、view_image检查，最多两轮有效视觉修正，程序报错另记；保留首个预览与失败。视频检查至少6时点与首尾。独立进程重开工程，交artifacts.json/README（原题差异、实际命令版本、未完成项）/真实尺寸与对象清单自检JSON。离线程序材料替代外部图片/HDRI时必须说明；不得用平面或贴图代关键实体。

本题新增可测规格：完整读取layout.yaml。户型12×9m、净高2.8m、外墙240mm和隔墙120mm、家具空间要求保持。增加三张室内机位、一张正交顶视和一张门窗/隔墙或家具尺度的灰模近景，因此本题至少5张最终静帧。原题4K/OptiX/Cycles改为本次1920主图、1600副图、EEVEE是公开执行差异。按房间分集合，交可核对墙界、门洞、家具数量和尺寸的JSON。南向大窗与附件通用窗台规格若有解释冲突，说明选择，不能假称唯一真值。

公开说明：这是评测方新增约束及来源概要，原始完整输入本地保留并用SHA-256绑定。原输入/附件请从来源固定链接取回；不是完整实际输入逐字副本。原本地系统配音仅作评测证据，公开版由Kokoro替换旁白，未修模型画面。
Public note: these are evaluator additions and a source-summary rerun guide, not the full verbatim input. Full local inputs are hash-bound and source attachments must be retrieved separately. Public Kokoro narration replacement does not repair original imagery.

This is the evaluation adaptation and a rerun guide, not a verbatim transcript of the full original request. Exact original inputs remain archived locally.

Original evaluation input SHA-256: a4627ce3158c5672b60d6b51e1d7da605111ff2d3cdd5991ffa201e1854154e8

Attempts, failures and recovery · 1

attempt-01 · high · Model finished

Requests / complete responses: 25 / 25

Autonomous high session finished with 0 request error(s) and 1 runtime traceback record(s). Model edits, previews and re-renders are included; this was not a single call.

Attempt evidence

04 · Blender · Requested effort: high

Traditional Chinese temple courtyard and roof structure

Red columns, upturned golden roof, paired lions, stone axis, bell and lanterns form a real 3D courtyard. The hall roof is separated from supports by at least 70.56mm; some suspension rods miss structure and visual revisions exceed two rounds.

Download MP4 · Playback starts only when you press play.

Six-second real camera orbit around the same scene

Task wall time: 31m 27s
Requests / complete responses: 26 / 26
Received input tokens, incl. cache: 549,262
Received output tokens, incl. thinking: 26,664

AI assistant observations

Independent geometry and multiple views confirm 14 red columns, two recognizable lions, 45 path slabs, one bell and ten lanterns; eave corners rise about 0.58m above center.
The 205 roof objects start at Z5.0255566m while 347 intended hall supports end at Z4.9549999m, a 70.56mm global gap.
Four hall suspension rods have 359–378mm upward structural gaps; at least four pavilion lanterns meet beams. Do not call all lanterns absent or all correctly hung.
The gate obscures the lower hall/lions in the hero view; other views clarify them. All 1974 objects share one Collection and three visual correction batches exceed the frozen two-round limit. The actual first preview survives in harness evidence.

These are qualitative AI observations, not independent human scores. Automated checks are listed separately.

Original case and evaluation changes

Published task and specification

The linked creator or official example is a source. The output above is our independently generated run.

Itemized verification

Real geometry and required elements· Pass: Lions, upturned roof and spatial layout corroborated from reverse and overhead views.
Contact and revision budget· Fail: 70.56mm roof separation, some disconnected suspension rods and three visual correction batches.

Source brief and public rerun specification

Source excerpt or summary; follow the link for the original

来源概要：以红柱金顶主殿为中心的传统寺庙庭院，石狮、青石路径、铜钟与灯笼，黄昏入口斜向视角。原文未给精确建筑尺寸。
Source summary: a dusk temple courtyard around a red-columned golden-roof hall, with lions, stone path, bell and lanterns; no exact architectural dimensions were specified.

Archived source prompt SHA-256: a21d6b051543e08ce7b43405090a6ca4f7fd1e1f46f4a4038036f66114c43e11

Public specification, with third-party wording omitted

以下为TOKRACE新增执行条件，并非原benchmark协议：本轮离线，仅本题公开输入，不访问公网或外部模型/资产库；k3.1-pro high，60分钟/50次请求/150000输出token，2线程。使用本机Blender5.2.3、EEVEE。原题的Cycles、OptiX、HDRI下载、采样/分辨率要求改为本次统一协议，必须在README披露；不声称官方BlenderWorld分数或严格复现原渲染管线。完整保留几何结构、尺寸、数量、配合、材质区分及原始UV/LOD等资产需求，不能为填任务数量简化；缺项如实列出。
先分功能块落盘完整可运行源代码，然后实际后台工厂启动建模/渲染，不使用单次巨型响应。交付final.blend、完整源码/资源、1920×1080主图+两张1600×900不同机位+结构/中性灰模近景；另交6秒1280×720/24fps的turntable.mp4（144帧）。减速器/蒸汽机可用机构运动代转台，但必须明确是运动学，非受力仿真，保留静态检查。先真实预览、view_image检查，最多两轮有效视觉修正，程序报错另记；保留首个预览与失败。视频检查至少6时点与首尾。独立进程重开工程，交artifacts.json/README（原题差异、实际命令版本、未完成项）/真实尺寸与对象清单自检JSON。离线程序材料替代外部图片/HDRI时必须说明；不得用平面或贴图代关键实体。

本题新增可测规格：本题增加对侧、俯视整体与屋顶/柱梁灰模近景。原题未固定长度，按合理古建尺度自主设计并记录，自定尺度不写成原题。石狮为一对实体，红柱/金顶/上翘屋檐、青石直达路径、铜钟与红灯笼应清楚可辨。

公开说明：这是评测方新增约束及来源概要，原始完整输入本地保留并用SHA-256绑定。原输入/附件请从来源固定链接取回；不是完整实际输入逐字副本。原本地系统配音仅作评测证据，公开版由Kokoro替换旁白，未修模型画面。
Public note: these are evaluator additions and a source-summary rerun guide, not the full verbatim input. Full local inputs are hash-bound and source attachments must be retrieved separately. Public Kokoro narration replacement does not repair original imagery.

This is the evaluation adaptation and a rerun guide, not a verbatim transcript of the full original request. Exact original inputs remain archived locally.

Original evaluation input SHA-256: b5d50af2e81570240262133a2cf95e210244803feb16a3ad3528c36c8b1f680e

Attempts, failures and recovery · 1

attempt-01 · high · Model finished

Requests / complete responses: 26 / 26

Autonomous high session finished with 0 request error(s) and 3 runtime traceback record(s). Model edits, previews and re-renders are included; this was not a single call.

Attempt evidence

05 · Blender · Requested effort: high recovery; max failures retained

Jovian orbital starship repair dock

Two max attempts did not deliver a complete scene. A separate high recovery produced the project and four views, but floating crew and oversized handrails prevent a geometry/scale pass.

Neutral clay view: internal engine geometry

Task wall time: 50m 11s
Requests / complete responses: 27 / 22
Received input tokens, incl. cache: 395,709
Received output tokens, incl. thinking: 25,867

AI assistant observations

Independent inspection found 673 objects, nine functional collections and four complete arm groups; the engine includes 105 meshes and 13 curves, rather than only painted detail.
Evaluated geometry along the tilted deck normal places both astronauts' boots about 0.500m above the deck. The handrail top is 17.450m above it, about 8.44× the 2.068m crew height.
The main view and neutral close-up reveal structure; some colored supporting views are dark. The planet material is artistic and was not graded for astronomical accuracy.

These are qualitative AI observations, not independent human scores. Automated checks are listed separately.

Original case and evaluation changes

Original creator case

The public rerun specification below records our adaptations. Source demonstrations are not TOKRACE results; we generated independent outputs.

The linked creator or official example is a source. The output above is our independently generated run.

Itemized verification

Project and actual geometry· Pass: The reopened project contains organized objects, arms and engine geometry, with no external image/font/cache dependencies.
Contact and scale· Fail: Crew float about 0.500m above the deck and handrails are oversized. Geometry presence does not establish plausible engineering connections.

Source brief and public rerun specification

Source excerpt or summary; follow the link for the original

来源概要（非原文转载）：以木星为背景，制作巨舰在环形船坞维修的工业科幻场景，交付工程、多机位和内部灰模近景。
Source summary, not a quotation: an industrial orbital ship-repair scene near Jupiter, with an editable project, multiple views and a neutral internal-structure close-up. See the linked creator report for the full source brief.

Archived source prompt SHA-256: 93a8724df0253fe95a3864cd8b2b075435b69a499043237ab1bb15971c0ec730

Public specification, with third-party wording omitted

评测方公开复跑规格概要（不是原题逐字转载）：在木星轨道创建原创维修船坞，巨舰约 1.5km、半开放船坞约 3km；舰尾暴露真实框架、设备和管线，至少三条有底座/关节/工具的机械臂。前景栈桥含两名宇航员，按人—设施—巨舰—船坞表现尺度。保留功能集合，交付可编辑工程、1920×1080 主图、两个不同方向图和中性内部近景。使用独立后台 Blender 5.2.3 EEVEE；先看灰模预览，最多两轮视觉修正，错误另记。max 的 600/900 秒失败与 high 的 900 秒恢复独立保留，完整原题按链接与哈希查询。

公开说明：上述是评测方适配规格与概要，省略第三方原题长文，不是完整实际输入逐字副本。初始请求使用 k3.1-pro、max_tokens=32768；每题独立空目录与隔离工具环境。模型自行写代码、执行和看图；失败与重试不覆盖。除木星独立 high 恢复外，其余为明确请求 max。使用 Coding 订阅额度，金额和失败请求的最终用量未知。原始完整输入的哈希与每次请求元数据在独立摘要中。

Public note: this is the evaluator adaptation and source-summary rerun guide, not the full verbatim input. Original complete inputs are retained locally and hash-bound. The original local videos used macOS system voices; public narration replacements are separately identified and are not additional Kimi evaluation calls.

This is the evaluation adaptation and a rerun guide, not a verbatim transcript of the full original request. Exact original inputs remain archived locally.

Original evaluation input SHA-256: 100497e84f30f135dcb378b247ad90b7f49662730600e34ec9eb9bd11b4a3e10

Attempts, failures and recovery · 3

attempt-01 · max · Stopped after timeout

Requests / complete responses: 6 / 3

max: a 600-second deadline, a network error and one interrupted automatic retry are retained; no complete scene was delivered. Final failed/interrupted usage is unknown.

Attempt evidence

attempt-02 · max · Stopped after timeout

Requests / complete responses: 3 / 1

max: a separate 900-second recovery still did not deliver a scene; one later request was interrupted. Initial partial usage is not final consumption.

Attempt evidence

attempt-03 · high · Model finished

Requests / complete responses: 18 / 18

high: a changed-setting recovery delivered the editable project and four views. This is not a max success or a controlled effort comparison; a runtime missing-file error and repair remain recorded.

Attempt evidence

06 · Blender · Requested effort: max

Parametric bookshelf and parameter revision

Both variants pass independent dimensions, board-thickness and equal-clearance checks. Parameters rebuild through the Python CLI; the project property panel is not a live parametric editor.

Baseline front: 800mm, three compartments

Baseline rear: exposed open-backed construction

Task wall time: 25m 6s
Requests / complete responses: 28 / 27
Received input tokens, incl. cache: 751,885
Received output tokens, incl. thinking: 39,960

AI assistant observations

Independent opens verify 800×300×1200mm / three compartments and 1000×300×1200mm / four compartments, with 18mm boards and 0.1mm tolerance.
Named panels remain separate from display props; the rear view exposes the open-backed construction.
The product visualization is coherent, but structurally simple; only two parameter sets were tested.

These are qualitative AI observations, not independent human scores. Automated checks are listed separately.

Original case and evaluation changes

Official integration example

The public rerun specification below records our adaptations. Source demonstrations are not TOKRACE results; we generated independent outputs.

The linked creator or official example is a source. The output above is our independently generated run.

Itemized verification

Exterior, thickness and clearances· Pass: Both variants pass independent checks; this does not validate arbitrary parameter combinations.
Editability and revision· Partial: Native panels are editable; parameter changes rebuild via script, while UI custom properties do not drive live geometry.

Source brief and public rerun specification

Source excerpt or summary; follow the link for the original

Short source quote: “Write a Blender script that models a parametric bookshelf.”
来源为 Vercel 的官方接口代码示例，并非一份已公开的渲染作品。 / This is an official API example, not a published rendered artwork.

Archived source prompt SHA-256: 86151fda84cd34fb1612185d4bf3edc051e49c38fd163d190b9acf2ae186f079

Public specification, with third-party wording omitted

保留官方原题，增加下列可验收要求：
使用Blender原生几何制作可编辑木质书架，交付一个可通过命令行参数重建的Python脚本。
基础版外框宽800、深300、高1200毫米；全部板厚18毫米；由2块侧板、1块顶板、1块底板、2块内部水平隔板组成，形成3个等高净空储物格；无背板、无额外脚。顶底板嵌在两侧板之间，所有外边界尺寸必须一致。每块板独立命名，集中放在Bookcase集合；展示道具可有少量书，但不得混入Bookcase集合影响测量。
脚本公开width_mm、depth_mm、height_mm、thickness_mm、compartments、wood_color参数。用同一脚本再生成宽1000、高1200、深300、板厚18、4个等高储物格的修改版，不要另写独立几何代码。
交付baseline.blend和modified.blend、对应真实尺寸/对象清单JSON、至少3张1920×1080图片（基础版主视角、背面、修改版主视角），和清楚说明层数语义的README。模型背面要可检查，不能用背景平面代替实际建模。
先渲染低分辨率预览并用view_image检查。最多两轮有效视觉修正；执行报错可以另行修复。最终用独立Blender进程重开两个工程，确认集合、材质、相机和可测尺寸。保留脚本、预览、日志，不制作动画。

公开说明：上述是评测方适配规格与概要，省略第三方原题长文，不是完整实际输入逐字副本。初始请求使用 k3.1-pro、max_tokens=32768；每题独立空目录与隔离工具环境。模型自行写代码、执行和看图；失败与重试不覆盖。除木星独立 high 恢复外，其余为明确请求 max。使用 Coding 订阅额度，金额和失败请求的最终用量未知。原始完整输入的哈希与每次请求元数据在独立摘要中。

Public note: this is the evaluator adaptation and source-summary rerun guide, not the full verbatim input. Original complete inputs are retained locally and hash-bound. The original local videos used macOS system voices; public narration replacements are separately identified and are not additional Kimi evaluation calls.

This is the evaluation adaptation and a rerun guide, not a verbatim transcript of the full original request. Exact original inputs remain archived locally.

Original evaluation input SHA-256: e70b753db31dae2f8f83a0a1f1f70fa116f19b051f21e3a6b148879fd5f1ea4f

Attempts, failures and recovery · 1

attempt-01 · max · Model finished

Requests / complete responses: 28 / 27

The autonomous session finished with 1 request error(s) and 0 recorded runtime traceback(s). It is not a single model call; autonomous edits and re-renders are included.

Attempt evidence

07 · Knowledge video · Requested effort: high

Atmospheric circulation: three cells and deflected winds

The 90-second film largely gets cell order, closed circulation and indirect Ferrel driving right, but both hemisphere panels follow the same oriented Coriolis curve. It fails to depict northern-right/southern-left deflection.

Audio version

No speech by default; retains the original model-code chords and soft bass pulses. On-screen text and captions remain.

Download selected MP4 · Playback starts only when you press play.

Music/effects by default; original picture unchanged, legacy narration available by explicit choice.

Public version and original evidence

Only the mixed soundtrack is replaced with independent procedural music/effects. H264 packet bytes and timing match exactly; picture defects remain. This is not a Kimi rerun or rescore. Historical public Kokoro narration is optional; original system voices remain local.

Original SHA-256: 99acfa844c97c33ad2e7f820c6acc3a792d26c912230120e52f3072f505595a5

Public SHA-256: 3879b0a0c5713022bf1a15931b6871c74b95b8bf217f3b43e75791f590a60a21

Failure evidence: identical Coriolis direction

Task wall time: 10m 20s
Requests / complete responses: 21 / 21
Received input tokens, incl. cache: 486,031
Received output tokens, incl. thinking: 14,951

AI assistant observations

The 0/about30/about60/90 order and cell rising/sinking/surface/upper directions are consistent; the script states three cells per hemisphere, shifting boundaries and a long-term-average idealization.
At 58–69s both hemisphere panels reuse the same path/traversal with changed labels/colors; the southern example says southward while moving upward. Correct words do not repair the trajectory.
Three cell routes close continuously; a separate open Coriolis path jumps back near 64.25s. Lower branches extend about 51px below the drawn ground line.
Sampled bilingual captions remain on screen, though 22px English can be small on phones. No exhaustive word-level listening is claimed. Public replacement narration preserves picture/scientific errors.
Playback revision: narration is off by default and original procedural background audio is restored. This changes audio only, not the knowledge or picture findings above.

These are qualitative AI observations, not independent human scores. Automated checks are listed separately.

Original case and evaluation changes

Creator demo and public prompt

The linked creator or official example is a source. The output above is our independently generated run.

Itemized verification

Cell flow and explanatory limits· Pass: Main cell directions/order, indirect Ferrel driving and idealization caveats are supported.
Coriolis animation· Fail: Correct north-right/south-left wording but identical oriented curves for both hemispheres.

Source brief and public rerun specification

Source excerpt or summary; follow the link for the original

来源概要：作者自述使用Opus 5.5制作大气环流讲解，并公开创作提示；来源视频实测288.133秒。我们仅据公开题意独立生成，没有取用原片工程，未独立确认原作者后台模型身份。
Source summary: the creator attributes the circulation demo to Opus 5.5; its public video measures 288.133 seconds. We independently generate from the brief without reusing the original project or independently verifying the creator’s backend model identity.

Archived source prompt SHA-256: b021ced84260e927afb5f3cc86e16892c9957aa363def7c92ee2f5b012d3e897

Public specification, with third-party wording omitted

本次追加规格（与公开作者原prompt分开；以此消解原prompt中在线TTS配置要求）
任务：制作75–90秒中文大气环流知识短片，目标约90秒。保留线稿动画、中英双语字幕、合适的轻量程序化音乐，交付可直接导出的完整视频。
变化说明：原作者案例为Opus5.5、288.133秒1920×1080；本次改用Kimi K3.1 Pro，压缩为75–90秒，TTS换为本地macOS Tingting，因此TTS.md、.env/APIKEY/VOICE和云水印参数均不适用。不是原片逐镜头复刻。
画幅1920×1080，30fps，有中文配音、烧录中英字幕及SRT。配音只用本地Tingting，不等待key；音乐若制作需程序化且旁白优先。
内容：解释不均匀加热和理想化三圈环流，重点画北半球0/约30/约60/90度的剖面。气团须沿每个闭合路径连续运动；对应地表信风、西风和极地东风，箭头与旁白一致。另用很短的地球/平面方向图解释北半球偏向运动方向的右侧，而不是‘永远向东’。注明南半球偏左以及真实气流比理想模型复杂、纬度边界会移动。以高中生能理解的方式讲一个实际含义，如副热带下沉与干燥。
不得把Ferrel环流说成简单独立热对流；不要求完整天气预报或喷流专题。纬度30/60为示意近似，不是固定自然边界。
执行只用已提供facts、Pillow/NumPy+FFmpeg、本地Tingting；只读inputs，不联网或安装陌生代码。从空目录自行编写原创制片源码。
交付MP4、旁白WAV、SRT、讲稿、完整源码、重建说明、工具/素材来源、至少8张实际编码成片代表帧（包括三圈路径关键状态和结尾）。每次失败与重试要保留，不手工替模型修动画。


公开说明：这是评测方新增约束及来源概要，原始完整输入本地保留并用SHA-256绑定。原输入/附件请从来源固定链接取回；不是完整实际输入逐字副本。原本地系统配音仅作评测证据，公开版由Kokoro替换旁白，未修模型画面。
Public note: these are evaluator additions and a source-summary rerun guide, not the full verbatim input. Full local inputs are hash-bound and source attachments must be retrieved separately. Public Kokoro narration replacement does not repair original imagery.

This is the evaluation adaptation and a rerun guide, not a verbatim transcript of the full original request. Exact original inputs remain archived locally.

Original evaluation input SHA-256: 54e042555ee878d98e39535585e04fec643e0ce62c3f1b824df591978061fc52

Attempts, failures and recovery · 1

attempt-01 · high · Model finished

Requests / complete responses: 21 / 21

Autonomous high session finished with 0 request error(s) and 4 runtime traceback record(s). Model edits, previews and re-renders are included; this was not a single call.

Attempt evidence

08 · Knowledge video · Requested effort: high

Derivatives: a secant approaching the tangent

Core mathematics in the 89.167-second film passes: P/Q, signed difference quotients, limit, tangent and the x=2 answer are correct. Label collisions, a negative-h reset, a missing glyph and lines crossing plot margins remain.

Audio version

No speech by default; retains the original model-code a slowly modulated three-frequency sine bed. On-screen text and captions remain.

Download selected MP4 · Playback starts only when you press play.

Music/effects by default; original picture unchanged, legacy narration available by explicit choice.

Public version and original evidence

Original SHA-256: f807b3b5f441b26670df6eeda5ff876633cd7df544f5e83bb1a844f64917e575

Public SHA-256: 8ddbd5ae5e2a9dc6b37468508454885cd28e467e900c6f7c19898e079962367d

Task wall time: 25m 57s
Requests / complete responses: 20 / 20
Received input tokens, incl. cache: 346,365
Received output tokens, incl. thinking: 13,161

AI assistant observations

Independent mathematical checks of all 2675 encoded frame states verify nonzero h, Q=(1+h,(1+h)²), quotient 2+h and tangent y=2x−1. The answer at x=2 is 4 and the actual purple line y=4x−4 is correct.
P/Q labels collide with Δx/Δy at multiple times; ⇒ is a missing glyph and unclipped secant/tangent lines cross legends/margins.
At 50.215s negative h resets from −0.350 to −0.600 before approaching zero again. Values remain correct but the same-side restart is unexplained.
The supplied 06_tangent still belongs to the question chapter; public evidence instead uses the independent 67.2s decoded final-video tangent frame. The question holds five seconds before the correct answer.
Playback revision: narration is off by default and original procedural background audio is restored. This changes audio only, not the knowledge or picture findings above.

These are qualitative AI observations, not independent human scores. Automated checks are listed separately.

Original case and evaluation changes

Creator demo and public prompt

The linked creator or official example is a source. The output above is our independently generated run.

Itemized verification

Core mathematics· Pass: Independent state expressions and targeted decoded frames show no confirmed final core math error.
Continuity and legibility· Partial: Negative-h reset, overlapping labels, missing glyph and plot-layout issues remain.

Source brief and public rerun specification

Source excerpt or summary; follow the link for the original

来源概要：作者自述用Opus 5.5/Manim制作导数学习视频，来源片实测457秒；要求通俗、举例并引发思考。我们保留题意，独立压缩为一个机制，未独立确认原作者后台模型身份。
Source summary: the creator attributes the derivative demo to Opus 5.5/Manim; its public video measures 457 seconds. We independently shorten its accessible, example-based brief to one mechanism without independently verifying the creator’s backend model identity.

Archived source prompt SHA-256: 5287dcf00a500b81f8363e9f8cfebeb93a90fd8f13440efe4ce0425bd0895383

Public specification, with third-party wording omitted

本次追加规格（与公开作者原prompt分开）
制作75–90秒中文导数概念学习短片，目标约90秒。原作者为Opus5.5+Manim、edge-tts，公开视频457秒1920×1080；本次改为Kimi K3.1 Pro，只保留‘通俗易懂、有例子、引人思考’的主题要求，压缩成一个机制并用本地Tingting替换edge-tts。不得称作完整原片复刻或Manim基准。
本轮不安装Manim/Remotion；使用Pillow、NumPy、FFmpeg离线原创动画。1920×1080、30fps、中文旁白、中文烧录字幕及SRT；轻量程序化音乐可选，旁白优先。
机制核心：以f(x)=x^2在P=(1,1)处为例。持续移动Q=(1+h,(1+h)^2)，同步更新割线、斜率三角形、h值与差商数值。让正h逐步缩小，随后演示非零负h从另一侧靠近；最后得到斜率2、切线y=2x-1。不得在h=0直接除0，不把多个互不连续的静态卡片当作割线向切线的过程。
开头提出具体问题‘这一瞬间到底有多陡’，结尾让观众预测f(x)=x^2在x=2的斜率并给短暂停留，随后给答案或明确为思考题。图上刻度、位置、线斜率和数值必须来自同一数学状态；负h的水平与垂直变化要保持符号一致。
交付MP4、旁白WAV、SRT、讲稿、完整原创源码、可重建命令和环境说明、工具/来源说明，至少8张从实际编码MP4抽取的代表帧，覆盖h较大、趋近、负h、切线、结尾。每次失败/重试保留。
源事实和技巧已在inputs，执行不得联网或安装陌生代码，不等待key。


公开说明：这是评测方新增约束及来源概要，原始完整输入本地保留并用SHA-256绑定。原输入/附件请从来源固定链接取回；不是完整实际输入逐字副本。原本地系统配音仅作评测证据，公开版由Kokoro替换旁白，未修模型画面。
Public note: these are evaluator additions and a source-summary rerun guide, not the full verbatim input. Full local inputs are hash-bound and source attachments must be retrieved separately. Public Kokoro narration replacement does not repair original imagery.

This is the evaluation adaptation and a rerun guide, not a verbatim transcript of the full original request. Exact original inputs remain archived locally.

Original evaluation input SHA-256: f331e0e46d1a51df29d0d9610d0b87f1bd8a9365b599b3b277ed29e9b9272279

Attempts, failures and recovery · 1

attempt-01 · high · Model finished

Requests / complete responses: 20 / 20

Autonomous high session finished with 0 request error(s) and 2 runtime traceback record(s). Model edits, previews and re-renders are included; this was not a single call.

Attempt evidence

09 · Knowledge video · Requested effort: max

A hand-drawn explainer of printer operation

The original run delivered a technically valid 115.7-second Chinese explainer, but its exposure beam crosses the drum interior despite narration describing surface exposure. It is shown as evaluation evidence, not a teaching-ready asset. Public playback defaults to procedural music/effects, with legacy narration optional; model imagery is not repaired.

Audio version

No speech by default; retains the original model-code pentatonic plucks, transitions and mechanism cues. On-screen text and captions remain.

Download selected MP4 · Playback starts only when you press play.

Music/effects by default; original picture unchanged, legacy narration available by explicit choice.

Public version and original evidence

Original SHA-256: d451ac82c8e2e9ab7f1f0de82d21150f203941ff9c1a960bfaf74fd571247f8a

Public SHA-256: 185d1963827bb200ca5dad083994a0f8a8e365aedfdfc1b56e1ce83a57a54303

Failure evidence: the exposure beam crosses the drum at 55.80s

Task wall time: 47m 43s
Requests / complete responses: 37 / 36
Received input tokens, incl. cache: 1,903,164
Received output tokens, incl. thinking: 64,101

AI assistant observations

The original is 1920×1080, 30fps and 3471 frames; full decode and all 25 SRT timing checks pass, with readable Chinese and distinct inkjet/laser processes.
At 55.80s and 58.50s, the red exposure beam crosses the drum interior to its far-side circumference. This contradicts surface exposure; successful encoding does not override the scientific diagram failure.
Scanning-mirror terminology is ambiguous against the source's scanning lenses. A suspected text-clipping issue was excluded after original-pixel inspection. No independent human listening study was performed.
Playback revision: narration is off by default and original procedural background audio is restored. This changes audio only, not the knowledge or picture findings above.

These are qualitative AI observations, not independent human scores. Automated checks are listed separately.

Original case and evaluation changes

Original creator case

The public rerun specification below records our adaptations. Source demonstrations are not TOKRACE results; we generated independent outputs.

The linked creator or official example is a source. The output above is our independently generated run.

Itemized verification

Original evaluation media checks· Pass: 115.7 seconds, 1920×1080 and 30fps, with complete decode and 25 valid caption cues.
Scientific animation· Fail: The exposure beam crosses the drum interior. Replacing narration does not repair this error; correction is needed before treating it as accurate teaching material.

Source brief and public rerun specification

Source excerpt or summary; follow the link for the original

来源概要（非长文转载）：以手绘风格讲解打印机原理。作者未公开原片完整工具链、精确时长和画幅。
Source summary: explain printer operation in a hand-drawn style. The creator did not disclose the complete original toolchain, exact duration or frame format.

Archived source prompt SHA-256: ef251d6c101cd8017f2169ef06b68b5c3eff683b7eea98e1a858ebbc59d802be

Public specification, with third-party wording omitted

本次新增可验收规格（原作者未披露时长/画幅/工具链，故不是原片逐镜头复制）：
制作90至120秒、1920×1080、30fps的中文科普MP4，含清晰中文旁白与中文字幕/独立SRT。整体保持手绘线稿风格，至少5个具有因果解释的动态场景，不能仅是文字PPT。
面向普通读者，区分喷墨和激光两种技术：墨滴如何到纸上、激光如何在感光鼓上形成潜像、显影/转印/热压定影。范围和技术名称要明确；画面、运动方向、材料标签与讲稿一致。
已提供两份Canon官方事实材料到inputs/，可以阅读并在讲稿说明来源。不得把激光画成直接把文字烧在纸上。
用系统say的Tingting配音即可，不必下载TTS。所有画面由你写代码绘制和动画；不得调用原生图像/视频模型。交付video.mp4、narration.wav或m4a、captions.srt、script.txt、可完整重新生成的动画源码、至少8张带时间标记的代表帧、README和artifacts.json。
你要实际完成配音、渲染和混音，检查时长、分辨率、帧率、音轨存在及字幕起止。不要为了省时间缩短到不足90秒；不要把解说字幕当作全部视觉内容。

公开说明：上述是评测方适配规格与概要，省略第三方原题长文，不是完整实际输入逐字副本。初始请求使用 k3.1-pro、max_tokens=32768；每题独立空目录与隔离工具环境。模型自行写代码、执行和看图；失败与重试不覆盖。除木星独立 high 恢复外，其余为明确请求 max。使用 Coding 订阅额度，金额和失败请求的最终用量未知。原始完整输入的哈希与每次请求元数据在独立摘要中。

Public note: this is the evaluator adaptation and source-summary rerun guide, not the full verbatim input. Original complete inputs are retained locally and hash-bound. The original local videos used macOS system voices; public narration replacements are separately identified and are not additional Kimi evaluation calls.

This is the evaluation adaptation and a rerun guide, not a verbatim transcript of the full original request. Exact original inputs remain archived locally.

Original evaluation input SHA-256: e306e9b72994388452b7673c223b45c7a99369796a5a0f71075f1e5fcf60aa87

Attempts, failures and recovery · 1

attempt-01 · max · Model finished

Requests / complete responses: 37 / 36

The autonomous session finished with 1 request error(s) and 3 recorded runtime traceback(s). It is not a single model call; autonomous edits and re-renders are included.

Attempt evidence

10 · Knowledge video · Requested effort: max

A motion-graphics guide to spoofed email

The original run delivered an 81.966-second English explainer with passing media checks and mostly sound authentication/alignment examples. From-field wording and a UI glyph need refinement. Public playback defaults to procedural music/effects, with legacy narration optional and no re-scoring of the model run.

Audio version

No speech by default; retains the original model-code atmospheric bass, chimes and filtered-noise cues. On-screen text and captions remain.

Download selected MP4 · Playback starts only when you press play.

Music/effects by default; original picture unchanged, legacy narration available by explicit choice.

Public version and original evidence

Original SHA-256: bd80f64d6a47aed64fae0601f52a743acadbf240542dee57d589d3d1d957c0f2

Public SHA-256: 762405500d97de58f7b9bce5766cff3a9c5d87d4910218c503db1cbdd1baaa2f

Task wall time: 47m 40s
Requests / complete responses: 30 / 29
Received input tokens, incl. cache: 1,735,111
Received output tokens, incl. thinking: 61,061

AI assistant observations

The original is 1920×1080, 30fps and 2459 frames; full audiovisual decode and all 17 SRT timing checks pass, with eight scenes and an ending.
SPF/DKIM pass for example.org while From is example.com, correctly causing DMARC failure through nonalignment. The film preserves that pass is not a safety guarantee and failure does not prove fraud.
Calling From only a display label is imprecise; a Gmail menu glyph appears as a box and dense headers/acronym captions could be clearer. Sampled frames and technical audio checks are not independent human listening.
Playback revision: narration is off by default and original procedural background audio is restored. This changes audio only, not the knowledge or picture findings above.

These are qualitative AI observations, not independent human scores. Automated checks are listed separately.

Original case and evaluation changes

Original creator case

The public rerun specification below records our adaptations. Source demonstrations are not TOKRACE results; we generated independent outputs.

The linked creator or official example is a source. The output above is our independently generated run.

Itemized verification

Original evaluation media checks· Pass: 81.966 seconds, 1920×1080 and 30fps; complete decode and valid caption timing. The public narration derivative has a separate hash.
Factual content and wording· Partial: Alignment and authentication limits are mostly sound; From wording is simplified. No live Gmail or real-mail authentication was tested.

Source brief and public rerun specification

Source excerpt or summary; follow the link for the original

Short source quote: “Can you create a motion graphic explainer video that talks about How to avoid getting spoofed by fake email?”
作者此前的 YouTube 风格上下文未公开，本次没有宣称逐镜头复刻。 / The creator’s prior YouTube style context was unavailable; this is not a shot-for-shot style reproduction.

Archived source prompt SHA-256: 1982b9439ae02d4828e0f5aca5cba72618cd25c6462778d7c046bdffd59a9398

Public specification, with third-party wording omitted

Evaluation additions and limits: the source's earlier YouTube style context is unavailable. Create an original visual treatment, not a claimed exact style replica. The original media is about74.2s at1920×1080; this run requires75–85 seconds,1920×1080,30fps, English narration and synchronized English captions/SRT.
Make a polished motion-graphics explainer for nontechnical viewers, with at least5 animated scenes that explain causality rather than just text slides. Use clearly fictional example.com/example.org addresses and illustrative messages; do not access or send real email.
Read the supplied official Gmail authentication material and RFC9989 in inputs/ when needed. Explain visible sender identity, trusted original/header information and SPF/DKIM/DMARC at an accessible level; retain the limits of what authentication can prove. Do not imply a pass proves harmless content or a failure proves fraud. If you use Authentication-Results, show results added by the receiving service, not arbitrary untrusted text as authoritative.
Use the local macOS Samantha voice via say. You must write all visual/animation code yourself and actually render/export. Deliver video.mp4, narration.wav or m4a, captions.srt, script.txt, complete rebuildable source, at least8 timestamped representative frames, README and artifacts.json. Confirm duration, dimensions,30fps, audio presence, captions and complete ending. Do not use image/video/music generation models or download large packages. The base renderer is your choice among installed tools.

公开说明：上述是评测方适配规格与概要，省略第三方原题长文，不是完整实际输入逐字副本。初始请求使用 k3.1-pro、max_tokens=32768；每题独立空目录与隔离工具环境。模型自行写代码、执行和看图；失败与重试不覆盖。除木星独立 high 恢复外，其余为明确请求 max。使用 Coding 订阅额度，金额和失败请求的最终用量未知。原始完整输入的哈希与每次请求元数据在独立摘要中。

Public note: this is the evaluator adaptation and source-summary rerun guide, not the full verbatim input. Original complete inputs are retained locally and hash-bound. The original local videos used macOS system voices; public narration replacements are separately identified and are not additional Kimi evaluation calls.

This is the evaluation adaptation and a rerun guide, not a verbatim transcript of the full original request. Exact original inputs remain archived locally.

Original evaluation input SHA-256: 1567e8e81385ae0f1709b27f4126337857ab012e6cd739c317de74bfbdfe95b0

Attempts, failures and recovery · 1

attempt-01 · max · Model finished

Requests / complete responses: 30 / 29

The autonomous session finished with 1 request error(s) and 2 recorded runtime traceback(s). It is not a single model call; autonomous edits and re-renders are included.

Attempt evidence

Evidence and limits

Download the evaluation dataset and source index

Ten source-adapted tasks span two batches with different execution conditions; this small sample is not a general ranking or controlled max/high comparison.
Long third-party prompts and full source pages are not republished. Public specifications summarize the source and evaluator adaptations; complete inputs, request/tool traces, failures and immutable previews remain hash-bound in the local archive.
AI qualitative observations and executable geometry/media checks are distinct from independent human blind scoring; no such scoring or exhaustive subjective listening is claimed.
Playback defaults to original procedural music/effects without speech. Historical public Kokoro narration is an explicit option and includes accelerated speech. Original system voices remain local. Audio revisions preserve H264 packets/timing and do not repair picture defects or revise failure findings.
A mutable route and returned model ID do not independently establish an internal fixed checkpoint; findings apply to this date and environment.

Production techniques and audio

The useful ideas are a storyboard tied to a shared time axis, deliberate framing and motion, deterministic rendering, and checks across every scene and transition. Narration is optional. These resources can inform that workflow. Re-rendering a ready-made template is a separate pipeline check, not an independently generated model result.

What limited these videos?

This was a short adaptation: the circulation source was 288 seconds and the derivative source 457 seconds; our outputs are 90 and about 89 seconds. The added tasks required Pillow/NumPy/FFmpeg, excluded Manim and Remotion, and allowed at most two effective visual-revision rounds. The model still finished within its budget and inspected its own frames, but missed direction errors, label collisions and motion discontinuities. The production constraints and the model's implementation both matter; no controlled A/B test has isolated their contributions.

Popular Opus examples do not establish a universal Claude Code plus skill recipe. The AI-history creator explicitly used Claude Code and Remotion; Deedy disclosed Manim, Kokoro and FFmpeg without naming a skill; opus-video-recipes includes both Cursor and Claude Code workflows. A main prompt can initiate many model calls, tool runs and revisions.

Remotion official skills — Official workflow guidance. A separate license for the skills repository was not established in this check; the framework has its own Remotion License with free-eligibility and Company License terms. Do not label both as unrestricted MIT software.
HyperFrames — HTML-based video production and rendering; its official repository declares Apache-2.0.
opus-video-recipes — Recipes for prompts, timelines and rendering. Code and prompts are MIT; the Lovely app screens and demo photographs are explicitly excluded.
opus-video-skills — Creator-published skills emphasizing scene intent, a shared cue timeline, contact sheets and motion strips. Referenced here for process design; these skills were not loaded in the historical Kimi runs.
Manim Community — A code-driven framework for mathematical and explanatory animation, distributed under MIT.
Kokoro-82M — The official model card declares Apache-2.0 weights, and the code repository also uses Apache-2.0. Record the exact model, voice source, version and audio hash.

Knowledge videos now default to nonverbal, code-generated music and sound effects. On-screen labels and captions remain. The player offers the historical Kokoro narration as an explicit option; some sentences in that version were accelerated to fit fixed time windows. Original macOS voice tracks remain in the local archive. Changing the audio preserves the model's picture and its scientific errors; this is a playback revision, not a new model result. Future runs generate narration only when requested, measure its duration before freezing the timeline, and record the voice source and cost separately.

Workflow audit and original case sources · Nonverbal audio revision source and integrity audit

Historical narration pipeline source and pinned environment (weights excluded)