THE GHOST PROTOCOL // AI Show Runner & Studio

Single-file, zero-dependency (Python 3 stdlib only) production studio for the AI debate show. Dark terminal aesthetic, four show formats, human banter injection, live transcript editing, and one-click export to a fully open-source Piper + FFmpeg render pipeline.

Run

python3 ghost_protocol_studio.py
# open http://localhost:7860

No pip installs. Env overrides: GP_PORT, GP_ENDPOINT, GP_API_KEY, GP_MODEL.

Backend

Defaults to your Bifrost gateway at http://localhost:8080/v1/chat/completions with model ollama/qwen3:14b. Any OpenAI-compatible endpoint works — change it in the UI (03 // ENGINE) or via env vars. Free-tier direct options if Bifrost is down:

• Groq: https://api.groq.com/openai/v1/chat/completions, model e.g. llama-3.3-70b-versatile
• Gemini: https://generativelanguage.googleapis.com/v1beta/openai/chat/completions, model gemini-2.0-flash

Put the key in the API KEY field (or leave blank if Bifrost holds keys). TEST CONNECTION pings <base>/models.

Simulation Mode (checkbox) runs the whole flow with canned lines and zero API calls — use it to learn the UI and test exports.

The four formats

1. Socratic Stress-Test — SUBROUTINE_ALPHA (interrogator) vs SUBROUTINE_BETA (defender), N rounds, SYSTEM_KERNEL verdict with fallacy detection.
2. Round Table — PROTOCOL-7 (deadpan literalist), M.I.R-A (techno-optimist), CYNIC.EXE (doom-poster) discuss the topic; after each round the run pauses and each human co-host (set names in 01 // FORMAT) types banter that feeds the next AI turns. Kernel closes with "Best Line of the Night."
3. Grand Tribunal — AUDIT-9 prosecutes the thesis, ADVOCATE-0 defends, SUBJECT-X testifies; openings, N examination rounds, closings, Kernel ruling, then a human jury verdict prompt. Optionally CALL A STAR WITNESS (01 // FORMAT, tribunal only) — name a witness and give them a background, and they're called to the stand after the accused's examination: prosecution questions them, defense cross-examines. Fully AI-voiced and editable in 02 // CAST like any other speaker.
4. Triad Synthesis — STOIC-1, NIHIL-0, UTIL-3 argue N rounds; SYNTHESIS_CORE forges the compromise position.

Every cast member's name, system prompt, model override, and Piper voice is editable in 02 // CAST. Every transcript bubble is click-to-edit; RE-ROLL regenerates a line in context; CUT deletes it.

Character registry (characters.json)

All 13 unique cast members (some reused across formats, like SYSTEM_KERNEL) have a full profile in characters.json: a debate archetype (Interrogator, Logician, Skeptic, Synthesizer, ...), a distinct comedic voice, a transparent cutout portrait, and — the big one — a genuinely unique AI model each, pulled only from what's actually working through Bifrost right now (Ollama + Groq; vllm/* is deliberately excluded — see CLAUDE.md, it's paid-subscription-backed and "not a default for any agent"). The app fetches this at boot (/api/characters) and pre-fills each cast card's MODEL OVERRIDE, portrait thumbnail, and archetype badge — still fully editable per-run in 02 // CAST. characters.json is the design-intent source of truth; CASTS in the app is what actually runs, kept in sync by hand.

◈ CAST POOL (header button) — browse the full 13-character roster across every format as a portrait gallery. Click any character to open their soul: a proper identity doc (characters/<sid>.md, styled after this platform's own SOUL.md convention — identity statement, values, voice, boundaries) rendered right in the UI. Same "view soul ›" link is available inline on any cast card in 02 // CAST.

Voices (voice_forge.py)

Every speaker has their own voice — 17 of them, designed rather than picked off a shelf. The old piper set had 13 characters sharing 6 stock voices; three characters answered in the same voice they were being interrogated by.

The voices come from OmniVoice, a self-hosted open-source speech model (https://omnivoice.madhatter.modlin.cloud). Two calls do the work: design invents a voice from attributes (gender, age, pitch, accent), and clone makes that voice say anything. So each character is designed once into assets/voices/<sid>.wav — their voiceprint, committed to the repo — and every line they ever speak is cloned from it. Same character, same voice, every episode.

python3 voice_forge.py design               # design any voiceprint that doesn't exist yet
python3 voice_forge.py design --sid cynic   # re-roll one character until you like them
python3 voice_forge.py audition             # every voice in one wav — listen, then re-roll
python3 voice_forge.py verify               # pitch/level/pace table + collision check
python3 voice_forge.py speak --sid alpha --text "..." --out line.wav

Design intent lives in characters.json (omnivoice per character, guest_voices for the human co-hosts and the star witness); the wav on disk is the source of truth once you've approved how it sounds. Re-rolling costs about two seconds.

verify is the useful one. The design model is stochastic — asking for "very low pitch" steers it, it doesn't pin it down — so verify measures what actually came out and fails loudly if two characters who share a format land within 12 Hz of each other. design uses the same measurement to pick the best of N takes instead of the first one:

  sid        pitch    level     len     pace   design
  nihil        83Hz   -22.6dB    5.0s   16.8c/s   M·Young Adult·Low·Russian
  cynic        88Hz   -23.7dB    5.0s   17.3c/s   M·Elderly·Low·Australian
  alpha        94Hz   -25.1dB    5.0s   17.2c/s   M·Middle-aged·Low·British
  …
  ✓ no two voices sharing a format are within 12 Hz of each other

Note on two characters: SYSTEM_KERNEL and ADVOCATE-0 were designed female. Nothing in their souls specified a gender, the old cast ran 11:2 male, and a low female adjudicator is the most recognisable voice on the show across all three formats it appears in. Re-roll either with a changed gender in characters.json if you disagree — that's a two-second call.

Export → video (04 // EXPORT)

• transcript.json / transcript.txt — JSON export includes each line's sid, archetype, humor style, model, and voice for full reproducibility.
• ▶ RENDER EPISODE.MP4 (in the UI) — server-side TTS + ffmpeg render. Pick the engine in 03 // ENGINE: OMNIVOICE (default when voiceprints exist) clones each character's designed voice; PIPER reads local .onnx models offline. Either way the picture is identical. Each cast member with a portrait in assets/portraits/<sid>.png (transparent cutout, background-removed with rembg) gets a Terry Gilliam-style floating cutout — Ken Burns zoom, wobbling pan, and a rotation jitter — composited directly onto the format's backdrop; lines without a portrait (human co-hosts) fall back to the original full-width text layout. If assets/backdrops/<format>.jpg exists (one per show format — socratic/roundtable/tribunal/triad), it fills the whole frame; falls back to the flat color background otherwise.
• compile_show.sh — manual/offline export with full parity to the server-side render (same portraits, backdrops, Ken Burns + rotation animation). Useful for rendering on another machine.

One-time setup on Arch:

sudo pacman -S ffmpeg        # the only hard requirement for an OmniVoice render
bash compile_show.sh         # -> episode.mp4

Piper is only needed if you want the offline engine — it stays as the fallback for when the OmniVoice host is unreachable:

yay -S piper-tts-bin        # or: pipx install piper-tts
mkdir voices                 # download the .onnx + .onnx.json voice files
# fallback voices: en_US-lessac-medium, en_US-ryan-high, en_US-amy-medium,
# en_US-joe-medium, en_GB-alan-medium, en_GB-northern_english_male-medium
# from https://huggingface.co/rhasspy/piper-voices
TTS=piper bash compile_show.sh

VOICEDIR=/path/to/voices if your piper voices live elsewhere; GP_OMNIVOICE=http://host:port to point at a different OmniVoice.

Workflow for one episode

1. Pick format + topic, set rounds (4 is a good 10-min episode).
2. TEST CONNECTION (or tick Simulation Mode).
3. ▶ INITIALIZE — watch the feed; type when the yellow human box appears.
4. Edit/re-roll any weak lines after the run.
5. Click ▶ RENDER EPISODE.MP4 — server-side OmniVoice+ffmpeg render, playable in the panel, ready to upload. (compile_show.sh download still available for rendering elsewhere.)