{ "schemaVersion": 1, "batchId": "pseo-batch2-20261014", "blueprintId": "blueprint-33", "kind": "blog", "slug": "video-demo-evidence-audit", "locale": "en", "evidenceClass": "observational-audit", "sourcePath": "content/pseo/batch2-20261014/blog/video-demo-evidence-audit.json", "sourceHash": "1f0d6df1a1a61b6cc3de4ad9a0ddc7cb8949b8aa43522e4f2ee5c117c78c5200", "contentHash": "8069a8e2acc24981b8c168d5d4fc1274870837664c2e96016824b4a447b7a20c", "contentVersion": "authored-8069a8e2acc24981b8c168d5", "rendererHash": "727572c29f88632162b30bf97d95dc1be416eb61ca002023ad1c1b1410d29fbd", "copy": { "title": "How to read AI video examples: an evidence audit of four fixed records", "description": "Inspect two MiniMax H3 source demos, an original practice and a local execution fixture to separate displayed clips, input completeness, interface tests and model-effect evidence.", "goal": "Determine which conclusions a displayed example supports before using it to choose a production method.", "assetCaption": "The cover is generated editorial artwork; the actual audited records and observations are identified below.", "preparation": [], "steps": [], "acceptance": [], "troubleshooting": [], "ending": "", "sections": [ { "title": "Define evidence before comparing demos", "paragraphs": [ "This audit fixes four records: the late-night laundromat example, the telescope fashion example, our reflective-bottle practice, and the task-resume recipe. The first two are source demonstrations, the third is an original teaching method, and the fourth is a local interface fixture. We deliberately chose different evidence levels. Four purposive records are neither a random sample nor a model leaderboard.", "We ask six separate questions: can the source be traced, can inputs be inspected, is the prompt visible, is the submitted request disclosed, can an output be tied to that request, and are all attempts recorded? “Not disclosed” means absent from the selected record we inspected. It does not mean the author never retained it, and unknown failures cannot become zero failures.", "An attractive clip answers what one displayed result looks like. A connected input and execution record starts to answer whether someone else can inspect how it was produced. Both are useful questions, but they require different evidence." ] }, { "title": "What each of the four records actually supports", "paragraphs": [ "In the built-in browser, the laundromat player reported 15.104 seconds and 1280×720 pixels. The visible prompt requests 15 seconds and 16:9, while the parameter panel labels the aspect ratio unspecified. These values come from different layers: the player describes a file, the prompt expresses an instruction, and accepted generation settings require a request record. Approximate agreement between the first two does not supply the missing third layer.", "The telescope page lists four reference inputs, and its player also reported 15.104 seconds and 1280×720. Linked inputs make the disclosed materials easier to inspect. The selected public record did not disclose a request ID, accepted parameters, all candidate outputs, or an editing log. Four linked images do not establish how the underlying system encoded, ordered, or reused those images." ], "table": { "headers": [ "Fixed record", "Directly inspectable", "Not established here", "Useful scope" ], "rows": [ [ "Laundromat source demo", "Visible prompt; player duration and dimensions", "Raw request, attempt denominator, independent reproduction", "Observe one publicly displayed file" ], [ "Telescope source demo", "Four input links; prompt and displayed file", "Submitted input roles, run request and selection history", "Inspect disclosed material completeness" ], [ "Reflective-bottle practice", "Original input, prompt and acceptance method", "Any generated result or success rate", "Learn an action-planning method" ], [ "Local task-resume fixture", "24 executed scenarios; code and receipt", "Provider video output, live completion or billing", "Inspect polling and download behavior" ] ] } }, { "title": "Complete materials do not license broader claims", "paragraphs": [ "The bottle practice provides an illustration, a reusable prompt and review criteria. Its page explicitly says model output has not been independently tested. Complete teaching materials help a reader begin; they do not prove that reflections, trademarks or real photos will remain faithful. Calling a downloadable method a verified effect would cross that boundary.", "The task-resume recipe did execute 24 local interface scenarios. Its receipt also says the download fixture is only 64 synthetic bytes with an MP4 header: it was not decoded as a real video, and no provider generation task was submitted. That supports particular status branches, file protections and download checks. It does not establish model speed, visual quality or production delivery success.", "Nor does a polished source demo supply a success rate. If only a chosen output is shown, we do not know the number of attempts, abandoned inputs or subsequent edits. A percentage without its denominator is speculation. We make no assertion about the sample authors’ actual private workflows and no ranking between models." ] }, { "title": "Make the next demonstration inspectable", "paragraphs": [ "For a future production, retain versions and hashes for input files, the prompt, accepted parameters and final output, then connect them with a request or task ID. Remove keys and privileged download links from public records. A hash establishes file identity, not visual correctness; it does not replace watching and reviewing the result.", "A reproduction claim needs an actual rerun of the same inputs and settings, with differences recorded. A stability claim needs predefined acceptance criteria, every attempt recorded and the sample selection described. Cropping, composited text and result selection belong in the delivery history instead of being attributed silently to the model.", "This article does not reproduce the templates’ operating instructions. The downloadable audit JSON and CSV identify our selected records, known fields and remaining unknowns. Follow the technical recipes or delivery workflow for execution steps, keeping evidence analysis and task execution clear in their respective roles." ] } ] }, "assets": [], "artifacts": [ { "path": "/learning-assets/pseo-batch2-20261014/video-demo-evidence-audit/audit.json", "label": { "en": "Fixed-sample audit JSON", "zh": "固定样本审计 JSON" } }, { "path": "/learning-assets/pseo-batch2-20261014/video-demo-evidence-audit/audit.csv", "label": { "en": "Evidence-scope matrix CSV", "zh": "证据范围矩阵 CSV" } }, { "path": "/learning-assets/pseo-batch2-20261014/video-demo-evidence-audit/browser-observations.json", "label": { "en": "Actual browser metadata observations", "zh": "实际浏览器参数观察" } } ], "execution": null, "hero": { "imagePath": "/learning-assets/pseo-batch2-20261014/video-demo-evidence-audit/editorial-cover.png", "label": { "en": "Symbolic editorial artwork for an evidence-chain audit", "zh": "证据链审计主题插画" } }, "sources": [ { "title": "Selected source catalog record (pinned revision)", "url": "https://github.com/AtlasCloudAI/awesome-minimax-h3-prompts/blob/ff00e714bf2f972ac9563f9ecfa0e1c3c0b325da/data/prompts.json" }, { "title": "Laundromat public example", "url": "https://minimaxh3.studio/zh/prompts/minimax-h3-late-night-laundromat-glow-animation-mix" }, { "title": "Telescope public example", "url": "https://minimaxh3.studio/zh/prompts/minimax-h3-telescope-peek-at-minimax-fashion" }, { "title": "Original bottle practice", "url": "https://minimaxh3.studio/zh/templates/reflective-bottle-light-sweep" }, { "title": "Executed local task-resume fixture", "url": "https://minimaxh3.studio/zh/recipes/async-task-resume-poll-download" } ], "shots": [], "demonstration": null }