Canon-first planning
Writing the synopsis (including the ending) before parallel drafting bounds plot drift across hundreds of pages.
Benchmarks
Measured results
Wall-clock timings and stability findings from local runs on Apple Silicon with llama-3.1:8b, plus sample generated book pages (redacted previews only).
Fiction pipeline timings
M2 Ultra · four Ollama servers · sixteen concurrent writers · Q6_K quantization
| Configuration | Director | Planner | Writer | Reviewer | Total |
|---|---|---|---|---|---|
| Fiction 90k | 25–30s | ~1 min | 45–90 min | 10–20 min | 1–2 h |
| Fiction 150k | ~30s | 85s | 95–130 min | 14–22 min | 110–155 min |
| Fiction 300k | ~45s | 130s | 190–245 min | 28–45 min | 220–295 min |
Key findings
KV-cache-aware concurrency
Reviewer concurrency stays at one quarter of writer slots, eliminating transient llama.cpp failures on long prompts.
Scene-granular resume
Atomic checkpoints after every scene mean a multi-hour run survives interruption with at most one scene of rework.
Automated media pipeline
Pictures mode calls Ollama image APIs; animated mode renders ManimRedux AVI→MP4; video mode reserves a text-to-video hook with ffmpeg stubs.
Native LaTeX publishing
All modes emit fragment LaTeX—graphicx figures, media9 clips, hardback 7×10″ layout, XeLaTeX for Devanagari language courses.
Structured language drills
Fill-in-the-blank, multiple choice, true/false, and mix-and-match exercises rotate per chapter with a separate answer key.
Sample generated pages
Preview pages from a Hindi language textbook run—vocabulary tables, lesson prose, and structured exercises. Full manuscripts and checkpoints are not published here.
Stability
Before retry-and-fallback logic, reviewer-stage KV-cache failures aborted roughly 24% of 90k-word runs. With tier-specific concurrency caps and transient-error retries, completion reached 100% across fifty trials, with graceful stitch fallback on at most one chapter per run in 18% of cases.