diff --git a/.gitignore b/.gitignore
index 30c58f7..8cc6fa5 100644
--- a/.gitignore
+++ b/.gitignore
@@ -22,3 +22,7 @@ docs/PACKAGES.md
 .vscode/
 .idea/
 Thumbs.db
+
+# AI build specs and session reports: kept locally, never committed
+MISSION*.md
+REPORT.md
diff --git a/MISSION.md b/MISSION.md
deleted file mode 100644
index c13755e..0000000
--- a/MISSION.md
+++ /dev/null
@@ -1,36 +0,0 @@
-# Mission: scribe
-
-Self-hosted, GPU-accelerated transcription service. An open-source alternative to Otter.ai that runs entirely on local hardware.
-
-## Goal
-
-A single service that takes audio or video files and produces speaker-labeled, timestamped transcripts, usable three ways: a drag-and-drop web UI, a JSON API, and a CLI. Everything runs locally on the RTX 3060 Ti (CUDA), no cloud calls, no accounts, no per-minute pricing.
-
-## Core decisions
-
-- **ASR**: faster-whisper (CTranslate2 backend) with `large-v3`, `int8_float16` compute on CUDA. Model size is configurable via env, default large-v3.
-- **Diarization**: pyannote.audio speaker-diarization pipeline on CUDA. Optional at runtime: if the model is unavailable (no HF token or terms not accepted) the service degrades gracefully to plain transcription and says so.
-- **Merge**: word-level speaker assignment by timestamp overlap, re-chunked into speaker turns.
-- **Server**: FastAPI + uvicorn, single background worker thread so GPU jobs serialize. SQLite job store so jobs survive restarts.
-- **UI**: single static page (vanilla JS), upload, job list, transcript viewer with per-speaker colors, export TXT / SRT / VTT / JSON.
-- **Python**: 3.12 in a uv-managed venv inside the repo. All caches (uv, Hugging Face models) live on D:, never C:.
-
-## Milestones
-
-1. **Scaffold**: repo, license, config, docs, environment bootstrap script.
-2. **Engine**: audio decode, transcriber, diarizer, merge, export formats. Unit tests for the pure-Python parts.
-3. **Service**: job queue, FastAPI API, web UI, CLI.
-4. **Verified end-to-end**: synthetic two-speaker test clip transcribed and diarized on CUDA, outputs checked.
-5. **Report**: REPORT.md with what works, what is deferred, and how to run it.
-
-## Constraints
-
-- 8 GB VRAM budget shared with other workloads: models load lazily and can be released.
-- Disk on C: is tight, so nothing heavy is written there.
-- No third-party runtime services. Hugging Face is contacted once per model download, then cached.
-
-## Out of scope (for now)
-
-- Live microphone / streaming transcription.
-- Multi-user auth. This is a LAN/Tailscale service behind nginx like the rest of the homelab.
-- Docker image (planned later for batcave deployment with CPU fallback).
diff --git a/REPORT.md b/REPORT.md
deleted file mode 100644
index cad0f91..0000000
--- a/REPORT.md
+++ /dev/null
@@ -1,41 +0,0 @@
-# Build report: scribe
-
-Autonomous build completed 2026-07-13 on the RTX 3060 Ti box. Working tree: `D:\projects\scribe`.
-
-## What was built
-
-A complete self-hosted transcription service, per MISSION.md:
-
-- **Engine**: PyAV decode to 16 kHz mono, faster-whisper 1.2.1 (`large-v3`, `int8_float16`) on CUDA, pyannote.audio 4.0.7 diarization, word-level speaker merge, five export formats (txt, srt, vtt, md, json).
-- **Service**: FastAPI server on port 8323 with upload, job list, status, result, export and delete endpoints. SQLite-backed queue with a single GPU worker thread, jobs survive restarts.
-- **Web UI**: single static page, drag and drop, live progress polling, color-coded speaker transcript, export buttons, dark and light theme.
-- **CLI**: `python -m scribe.cli transcribe file --diarize -f srt -o out/`.
-- **Environment**: Python 3.12 venv via uv, torch 2.11.0+cu128. Everything (venv, uv cache, model cache, runtime data) lives inside the repo on D:, nothing on C:.
-
-## Verified end to end
-
-- 13 unit tests pass (merge logic, export formats).
-- CUDA sanity: torch sees the 3060 Ti, CTranslate2 reports 1 CUDA device through the `cuda_dlls.py` shim, so no system CUDA Toolkit install was needed.
-- Synthetic two-speaker dialogue (46.9 s, generated with Windows SAPI voices by `scripts/make_test_audio.ps1`) transcribed **word-perfect** by the CLI in 62.8 s including first model load. Later runs with the model warm are much faster.
-- Same file through the HTTP API: upload accepted, job queued, progressed through stages, result JSON correct (8 utterances, language en), SRT and VTT exports well-formed, UI and static assets serve, delete removes the job plus its files on disk.
-- Graceful degradation confirmed: with no Hugging Face token, both CLI and server complete transcription and report exactly why speaker labels are missing.
-
-## Known limitations
-
-1. **Diarization is not yet live-tested.** The merge logic is unit-tested and the pipeline code follows the pyannote 4.x API, but the gated model needs a one-time manual step: create a token at https://huggingface.co/settings/tokens, accept the conditions at https://huggingface.co/pyannote/speaker-diarization-community-1, put `SCRIBE_HF_TOKEN=hf_...` in `.env`, then re-run the sample and confirm two speakers come back.
-2. Whisper hallucinated nothing on the clean TTS sample, but real-world noisy audio will benefit from tuning (VAD parameters are currently defaults).
-3. Single worker by design. Fine for one user, revisit if it ever serves the whole homelab.
-
-## Deferred (intentionally)
-
-- GitHub push (`jay739/scribe`), Discussions, icons and landing page per the usual project conventions.
-- Docker image with CPU fallback for batcave deployment behind nginx.
-- Live microphone streaming, folder watch, Jellyfin subtitle integration.
-
-## How to run
-
-```powershell
-cd D:\projects\scribe
-.venv\Scripts\python.exe -m scribe.server    # web UI on http://localhost:8323
-.venv\Scripts\python.exe -m scribe.cli transcribe path\to\file.mp3 --diarize -f srt
-```