UI: clean EHR-style clinical theme; remove Peitho branding.

ZeroGPU: disable moshi torch.compile + CUDA graphs.

ZeroGPU: load on CPU, place on GPU inside @spaces.GPU, fresh ChatState per turn.

ZeroGPU attempt: boot-safe loader + @spaces.GPU realtime handler.

Hackathon: full tags, Field Notes report, custom UI for Off-Brand.

Redesign UI: dark teal theme, hero, schedule cards, question chips.

Add few-shot example Q&A turns to coax v6 into plan-grounded answers.

Ground each question with the plan in the user turn; assertive prompt.

Make it the Praxis-Briefing medical demo on the working v6 realtime stack.

v6 TURN: use own Cloudflare Calls keys (free fastrtc/HF TURN is down).

Fix v6 demo TURN: use live HF Hub endpoint, lazy fetch, STUN fallback.

Roll back Space to Peitho v6 demo (state of 8360af0).

Route TURN through HF Hub API when turn.fastrtc.org DNS is down.

Align Space S2S with official liquid-audio-demo WebRTC stack.

Fix mic audio shape and add realtime WebRTC S2S tab.

Pivot to Praxis-Briefing: GP voice schedule lookup, no WebRTC.

Use HF API for TURN credentials instead of dead fastrtc.org DNS.

Fetch TURN lazily when mic connects instead of at boot.

Fix Space startup crash when TURN DNS is not ready.

Pivot to Nachsorge-Stimme medical discharge voice check-in.

Add Morgenruf morning check-in for Oma.

chat: restore v6 routing, document why base-only failed

chat: route through base_model, drop transcribe pre-pass, new persona prompt

chat: route transcribe pass to base_model, chat to v6_model

Upload chat.py with huggingface_hub

Try short persona prompt + temp=0.3 topk=10 to combat verbosity

Upload folder using huggingface_hub