Describer Workbench 0.8.4

Describer Workbench 0.8.4 makes human review optional.

What changed:

• Review is optional. Descriptions that the AI marked as unsure are now spoken in MP4, MP3 and WAV exports. Before, one Needs review description stopped the whole export.
• The Needs review tag stays. Unsure descriptions keep the Needs review tag, the dashed border and the filter. You can still find and check them, but you do not have to.
• New setting: Require review before media export. It is off by default. Turn it on in Settings > Export voice to get the old behavior: any Needs review description blocks MP4, MP3 and WAV export until a person checks it.
• Subtitle exports did not change.

Why: in the 0.8.1 test (13 meeting videos, 105 descriptions graded by hand, gemma4:12b), 67.6 % of descriptions were correct and only 3.8 % were wrong or made up. Most drafts can be used without a person checking each one.

The drafting engine and the narration loudness are the same as 0.8.3. See the 0.8.3 release notes.

Updating on Windows:

• From 0.8.3: use File > Check for updates.
• From 0.8.2: update by hand. Close the app, unpack the Windows ZIP below into a new folder and start the new copy. The 0.8.2 in-app updater fails when the app was opened by double-click.

Known limits:

• A description without the Needs review tag can still be wrong. The tag only marks what the AI said it was unsure about.
• The model can still call something a vote when there is none.
• Results change from run to run on the same video.
• Preview playback in the app still uses the old volume. Only exports match the video loudness.

Downloads:

• Windows x64 portable ZIP: signed as Scott Converse with Azure Trusted Signing, timestamped, and checked with Authenticode and signtool.
• macOS Apple Silicon DMG: ad-hoc integrity signature only. No Apple publisher certificate or notarization.
• Linux x64 DEB and portable archive: unsigned.
• unix-media-sources-arm64.tar.gz: source code for the media tools inside the Mac app.
• Python wheel, source package and offline documentation.

All files come from commit 22948ea. See SHA256SUMS.txt and release-verification.json for file hashes and provenance.

Describer Workbench 0.8.3

Describer Workbench 0.8.3 fixes the narration loudness and two Windows updater bugs.

What is fixed:

• Narration loudness. Exports now set each spoken description to the loudness of the video's own sound in the 30 seconds around it. A quiet room in a long meeting gets a quiet voice. Before, the voice on a quiet meeting video could be up to 16 dB louder than the room. Speech volume is now a trim on top of that match.
• No clipping. A limiter keeps finished exports below -1 dBFS (full scale). The mix steps use float audio files, so nothing clips before the limiter runs.
• Updater: install folder lock. The update helper now leaves the install folder before it swaps the folders. Before, the swap could fail with "being used by another process". The app then put the old version back.
• Updater: pending download. A download that waited for a busy job now starts as soon as that job ends. Before, it could wait for an unrelated job.

Updating from 0.8.2 on Windows:

• Best way: update by hand. Close the app. Download the Windows ZIP below. Unpack it into a new folder. Start the new copy. Your settings and projects stay, because they are not in the app folder.
• The in-app update in 0.8.2 has the two bugs that 0.8.3 fixes. For this one update, the old 0.8.2 updater runs:
  • If you opened the app by double-click (the usual way), the update cannot replace the app folder. The app closes and does not start again. Your 0.8.2 copy is not changed. Update by hand. You can delete the folder DescriberWorkbench.incoming-<number> next to the app folder.
  • If Check for updates waits and does nothing, choose Check for updates again and click Cancel in the second folder picker.
• From 0.8.3 on, the in-app update works when you open the app by double-click.

The drafting engine is the same as 0.8.1. The measured drafting quality is the same. See the 0.8.1 release notes.

Known limits:

• Preview playback in the app still uses the old volume. Only exports match the video loudness.
• The model can still call something a vote when there is none.
• Results change from run to run on the same video.
• AI output is a draft. Check every draft before you export.

Downloads:

• Windows x64 portable ZIP: signed as Scott Converse with Azure Trusted Signing, timestamped, and checked with Authenticode and signtool.
• macOS Apple Silicon DMG: ad-hoc integrity signature only. No Apple publisher certificate or notarization.
• Linux x64 DEB and portable archive: unsigned.
• unix-media-sources-arm64.tar.gz: source code for the media tools inside the Mac app.
• Python wheel, source package and offline documentation.

All files come from commit 42c3e38, except the documentation ZIP: it has the fixed update advice from commit b87afcc. See SHA256SUMS.txt and release-verification.json for file hashes and provenance.

Describer Workbench 0.8.2

Describer Workbench 0.8.2 lets you turn on Public meeting mode and choose a captions file in the app window. Before, you had to edit settings.json for both.

What is new:

• Draft with AI has a Public meeting mode box. Tick it for city, council and board meeting videos. The app keeps your choice.
• Draft with AI has a Captions file for this video row with Choose… and Clear. Pick the SRT or VTT file for the same video.
• A captions file chosen in the window is used for that video only. When you open another video, the app clears it.
• A captions path typed by hand in settings.json is kept for every video, as before.
• If the chosen captions file is gone, Start draft stops and tells you.

The drafting engine is the same as 0.8.1. So the measured quality is the same:

Measure	0.8.1 result	Target
Correct	67.6 %	60 % or more
Wrong or made up	3.8 %	5 % or less
Trivial (true, but adds nothing)	10.5 %	15 % or less
Spoken in a gap, or a pause for a meeting event	100 %	70 % or more
Extra length added (meetings)	+1.6 s per minute	under 15 s per minute

Measured on 13 test videos, 105 cues graded by hand, model gemma4:12b. See the 0.8.1 release notes.

Known limits:

• The model can still call something a vote when there is none.
• Results change from run to run on the same video.
• The new labels show in English in the other app languages.
• AI output is a draft. Check every draft before you export.

Downloads:

• Windows x64 portable ZIP: signed as Scott Converse with Azure Trusted Signing, timestamped, and checked with Authenticode and signtool.
• macOS Apple Silicon DMG: ad-hoc integrity signature only. No Apple publisher certificate or notarization.
• Linux x64 DEB and portable archive: unsigned.
• unix-media-sources-arm64.tar.gz: source code for the media tools inside the Mac app.
• Python wheel, source package and offline documentation.

All files come from the same commit. See SHA256SUMS.txt and release-verification.json for file hashes and provenance.

Describer Workbench 0.8.1

Describer Workbench 0.8.1 makes Public meeting mode (city and council meeting videos) write fewer, better cues.

What is new:

• The model is told to skip small body moves (a nod, a hand on the table).
• A filter drops cues that only say a small move or "nothing changed".
• Each cue gets a second look: the model checks it against the picture and drops it if it is not true.
• When a slide cue and a cue about a person overlap, the slide cue is kept.
• Only meeting events (votes, speaker changes, slides, on-screen text) may pause the video.
• A blank row from the model no longer breaks a whole chunk.
• A vote found from the captions is not dropped by the second look.

Quality on 13 test videos (105 cues graded by hand, model gemma4:12b):

Measure	0.8.0	0.8.1	Target
Correct	29.5 %	67.6 %	60 % or more
Wrong or made up	15.4 %	3.8 %	5 % or less
Trivial (true, but adds nothing)	36.5 %	10.5 %	15 % or less
Spoken in a gap, or a pause for a meeting event	45 % in a gap	100 %	70 % or more
Extra length added (meetings)	+6.0 s per minute	+1.6 s per minute	under 15 s per minute

The "spoken" target now counts a cue that pauses the video for a meeting event. A pause cue never talks over a speaker.

Known limits:

• The model can still call something a vote when there is none. On one test video it said "An audience member raises a hand." and no hand was up.
• Results change from run to run on the same video.
• Some videos get very few cues (2 in 5 minutes on one test video).
• AI output is a draft. Check every draft before you export.

Downloads:

• Windows x64 portable ZIP: signed as Scott Converse with Azure Trusted Signing, timestamped, and checked with Authenticode and signtool.
• macOS Apple Silicon DMG: ad-hoc integrity signature only. No Apple publisher certificate or notarization.
• Linux x64 DEB and portable archive: unsigned.
• unix-media-sources-arm64.tar.gz: source code for the media tools inside the Mac app.
• Python wheel, source package and offline documentation.

The Mac and Linux files were added after the Windows release. They come from commit 7a2c402 (PR 58). That commit only fixes the version number in the Mac and Linux packaging scripts. The app code is the same as the Windows build.

See SHA256SUMS.txt and release-verification.json for file hashes and provenance.

Describer Workbench 0.8.0 (Windows)

Describer Workbench 0.8.0 is a Windows release. It improves Public meeting mode (city and council meeting videos).

What is new:

• Meeting mode describes only what changed on screen. It no longer repeats the same cue.
• It reads slide titles and on-screen text.
• It finds votes, such as a show of hands.
• It places spoken cues in the pauses between talk more often.
• It never speaks internal labels such as "Speaker 1" or "man_plaid".
• Long descriptions are not cut in the middle of a phrase.
• A slide cue starts when the slide shows, not before.
• Fixes for F5, F6, F7, F8 and F10 from the audit (each with a test).

Quality is still below its targets. On 13 test videos (156 cues graded by hand, model gemma4:12b):

Measure	Now	Target
Correct	29.5 %	60 % or more
Wrong or made up	15.4 %	5 % or less
Trivial (true, but adds nothing)	36.5 %	15 % or less
Spoken in a pause	45 %	70 % or more
Same cue said again	0 %	under 5 %
Internal label spoken	0 %	0
Extra length added (meetings)	+6.0 s per minute	under 15 s per minute

So check every draft before you export. AI output is a draft.

Downloads:

• Windows x64 portable ZIP: signed as Scott Converse with Azure Trusted Signing, timestamped, and checked with Authenticode and signtool.
• Python wheel, source package and offline documentation.

Mac and Linux stay on 0.7.0. There is no 0.8.0 Mac or Linux build. Use the 0.7.0 release on those computers.

See SHA256SUMS.txt and release-verification.json for file hashes and provenance.

Describer Workbench 0.7.0

Describer Workbench 0.7.0 contains the eight approved stabilization packages while preserving existing features, providers, models, accessibility routes, export formats and safety protections.

Downloads now available:

• Windows x64 portable ZIP: Azure Trusted Signed as Scott Converse, timestamped and independently verified.
• macOS Apple Silicon DMG: ad-hoc integrity signature; no Apple publisher certificate or notarization.
• Linux x64 DEB and portable archive: unsigned.
• Python wheel, source distribution and offline documentation.

There is no Intel Mac DMG for 0.7.0. Its native build did not complete (one timing-sensitive test failed on the Intel runner, then the test step reached its time limit). Intel Mac users can install from the repository with tools/install_unix.py (see the platform guide) (this route was not tested on Intel for 0.7.0) or use a later release.

Installation and manuals. See SHA256SUMS.txt and release-verification.json for artifact hashes and provenance. AI output remains a reviewed draft; check facts, missed events, text and timing before export. Existing provider entitlement limits remain documented.

Describer Workbench 0.6.1 - macOS and Linux add-ons

Describer Workbench 0.6.1 — macOS and Linux add-ons

Additional native distributions of Describer Workbench 0.6.1. The signed Windows 0.6.1 executables and application version remain unchanged; Windows remains the default stable release. External documentation has been refreshed across distributions. These packages have a separate source tag, v0.6.1-unix.1; they do not replace the Windows release.

• Apple Silicon Mac: download DescriberWorkbench-0.6.1-macos-arm64.dmg.
• Intel Mac: download DescriberWorkbench-0.6.1-macos-x64.dmg.
• Ubuntu 24.04 x86-64: download DescriberWorkbench-0.6.1-linux-x64.deb, then run sudo apt install ./DescriberWorkbench-0.6.1-linux-x64.deb.
• Other Linux desktops: the tar archive requires equivalent system dependencies; compatibility outside the tested distribution is not guaranteed.
• Repository installation: clone this tag, install the platform prerequisites in the installation guide, then run python3 tools/install_unix.py.

macOS targets version 15 or later. Open the DMG and drag the app to Applications. No Apple developer certificate or notarization is supplied. The app has local ad-hoc integrity signatures, which do not verify a publisher. Use the normal macOS Privacy & Security approval if required; do not disable Gatekeeper globally. Linux packages are unsigned. Compare all downloads with SHA256SUMS.txt.

The add-ons provide native measured narration through macOS voices or eSpeak NG, native secure credential storage through Keychain or Secret Service, per-user data locations, desktop launchers, and the existing shared engine, CLI, HTTP and MCP interfaces. Unix updates are manual. Existing AI routes remain subject to the external CLI/service being installed, signed in, reachable and entitled on the destination computer; the release does not bundle subscriptions or model weights.

Use the platform verification receipt for exact package/source identity, native installation checks and test evidence. CI does not establish human listening quality, VoiceOver/Orca usability or live subscription entitlement. Factual descriptions remain reviewed drafts with the existing documented accuracy limits.

Mac FFmpeg/x264/LAME corresponding source, build scripts and GPL/LGPL notices are supplied in the matching unix-media-sources-*.tar.gz assets. Linux media tools come from the distribution's packages. The wheel/source archive are Python distributions; they are separate from the native desktop installers.

Installation and removal · Manual · Unchanged signed Windows release

Documentation refreshed November 5, 2026 across the README, nontechnical quick start, full user and technical manuals, architecture drawings, landing page and both hosted sites. All native packages now contain these current guides, identified by documentation-manifest.json. Download DescriberWorkbench-0.6.1-documentation.zip for standalone offline HTML/Markdown/drawings.

This refresh changes documentation and corrects the Mac bundle launch entry point, with no app/media rebuild or version change. Linux preserves every other payload byte/mode. Mac Info.plist now launches DescriberWorkbench rather than DescriberCLI; all other plist fields, application code and nested payloads are preserved. The resource envelope is resealed with ad-hoc integrity only. Strict codesign, CLI startup and a real LaunchServices app-bundle launch/normal quit are checked. No Apple publisher certificate or notarization is added. Package/archive checksums changed; use the refreshed SHA256SUMS.txt. documentation-refresh.json records old/new package hashes and preservation; historical native acceptance remains separately identified. Windows executable signing, installed runtime/settings and desktop launcher are preserved.

Describer Workbench 0.6.1

Version 0.6.1 fixes all eight findings from the 0.6.0 whole-repository audit-lite review:

• Reject malformed observation metadata as a whole chunk, avoiding successful results that silently lose descriptions.
• Retry rejected chunks with a fresh provider request while retaining validated prior chunks and failure evidence.
• Restore visible subscription/API connection tests, exact typed model IDs and accurate transmission disclosure.
• Recover safely from malformed nested settings while retaining valid existing connection profiles.
• Retain and display available Ollama model family, parameter size, quantization, size and context metadata.
• Correct model-entry, timing-edit and speech-volume instructions throughout maintained documentation.
• Import valid VTT timestamps with optional hours.
• Rename prompt presets explicitly without creating duplicates.

Windows GUI and CLI are signed as Scott Converse through Azure Trusted Signing with timestamps. Compare downloads with SHA256SUMS.txt; release-verification.json records exact source identity, signing, tests and delivery verification. The portable ZIP includes FFmpeg, manuals and notices; Ollama/models/media remain separate.

Source and required Windows CI: 182 tests and 3 subtests passed. Packaged GUI startup, CLI, authenticated HTTP/OpenAPI and 17 MCP tools passed. New deterministic provider regressions exercise real normalization, checkpoint reuse, job completion, Qt dialogs and settings persistence with external boundaries mocked. Existing live-provider evidence and limits remain documented in the acceptance and connection guides.

Descriptions remain reviewed drafts. Correct facts, omitted events, printed text and timing before final export. The accepted measured Windows narration policy remains unchanged. Windows acceptance does not establish full-film factuality, active screen-reader certification or every-model entitlement. Native Mac/Linux downloads and their own evidence are published separately in the add-on release.

Documentation refreshed November 5, 2026: all-platform README, nontechnical quick start, full user and technical manuals, architecture drawings and landing page. Download DescriberWorkbench-0.6.1-documentation.zip for offline guides, or use the current website. The portable ZIP and Python archives now include current documentation; both Windows executables, application code, version and publisher signatures are unchanged. Package checksums changed. documentation-refresh.json records original/new archive hashes, documentation source and preservation checks; original application acceptance remains historical in release-verification.json. Existing installed copies can use the online/offline guides without replacing the application.

Describer Workbench0.6.0

Version0.6.0 completes the bounded22-item remediation: subscription gateway plus native Codex/Claude/Antigravity tools, explicit compatible/API connections, authored cue timing, incomplete-response recovery, measured Windows timed narration, corrected gain, safe shortening, original nonsecret settings, durable progress/usage and redesign controls.

The GUI and CLI were Azure Trusted Signed as Scott Converse. Both independently passed Authenticode and signtool verification with Microsoft timestamps. Unchanged signed-build source: ba04cc3. Later commits refresh documentation only.

Descriptions are reviewed drafts. Correct visual facts, omissions, printed text and timing before export. The360-second comparison retains factual errors and omissions. Native and gateway subscription routes have bounded image-backed description/Scene checks; this does not establish every listed model entitlement. Gateway Gemini reported a different canonical ID from its high alias; native Codex/Antigravity did not report actual model identity. Separate keyed APIs and local compatible servers retain documented live-testing limits.

Extract the entire Windows ZIP and retain _internal. Native tools, sign-in and model weights are separate. See the included manuals, docs/AI-CONNECTIONS.md and docs/V060-ACCEPTANCE.md. Download SHA256SUMS.txt and release-verification.json for artifact and delivery evidence. Preserve existing projects, source media and rollback copies.

Delivery status: signed application and installed desktop launcher verified0.6.0; settings unchanged and rollback retained. Both hosted sites and all maintained/generated/packaged/installed documentation surfaces verified. Main/tag Windows CI and GitHubPages deployment passed. See release-verification.json.

Describer Workbench 0.5.2

Describer Workbench 0.5.2

This release fixes the 25 bounded first-use and recovery findings across the Windows workbench and shared service.

• Preserve review evidence, uncertainty and visual-event identity while editing.
• Connect explicit pinned speech-detector setup to Draft, with verified continuation and cancellation.
• Keep exact model selection and cancelled Draft settings consistent; run prerequisite checks without blocking the UI.
• Validate narration before vision and obtain consent for each actual external destination.
• Protect typed export paths, handle unavailable voices truthfully, and keep destination errors in the dialog.
• Repair missing narration and resume saved interrupted jobs from the desktop while retaining valid checkpoints.
• Isolate damaged job records, clear stale Scene state and report actual API startup readiness.

Native acceptance covered real Windows speech repair/discovery, detector download/validation, typed overwrite refusal and recovery across an app restart. Prior completed speech checkpoint files retained their hashes and modification times. The source video remained unchanged. See the release-verification.json asset and docs/V052-ACCEPTANCE.md for provenance and limits.

Download the Windows ZIP, extract the whole folder, and launch DescriberWorkbench.exe. Keep _internal beside it. The GUI and CLI are signed as Scott Converse using Azure Trusted Signing with Microsoft timestamps. User settings/projects are separate from the program directory; preserve your old version as rollback. Older jobs may reject resume when implementation identity differs, preserving their files.

No new claim is made for live paid/cloud inference, full-film AD accuracy, screen-reader certification, clean-machine installation or public YouTube extraction. AI descriptions still require editorial review.

Signing correction, 2026-11-02: the initially unsigned Windows ZIP was replaced by the same application with verified publisher signatures. The v0.5.2 source tag is unchanged. Download the refreshed ZIP and checksum manifest.