Verify a diagram against its rule

Installed workflow resources

For a Claude installer layout, use the adjacent managed support directory ../video-research-resources/ and its ../video-research-resources/docs/integrations/qwen-education.md. The ../video-research-resources/integrations/qwen/education.json records the selected component and requirements. The local scripts/lesson.py requires the separately installed video_research_mcp.education Python package from the accepted candidate. Check its normalized metadata version and wheel receipt as described in the component contract. Installing this skill does not install that runtime or its optional native dependencies. The repository-relative references below apply when using a source checkout or unpacked npm package. Installing resources does not activate optional runtimes, providers or external source components.

Read the component contract before authoring the required source/analysis/script/storyboard ledger. This skill uses the installed first-party lesson CLI, with no public MCP tool, foreign renderer, provider generation, font package or automatic installer.

The admitted lesson has three scenes: right 3-4-5 triangle, input-reflected square and one battery/resistor series cycle. Bind exact source JSON and supplied regular 48kHz mono PCM16 WAV with independent full SHA-256 values. Preserve the exact transcript. Schema1 retains caller-declared two-second intervals and its exact six-second WAV. Schema2 requires three ordered hash-bound narration_segments (triangle, curve, circuit) and omits declared scene/caption times. Its measured sample counts set all whole-scene caption and page boundaries; combined duration is at most30 seconds. No default voice or synthetic audio is generated. Supplied tones can check the component's sample contracts; they do not satisfy actual spoken lesson acceptance.

Run lesson.py validate with --spec and --spec-sha256. Schema1 also requires --audio and --audio-sha256; omit both for schema2. Relative segment paths are resolved beside the source JSON. It checks every required scene, numeric rule, supported glyph, caption timing/layout and complete audio sample population without native execution. Failed, missing or skipped applicable gates are refusals, never a source PASS.

Run lesson.py build with the same exact inputs and an absent --output directory. The explicit workflow uses optional Pillow and operator-installed FFmpeg/ffprobe; a missing dependency is reported rather than installed. It writes the self-contained page, the complete admitted12fps frame population (72 for schema1, at most360 for schema2), lossless local MP4, complete decoded RGB/PCM evidence and a committed receipt. One deadline covers the operation. Cancellation joins owned work and preserves inputs/prior artifacts. Retained failed-attempt receipts do not authorize regeneration or overwrite.

Schema2 preserves all original WAV bytes and concatenates only their exact PCM. Page scene/seek/reset controls use the measured sample boundaries. Video chooses the scene at each frame's start sample; transitions and the video end may round up by less than one frame. Inspect the recorded visual_delay_samples and video_tail_samples; audio stays at its exact sample count. A scene with no scheduled frame is refused. No word timing is estimated and supplied text/speech alignment remains unverified. Version2 receipts are bounded to512KiB and full decoded RGB to256MiB; schema1 receipts and caller artifacts remain supported.

Retain the returned receipt hash outside the mutable output directory. Inspect the actual page's scene, seek/reset and reflection/shift controls, visible numeric samples and parenthesized square equation. The fixed x=-2..2,y=0..4 viewport may clip changed curves; numeric rules remain visible. Page/video text uses the same authored ASCII uppercase cells, with exact source text preserved. CJK/unsupported glyphs are explicit refusals.

Run lesson.py check OUTPUT --receipt-sha256 RETAINED_SHA256 after restart. Do not derive that expected hash from an edited self-receipt. Checking repeats source predicates and complete decoded RGB/sample comparisons, including every frame's source lineage, glyphs and captions. Read back the original and output byte commitments and report separate source/finished-output states.

Keep spoken semantics, actual speech/caption alignment, physical/arbitrary-domain correctness, human lesson acceptance and multilingual assets unverified. All 12 upstream templates, HyperFrames/GSAP/KaTeX/fonts, foreign validators, provider narration and broader scientific animations remain unexecuted/unqualified. A three-domain component receipt does not close those workflows or the broader leaf. Native MP4 is a lossless master; browser codec delivery still requires a separate qualified runtime journey.