# iPolloWork Video Acceptance Load applicable checks for delivery/delegated review; small edits check affected claims plus playback. Batch independent source and temporal checks. Run one initial check and at most two targeted repair rounds, repeating only affected checks. If the same issue remains after two repairs, preserve the editable draft and report the exact field, scene/time and blocker; do not repeat unchanged checks, search installed program/cache trees, or shorten source media to pass. The native main Agent calls the existing component/project/voice and preview-check tools, reads their real results and repairs defects; reuse the existing Studio service. Technical health and actual expression are separate verdicts. Default generation/editing acceptance is `acceptance=studio-preview`, including direct conversation and uploaded-video enhancement. Run `media/video_component_check` for source structure and `media/video_preview_check` with the exact `sourcePath` for executed timing/layout, continuity and direct/reverse seek. Then seek and play changed scenes and seams in the current built-in Studio canvas, listening to required audio. These checks replace rendered temporal samples, motion-audition clips, contact sheets and exported audio/video review. Do not render, export, record, start another browser/service or require an MP4 receipt for validation. Return the editable project and the checks actually observed; `playback=unverified` from the technical tool is not a watched/listened verdict. Only explicit MP4/file delivery or authorized publication enables the separate export stage. Legacy `reviewOnly:true` also performs preview inspection without creating a video. ## Source and timeline - One valid root owns stage/aspect ratio, FPS, duration, tracks and paused timeline. Scripts/media load without console/decode/path errors; scene/clip/transition/caption/audio/animation windows agree within integer-frame bounds. - Full scenes retain stable IDs, `data-ipw-scene`, track/windows, motion pattern, truthful timing source and continuous executable beats. Registry hosts retain unique composition IDs, real source/variables/registry marker and `data-ipw-timing-owner="host"`; inner roots have no competing window, Land persists to host end. Custom exceptions retain policy/evidence; variable IDs, tokens, brands, hooks and paths remain valid. - Direct/reverse seek, replay/reopen/export are history-independent. Invalid timeline, zero/inconsistent duration, out-of-bounds clips, broken variables or required missing assets fail; source markers do not prove execution. ## Content, recipe fit and sequence Review the approved source/brief, latest native storyboard, actual selected card/variant/source and mounted implementation against current Studio evidence or the separately requested export, without assuming construction succeeded. - Required facts, names, units and qualifiers are spoken or visibly explained in the correct order. Each scene has purposeful primary content at its reading state; blank/placeholders fail except an explicitly approved blank beat. Branding, captions/transcripts and persistent chrome alone are not scene content. Corrected subjects must appear in actual speech/text/pictures/captions, not merely replacement imagery. - Check final-font fit, hierarchy, contrast, safe margins, crop and scale at output dimensions. Preserve supported editable data/variables, theme inheritance and selection hooks. Label illustrative imagery and retain factual evidence attribution. - Recipe useWhen/avoidWhen, inputs, capacity, reading order and usage limits hold. Compare real choreography/easing/framing/landing with accessible reference footage; a card name or moving wrapper does not prove explanation. A preview HTTP success is not proof of having watched it; missing reference playback leaves fidelity unverified, never pixel-parity claims. - Imports record exact card/variant, pinned rule/source and actual installed component. Reference-only/unported entries and recipe labels on original graphics are not executable reuse. Under recipes-only, missing capacity/input support needs another valid recipe, semantic split or gap disclosure; custom graphics fail that policy. - Sequence serves the source/audience: required facts remain, optional slots may merge with reason. Preserve dependency order, comparable units/viewpoints and genre energy; reject unsupported causality/winners, duplicated title cards and promotional rhythm imposed on teaching. - Real-page/camera work retains sharp texture/cutout/backplate dimensions and `assets/capture-layout.json` coordinates for actual landing slots. Reject blurry closeups, detached annotations, floating destinations or fabricated interfaces substituting supplied evidence. Representative styleframes or their justified skip remain traceable. ## Temporal samples and continuity Seek Establish, Develop, and Land states in Studio for default acceptance. Render Establish, Develop, and Land samples only from an already requested export when native-size file review is required; never generate a second review movie. Preserve scene/time/frame identities; time ranges, visual focus, visual action, and result or hold plus targets/references cover relevant speech/reading without gaps/overlaps. Speech uses measured provider timing/project FPS; silent work uses its real driver. Purposeful teaching, subtitle reading and final holds remain valid. Samples show actual focus/state/evidence/path/framing progression, with information readable when discussed. Generic entrance followed by an unintentionally static long explanation fails. Check the selected pattern's real development: montage states, camera waypoints/subject-follow, dialogue responses, kinetic meaning or measured audio-driven accents. A component name, metadata label or tiny continuous easing cannot substitute for these events. Inspect outgoing Land + seam + incoming Establish at changed/flagged boundaries. Confirm real cut/preset/custom handoff execution and recognizable continuity or intended reset. Windows meet; incoming motion consumes Establish. Outgoing meaning, non-empty incoming state and fixed captions/identity/audio survive. Reject empty fades, hidden qualifiers or replayed entrances that restart continuing action. Play the current Studio canvas at ordinary speed. If an export was separately requested and already exists, also watch its `pixelReview.evidence.videoPath`; do not create one merely for validation. Evaluate short action → useful result → next event, late-window dead time, repeated tails, focus handoffs and subject-follow against genre; long tracking/micro-scale drift is not development. Use timed frame paths/scene IDs/absolute frames for defects and targeted missing samples. Technical pixel success cannot certify rhythm/expression; unavailable playback stays unverified. Draft frames cannot certify export sharpness: inspect requested native-size output and dense text/texture regions. For initial generation or substantial motion redesign, review Compose's continuous motion audition in Studio before expanding, then replay the completed sequence after measured retiming. Judge the artifact against the storyboard's actual verbs: arrangement changes for organize, original-object bounds change for morph, inputs contribute to synthesis, subject/focus transfers for handoff, and framing follows a moving subject for a follow shot. A tagged entrance with correct timing can pass technical checks while failing expression. Keep the verdict and observed event/time in existing frame notes; unresolved required development keeps the film in draft. No global effect count, camera quota or pixel-activity percentage establishes quality. ## Media, narration and soundtrack - Every planned need has a truthful final outcome (reused/generated/declined/unavailable/failed/intentional editable replacement) with actual file/receipt/provenance. Delivered ordering, script choices and timing agree. Visible images/clips use intended crop/resolution/window and decode correctly; hidden dummy media or URL-only declarations do not count. - Enabled/required narration is real, audible and scene-bound; measured duration/alignment drives beats/captions and dependent timing. Disabled/unavailable speech permits disclosed visual delivery, not a completed narrated result. Timeline nodes/waveforms do not prove audible output. - Finished script has explicit `music_prompt` or deliberate authorized `music_prompt: none`; actual `music_asset` matches the mounted bed and documented source/license/content-fit. Failed sourcing cannot silently become no music. Requested BGM/SFX are separate local music/sfx clips with valid source/trim/start/duration/track/levels and planned visual cues. Muted/missing/wrong-path/out-of-bounds audio fails its requirement. - Source/receipt checks verify references/timing, not licensing, sound waveform, device output, mix quality or synchronization. Listen to narration/music/SFX together and to exported mix when requested; verify spoken phrase/visual-event pairs, effects onset/trailing sound, intelligibility and masking. Music-led events need measured cues, not filenames or decorative waveforms. - Host `audioReview` decode detects missing/silent mix and sample peak warnings. Non-silent BGM can hide absent speech/effects; it is neither speech recognition nor audible-sync proof. Near-full-scale peaks require inspection, not automatic clipping claims. If supported, diagnose masking with same-timeline music-disabled render retaining speech/SFX; no parallel project or compulsory second export for music-free work. Host `temporalReview` measures sampled activity/stillness/cadence; `carrierReview` inspects actual visible carrier identity/geometry and `deterministicSeek` compares direct/reverse state. Inspect required continuity at beat boundaries, contacts/landings, intentional holds and requested shutter output. Diagnostics do not certify meaning, perceived quality, blur or audible phrase sync. Language variants also require source-language identity and real child timing, not only matching root duration. ## Editor, export and publication For delegated runtime review use the exact built-in Studio project: scrub starts/action/Land/seams both ways, play changed scenes, check picture/audio and stop/replay/reopen/theme/variable stability. Preserve unrelated work and the existing server. A moving playhead proves neither picture nor sound. Requested export needs its completed host-render receipt and separate file checks for dimensions, duration, representative native frames, media and audible mix. Intermediate generated assets are process evidence, not final requested deliverables. Publication additionally needs the authorized publisher result/receipt; render preparation or upload submission is not publication success. Do not claim real-model generation, source verification, playback, independent review or exported quality when not observed. ## Artifact-first verdict and bounded repair An available authorized independent reviewer receives only approved brief/script, selected card/variant/source/reference, current Studio preview and timed evidence (or the explicitly requested existing export), without maker verdict/debugging explanations. Otherwise record that review was not independent; never invent or spawn unavailable/unauthorized agents. Report **technical health**, **expression/recipe fit**, and **ordinary-speed audiovisual/export review** separately in existing outputs. Automatic receipt leaves `expression` and `audibleSync` unverified; only observed current-version evidence supports those claims. Failures/unverified claims retain scene ID, frame/time, issue/missing evidence and next action. Separate must-fix, optional and unverified findings. - **Passed:** every applicable required claim has observable supporting evidence. - **Partial:** usable scoped artifact with disclosed declined/unavailable/failed optional capability or review; unresolved required audio/export is an unfinished required scope. - **Failed:** required scene blank, selected component not mounted, unsupported custom exception, invalid timeline/media, undeveloped temporal story, required inaudible audio or unverified requested export. Do not label an unperformed review Passed. Repair source/selection errors in script/recipe, implementation errors in mounted source and sound errors in mix. Preserve prior evidence, obtain current-version evidence and replay whole sequence after retiming to catch shifted cues/seams. Host fingerprint includes scripts/components/styles/assets, excludes generated renders, and source changes invalidate prior receipt. Disclose outside dependencies not covered by fingerprint. After one consolidated repair and one targeted follow-up, remaining must-fix issues keep delivery in draft instead of indefinite polish. ## Host export and publication continuation For requested MP4/export or publication call `ipollowork_extension_call`, `extensionId=media`, `action=video_render_start`, `args={sourcePath:"video//index.html",operationKey:""}`, then `video_render_status` with the same args at returned intervals until complete/failed. Host starts bundled Studio; no guessed versions, CLI installs, manual Export, endpoint exploration, delegated probe agent or schedule/calendar for immediate publishing. Actions remain discoverable outside video-typed chats. Preparing/rendering is progress, not completion; use completed `outputPath`, or report exact failure. Export-only does not authorize publication; edit-only does not authorize export. Reuse the original explicit publish authorization and exact draft/job/key on continuation; never resubmit an uncertain publication. Douyin uses its installed worker import-media → save-draft → publish-draft; WeChat Channels uses its worker with the same verified MP4/account-specific browser. Publisher MCP/session bridge names and bindings come from the runtime contract, never cloud search or an OS app. Publish only within the user's authorized scope; obey the publisher's consequential-action confirmation and approval controls. If publisher returns `browserTask`, continue the same task: claim it, reuse its exact persistent account profile/`extensionId` and upload exact `mediaPath` through host browser upload. API settings/scope gaps can be browser fallback, not manual-upload handoff. Pause only for actual login/QR/SMS/captcha, denied approval or platform error; after submission verify/save receipt before success. Derive account/profile IDs only from installed publisher state and claimed task. Never use an OS social app, ambient profile or ask to reconnect an already verified account.