AV event and music evidence

Installed workflow resources

For a Claude installer layout, use the adjacent managed support directory ../video-research-resources/ and its ../video-research-resources/docs/integrations/AV_EVENTS.md. The ../video-research-resources/integrations/qwen/av-events.json descriptor records the selected component and requirements. 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.

Use the root tools media_caption_events, media_count_events, media_ground_events and media_analyze_music. Read the integration contract when choosing budgets, interpreting support or handling incomplete populations. Existing media_perceive remains available for its published joint spoken/visible timeline contract.

Bind one regular local file with its full expected_source_sha256, choose a bounded interval, and prepare with dry_run=true. Inspect the exact source, actual frame PTS/indices, decoded audio interval, gaps and planned windows. A dry plan can perform local native preparation, but makes zero provider calls.

Choose finite fps in 0.1..30 inclusive; the default is 1. The extractor samples available source frames without interpolation. Set explicit frame budgets when requesting higher rates: max_frames_per_window defaults to 32 and caps at 48, with at most 128 aggregate frames. A reached frame budget reports partial visual sampling and coverage.stop_reason=frame_budget; inspect actual PTS, sampled points and gaps. A higher rate does not enlarge byte, deadline or provider budgets, and sampled points do not establish continuous watched coverage.

Submit with dry_run=false and authorize_submission=true only when the workflow has authority for that source upload and spend. Use the configured Gemini account; never add endpoints, credentials, uploads or alternate services to a request. The adapter has no installer or foreign Qwen execution route.

Choose the concrete task:

• Caption supported chronological audio, visual and fused occurrences.
• Count an explicit nonempty target; use only the server's admitted record count.
• Ground a nonempty query; choose positive integer top_k and retain the disclosed available/selected/truncated populations. Scores are uncalibrated model labels.
• Analyze music as audio only, including from a video container. Retain timed sections, instruments, moods and tags. Tempo, key, meter and section boundaries are inferred; decoded sample clocks measure the submitted waveform.

Report records with their source SHA, stable record ID, absolute and local times, window index and actual audio/visual support. Both modalities supporting one fused record count once. Keep distinct overlapping records and cross-window boundaries; do not invent a physical identity merge or continuous watched coverage. Do not rename nonspeech audio as a spoken claim.

Keep planned, valid empty, abstained, partial and failed outcomes distinct. A dry count is unknown. An abstention does not prove absence. A later window failure retains earlier records and count_so_far, while the final total remains unknown. Include all planned and completed windows, measured gaps, attempts and unknown usage. Do not repair JSON, silently truncate records or broaden retries after a terminal failure. Cancellation must finish owned cleanup before returning.

Decoded clocks and byte commitments establish submitted lineage, not event or music correctness. The programme's reviewed live semantic journey remains RESOURCE_UNAUTHORIZED; source and mocked SDK checks do not close that gate.