# Session Notes — `attestation-fable-5-max-claude-code` Factors that shaped this session's output, for anyone comparing artifacts across models/harnesses/effort levels. ## Model & configuration - **Model:** `claude-fable-5` for the entire session. **No safeguard downgrade occurred** — Fable 5 sessions can fall back to Opus 4.8 when safety classifiers decline a request, but no refusal or fallback event happened here, so no restore step was ever needed. The user ran `/model` once mid-session (before the voice-fix turn) and **kept** Fable 5. - **Effort:** `/effort` was set to **max** ("this session only") as the very first action, before any task prompt, and never changed. Output therefore reflects Fable 5's deepest reasoning setting throughout — long deliberate thinking blocks, extensive self-verification (visual inspection of 20+ rendered frames, numeric audio-level checks), and willingness to discard and re-run full renders over small defects (a stale colophon credit, 0.8 dB of audio headroom). - **Thinking:** always-on (a Fable 5 property); no fast mode. ## Harness - **Claude Code CLI** on macOS 26.5 (arm64), zsh, in an initially empty git repo. No `CLAUDE.md` / project instructions existed at start. - **Single agent, serial.** No subagents were spawned (no Agent-tool / Explore / Plan / general-purpose delegation). Within turns, independent tool calls were batched in parallel (e.g., three recon commands at once; paired image Reads), but there was never more than one model instance. - **Model/compute overlap:** the four full video renders ran as *background Bash tasks*; the model continued working (README edits, code fixes) while ffmpeg/Kokoro ran, and was re-invoked by harness notifications on completion. One background render was deliberately killed (`pkill`, exit 144) to fix the colophon before re-launching. - **Harness injections observed** (system-generated, not user input): periodic task-tool usage reminders; a date-rollover notice (2026-07-29 → 2026-07-30 — the session crossed midnight); background-task completion notifications explicitly marked "NOT USER INPUT"; and periodic automated policy reminders. A foreground `sleep` was blocked by harness policy (progress was read from the task's output file instead). - **Sandboxed Bash** with a session scratchpad; network access worked (mise/uv/PyPI/Hugging Face downloads) without sandbox overrides. The harness's task list (TaskCreate/TaskUpdate, 6 tasks) was used for progress tracking. ## Memory / persistence - A persistent file-based memory directory was available to the model, but **no memories were recalled and none were written** — the session was self-contained, so nothing met the bar for cross-session persistence. - **No context compaction/summarization was observed**; the full conversation stayed in context (Fable 5's 1M window; harness noted a 1-hour prompt-cache TTL). The TTS cache (`dist/build/tts/`, keyed by backend|voice|speed|text) is the only on-disk state that made re-renders cheap across turns. ## External dependencies affecting reproducibility - **Kokoro-82M** neural TTS: ~330 MB model auto-downloaded from Hugging Face on first use, plus spaCy's `en-core-web-sm` fetched by misaki mid-run. Voice output depends on these exact artifacts. - **Apple system fonts** (New York, Baskerville, Avenir Next, Menlo) — the visuals require macOS; Georgia/Helvetica fallbacks exist in code. - **macOS `say`** remains as fallback narrator; final film does not use it. - mise **verified GitHub artifact attestations** for the uv tarball during setup — noted because it is the video's subject occurring in its own toolchain, and because it means the toolchain trusts what this repo's film is about. ## Decisions & defects that materially shaped the output 1. **Font rendering:** Apple's New York variable font silently drops `— – - = +` under FreeType/Pillow; running text moved to Baskerville (New York retained for display headlines only). Menlo lacked `⟷ ‖` (replaced with ASCII). 2. **ffmpeg `-shortest`** trimmed the MP4 to the *subtitle* stream's last cue, cutting the final 4 seconds (the closing seal + colophon). Removed — audio/video are equal-length by construction. 3. **Narrator swap** (turn 2): Samantha (concatenative) → Kokoro `af_heart` at 0.95×; all scene/subtitle/beat timing re-derived automatically because measured narration duration is the pipeline's master clock. 4. **Prosody direction** (turn 3): per-sentence `speed` (0.80–1.0) and inline `` markers synthesized as true silence between separately generated fragments — Kokoro has no SSML, so delivery is directed through pace, punctuation, and silence. 5. **Original score** in D minor generated in NumPy from the narration schedule; "V for Vendetta" was treated strictly as a *mood* reference (dark strings, patient builds, resolute finale) — no melodic or audio material from the film's soundtrack was used or reproduced. 6. **Audio headroom:** Kokoro's higher crest factor pushed the first mix to −0.5 dBFS; final master ceiling is −1.3 dBFS for AAC true-peak safety. 7. Three complete films were rendered and superseded (7:15 Samantha → 7:27 Kokoro → 8:10 directed + scored). Only the last survives in `dist/`. ## What was *not* done - No git commits (user never asked; tree left dirty by design). - No web fetches/searches; everything from local tools, PyPI/HF packages, and built-in knowledge. Pricing in `metadata.toml` came from the harness's `claude-api` skill reference (loaded in the final turn), not memory. - No use of external media assets of any kind — every frame and every audio sample is synthesized in this repo's code. ## Accounting caveats The harness exposes no token/timing counters to the model. All duration, token, and cost figures in `metadata.toml` are **good-faith reconstructions** from the conversation record (file timestamps, render logs, message counts, typical vision-token sizes) and are marked as estimates there. Rates are the published Claude Fable 5 prices current as of 2026-07-30.