# Engineering demonstration media Updated 20 October 2026. All people, dialogue and project details are fictional. No customer meeting, cloned personal voice or visitor recording is included. ## Clean website sample `conversation.mp3` is a 64.88-second English edit of the normal-difficulty `normal-079` engineering release-planning case. Seven existing Gemini 3.8 Flash TTS clips (T0025, T0029, T0033, T0035, T0037, T0039, T0041) are joined with 280 ms pauses. There is no added static, packet loss or overlap. This is an edited sample, not an uninterrupted excerpt from the original bilingual meeting. The fictional cast is Maya, Leila, Jonas and Nina. The action is **Maya's revised Orion 2 release plan, due 28 June**. That is not an approved launch date. The proposal to skip review is challenged; additional support remains undecided. `prepare_demo.py` uses cached synthesis only. It encodes mono 80 kbps MP3 and aligns words locally with the existing Omni CTC checkpoint. `demo.json` retains source and clip hashes, with no credentials or local paths. Waveforms are RMS envelopes of the actual audio. Names and prepared notes come from the synthetic reference, not an assertion of automatic speaker-name recognition. ## Multilingual stress-test film The separate 16-second stress film (`stress.mp4`, `stress-portrait.mp4`) plays source seconds 23.70–36.15 of the existing `very-hard-natural-preview` mixed recording: English and Hindi, simulated remote packet loss, background noise and brief heated cross-talk. The audio comes from `meeting.wav`, not from clean voice clips, and starts 0.5 s into the film. Its only edits are 40 ms/80 ms fades at the cut points, music underneath at reduced level, and the final loudness gain. It is separate from the clean website listening sample and no longer appears in the core film. `video/prepare_stress.py` copies the **saved Qwen3-ASR 1.7B output** for that exact interval. Text and speaker labels are not corrected to match the reference. The final word stays **Unassigned**, and the film points to it ("When it isn't sure who spoke, it says so."). Rows appear at their saved times. This is a replay of an existing result, not live inference or a claim of accuracy. The film says so on screen: "Saved model output, replayed. Accuracy varies by language and recording." Source audio and transcript hashes are kept in the generated `stress.json`. The English text under the Hindi row is **not model output**. It is a gloss written for viewers of this film and labelled "Translation for this film" on screen and "[English translation]" in `stress.vtt`. The saved output has row timing only. The captions therefore split the long Hindi row at its own sentence ends with proportional cue times, and pair each part with its gloss. All other caption rows keep their saved text and times. The stress film makes no language-count claim. Qwen3-ASR's supported languages are listed on its model card: https://huggingface.co/Qwen/Qwen3-ASR-1.7B ## Saved notes shown beside the sample The "Overtalk's saved notes" panel is the app's own unedited output, not hand-written. It comes from a local evaluation run (`fast-full200-v2`) over the original 6½-minute English–Arabic recording of `normal-079`: speech by Qwen3-ASR 1.7B, notes by Qwen3.5 4B, on one computer. `prepare_output.py` copies the Orion Two decision, Maya's action and the open question verbatim into `app-output.json`, with the notes file's SHA-256, and links each cited transcript line to the turn of the English sample that starts inside it (04:50 → 00:29, 04:59 → 00:38, 05:23 → 00:45, 05:39 → 00:57). The decision's sources were spoken in Arabic and are not in the sample; the page shows the saved Arabic transcript line with an English translation written for the site. "Orion Two" and the euro ceiling come from that original meeting. The same saved notes also cover a second release, where one action has no named owner; Overtalk wrote "Not stated" rather than guessing. ## Native screenshots and animated films `make_native_fixture.py` derives a disposable fixture from the clean public sample. Real native app captures (2× density) show the same meeting, people and project. The project also lists two earlier fictional meetings, "Release readiness check" (17 Oct) and "Payment retry investigation" (18 Oct), which exist only as list entries with no audio or transcript. The notes inside these captures come from the fixture's prepared summary, not from a model run; the page says so. Captures never open or alter the user's database. The project screens (`coming-up*.webp`, `brief*.webp`) are the same native app with the fixture's calendar section: three upcoming events with their filing reasons and one brief for the Orion 2 release review. The brief's lines and talking points are written into the fixture to match the sample's notes; no model wrote them. Times are rendered with `TZ=UTC`. ## Illustrations and share image The route and project archive illustrations were generated with Codex's built-in image generation from one approved style reference (`art/route-local.png`) and resized to WebP. Each prompt is stored in `art/prompts.json` and beside each shipped file. `og.png` combines generated artwork with HTML-set type and the Overtalk mark. The illustrations depict ideas (a local laptop, a sealed EU worker, the same sealed worker with the subscriber's own key in its lock, a project drawer); they are not product screenshots. The 76.5-second demo film is an authored animation of the product interface, not a screen recording. It and the stress film each have 1920×1080 and 1080×1350 H.264/AAC exports at 30 fps (yuv420p, BT.709). No captions are burned in. Speech captions ship only as WebVTT (`promo.vtt`, `stress.vtt`), with speech-only cues that follow the trimmed audio. The films contain no buttons or URLs. - **Demo film** (`promo.mp4`, `promo-portrait.mp4`; label: "Fictional meeting · Synthetic voices"). It follows the fictional "Orion 2: release plan" meeting of 19 Oct in the order it is lived, in the free, open-source app, and names no price or subscription. - *Before the meeting.* The calendar event is filed into Orion 2 with the reasons the app's filing matcher writes for the fictional evidence on screen; a meeting with too little evidence becomes a suggestion rather than being filed. The brief uses the app's sections for the second meeting in a series. Unlike the native brief captures (for the 21 Oct review), its lines are written for the film from the fictional 17 Oct "Release readiness check" (target 24 June if the missing information arrives early) and the 18 Oct ticket ORI-131. That meeting has no audio, so these lines cite notes, not a playable moment. - *The meeting.* It plays trimmed parts of the cached clips in meeting order: Leila 00:11 ("If we call this finished… cannot support."), Maya 00:29 (first sentence) and Leila's full 00:38 correction. Jonas's 00:19 question stays in the transcript but is not played. - *After it.* A notes chip replays "It is Maya, not Jonas." The notes paraphrase the app's saved output for the full meeting: an Action for Maya due 28 June (sources 00:29 and 00:38) and a decision that the release stays conditional on passing the review. The film links that condition to Leila's 00:11 line; the saved notes cite passages spoken in Arabic that are not in the English sample. A ticket draft is copied, not sent. Two later calls without an invite are filed by the same matcher, one only after the user accepts the suggestion. - *Later.* The project answer cites two dated meetings. "Release readiness check · 17 Oct" appears only as a text quote. "Release plan · 19 Oct" plays Maya's cached 00:29 sentence again when its source chip is clicked. Remembering Maya's voice shows the app's consent step: "They agreed in this meeting" and the checkbox "They explicitly agreed to Overtalk recognising their voice on this computer". - *Privacy.* Three routes: on this computer (default; "No bot joins the call"), the user's own Privatemode account, and optional private AI summaries, marked **Planned** ("Advanced private AI models, encrypted on your computer with your own key", adding "Overtalk never sees the text"). That describes the planned end-to-end route: Privatemode's own proxy runs on the user's computer with a key issued to that user, so Overtalk's account service handles accounts and keys, never meeting text. The film does not claim that this service is live. - It closes on "Free and open source." and "Everything here runs on your computer by default." The films use **The Vibe by Punch Deck**, the artist's 2023 recording, under **CC BY 4.0**. It is not AI-generated music. Official song: https://punchdeck.bandcamp.com/track/the-vibe License: https://creativecommons.org/licenses/by/4.0/ Edits: source seconds 35–95, a 0.3 s entry fade, a 2 s closing fade, level adjustment and smooth speech ducking (0.25 s attack, 0.75 s release), with extra attenuation under the noisy stress recording. The demo film is longer than the excerpt, so one 16.744 s passage repeats once at a matching beat, with an 80 ms crossfade; nothing else is looped. Final mastering applies one linear gain per film (target −16 LUFS, −1.5 dBTP), with a look-ahead limiter on the few loudest peaks where needed, so the closing fade is preserved. Credit appears beside the player and in each video's metadata. Retain the title, artist, song link, license and edits when reposting. Detailed source and excerpt SHA-256 hashes are in the repository's `video/assets/MUSIC-LICENSE.md`. The films use OpenDesign Motion and Claude Motion guidance with the existing Remotion renderer. Their upstream implementation is not copied. Fonts, logo and connection marks keep their notices. No model weights are distributed. No new paid generation, evaluation rerun or user-data mutation was performed.