ClipDeck 0.1.0 Preview 4 — real video editing

Preview 4 improves the first-use workflow and provides a new packaged real-media verification run.

• Transcription starts with clear waiting/progress feedback. Cancellation, retries and replacement transcripts no longer inherit stale completion or cancellation messages; current errors remain visible.
• The extracted app completed real interview import, local recognition, reversed passages, Source/Assembly word and cut navigation, expanded playback, native save/export and reopen. A separate fresh run completed under defined restricted-network controls with an explicitly selected local model.
• Native file pickers verified missing-source recovery and content identity: equal duration does not make a replacement valid. Changed bytes invalidate old transcripts and require cut review before export.

Downloads: use the new app ZIP, its matching source archive and the Chinese getting-started guide. Watch USGS-first-response-final.mp4, a 116.8-second real-camera interview story. Download the unchanged editable three-source/twelve-cut demo to inspect and rearrange that story yourself; keep the extracted project and its originals together. The film/demo retain their historical Preview 3 validation; the new native workflow uses a fresh two-cut interview project. The 50-second walkthrough is historical Preview 2 footage.

Application source: 8b057f9e716861a2f01a6bb3f76d42d75a00e03d. Exact-source CI passed both jobs. Versioned verification and the attached metadata explain performed checks, evidence-attribution corrections and remaining limits. SHA256SUMS binds every supplied asset except itself.

The app bundles local media/Python runtimes and the Chinese punctuation resource. A speech-recognition model requires explicit preparation; recognition errors still need review. This Apple Silicon Mac preview is ad-hoc signed, without Developer ID signing or notarization. If Gatekeeper rejects it, use the exact source build instructions; do not disable system security. Full product acceptance, universal recognition accuracy, human lip-sync certification, macOS 14 execution and other platforms remain unverified. The restricted-network method is not whole-application kernel denial. Private recordings, transcripts and project files are excluded.

中文：这版重点修正“加进素材后下一步不清楚”和转写状态混乱的问题，重新跑通真实视频剪辑、放大预览、导出及重开，并实际检查了素材丢失和错误替换。下载中文说明可直接开始；真实成片和可编辑工程一起提供。识别仍需核对，Apple 开发者签名与公证尚未完成，具体已测范围公开列明。

ClipDeck 0.1.0 Preview 3 — faster verified preparation

Experimental Apple Silicon Mac preview built from b4069a9138997d0207788abfdd2f3b9797553ce3.

First preparation of the included 116.8-second real-camera-footage story took 18.44 / 18.33 / 18.21 seconds in three fresh production service/process/cache runs on the measured Apple M5 Pro Mac, compared with the two 32.75 / 32.81-second baseline observations. The fix requests four threads per validation decoder context and keeps every frame/sample/duration/timestamp/source/cancellation check. These instrumented filesystem-warm service observations are not a universal speed guarantee or click-to-visible benchmark.

This package was actually opened with a native project picker, prepared, navigated using words/cuts in the chosen Source/Assembly mode, expanded, played, paused, exported through a system Save panel and closed normally. Its newly exported 1920×1080/30 fps story validated at 3,504 frames and 5,606,400 audio samples, with bytes identical to the saved demo final. The 60-marker mixed-media reverse-order matrix passed in all three new runs. Exact-source CI passed both jobs.

Download the app ZIP and its matching source archive. The editable demo ZIP contains three unmodified USGS camera-footage originals, the portable project and the previously saved final MP4; its unchanged hash is recorded. The 50-second walkthrough is explicitly historical Preview 2 native capture footage with the same renderer, not a new speed measurement. SHA256SUMS and preview-verification.json bind all supplied artifacts.

The app bundles the media/Python runtimes and local punctuation resource. A Whisper recognition model still requires explicit setup. This is ad-hoc signed, without Developer ID signing or notarization; macOS 14 execution and other platforms are unverified. Use the exact source develop-and-verify instructions as the build alternative. Human listening/lip sync, a fresh large-project renderer campaign and full R01–R18 acceptance remain separate.

中文：这版修复了首次成片预览检查的单线程瓶颈。指定真实示例的三次独立准备为 18.44 / 18.33 / 18.21 秒，原来的两次记录约 32.8 秒；所有完整性检查保留。新程序已实际完成选段定位、放大播放和导出，提供可编辑真实工程。识别模型需明确准备；程序尚无 Apple 开发者签名与公证，人工音画审核和完整产品验收仍未宣称完成。

ClipDeck 0.1.0 Preview 2 — local speech editing

ClipDeck builds a rough cut by selecting words from your own interview, lesson or spoken video, then arranging the passages in the order you choose.

This experimental Apple Silicon Mac preview fixes navigation in the selected Source / Assembly mode. Clicking a bottom cut locates its beginning in that video, including the actual first frame of the next assembled passage. Word navigation keeps playback running or paused. Expanded preview, cut loops, compact-window navigation, project reopening, relinking and an actual native export were checked with real filmed interviews.

Try it

1. Download ClipDeck-0.1.0-preview-ded3e4d-macos-arm64.zip and extract the app.
2. Download ClipDeck-USGS-demo-20261007.zip, keep its folder intact, and open USGS-first-response.clipdeck in ClipDeck. It includes three original USGS recordings, saved transcripts, twelve editable selections and the actual 116.8-second exported film. No new recognition run is needed to edit the supplied project.
3. Watch ClipDeck-desktop-walkthrough.mp4 for 50 seconds of actual desktop actions. It is a silent capture, not an import/transcription speed benchmark.

To edit your own footage, import it and explicitly prepare the selected local speech model before transcription, or use manual time ranges. Projects reference the original files; retain them. Review names, punctuation and timing before using a transcript.

Exact version and verification

The application and matching application-source archive come from ded3e4d. Its unit/native CI run passed. SHA256SUMS identifies the supplied files; preview-verification.json summarizes the bounded native checks. The app retains the application source, media/decoder corresponding source and third-party licenses inside its bundle. The separately pinned development resources remain in the runtime resources release.

The exported example contains 3,504 decoded video frames at 1920 × 1080 / 30 fps and 5,606,400 decoded/presented stereo audio samples at 48 kHz. Structural media checks passed and original file hashes remained unchanged. USGS footage is public domain; logos and burned captions belong to the originals. The footage and interviewees do not endorse ClipDeck.

Known limits

• Ad-hoc signed; no Developer ID signature or notarization. Gatekeeper may reject the download. The project does not ask you to disable security protections; source-build instructions remain available.
• Actual packaged checks ran on macOS 27.0.1 ARM64. The metadata floor is macOS 14; execution there, Intel Macs, Windows and Linux remain unverified.
• First assembly preparation still encodes the cuts and can take tens of seconds. Subsequent cache hits reuse the preview.
• Recognition models require explicit setup and are not bundled. The example reuses saved English transcripts from the earlier packaged recognition campaign; this candidate did not rerun recognition on those three inputs.
• This release does not claim complete R01–R18 acceptance, a human listening/lip-sync review, universal recognition accuracy, a fresh large-project performance benchmark, or a whole-Electron kernel network-denial test.

中文

这是用于访谈、课程和讲话视频粗剪的本地工具：按文字选段，再按自己的顺序拼接。此版已在真实桌面程序里检查“原素材 / 成片”定位、底部片段起点、播放状态保留、放大、循环、保存重开、素材重新关联和实际导出。示例包可以直接打开继续剪，成片与 50 秒真实操作演示一起提供。

仍是实验性预览：尚无开发者签名与公证，第一次成片预览可能需要几十秒，识别结果需要复核。私人视频、项目、账户资料、模型缓存和原始测试日志不在公开文件中。

ClipDeck 0.1.0 runtime resources

Exact macOS ARM64 runtime resources for building ClipDeck from source.

Use the resource list and hashes in packaging/resources.lock.json for your checkout. The current set has five archives: the standalone worker, media runtime, media source, maintained decoder source, and the separately pinned Chinese punctuation weights and vocabulary. Whisper recognition models still require explicit setup. User recordings, private projects and local audit logs are excluded.

The four runtime/source archives retain their original verified bytes. Those native runtime artifacts were checked on macOS 15.7.9 ARM64 and macOS 27.0.1 ARM64. The punctuation archive was added separately, checked against its pinned upstream files, and downloaded anonymously with a matching SHA-256. Its model card, Apache-2.0 terms, FunASR MIT notice and provenance are retained in the source and bundled notices. The 14.0 minimum-OS field remains build metadata; macOS 14 execution is unverified.

This is a resources-only prerelease. An accepted application ZIP and full desktop demo are still pending. Application source, full notices and preparation instructions are in the public repository.