Kaisetsu Studio Technical Documentation

Core Principle: An evidence-first studio that pairs AI production speed with newsroom-grade fact-checking and multi-role human sign-offs.

1. Architecture Overview

Client (Web SPA / Firebase Google Auth)
   │ Session Cookie + CSRF Header
   ▼
Cloudflare Workers (Hono API + Static Assets)
   ├─ D1: Topics, Sources, Claims, Scripts, Approvals, Hash-chained Audit Log
   ├─ R2: Source raw snapshots, Audio WAV, Render MP4s, ZIP packages
   ├─ KV: Policy overrides, RSS feeds cache, Rate-limit counters
   ├─ Queues: Asynchronous claim extraction, script drafting, TTS, renders
   └─ Containers: Hypit Render Engine (Node 24 + Chromium Headless + FFmpeg)

2. Primary Sources & Verifiable Claims

Sources undergo SSRF checks, robots.txt verification, and 5MB caps before freezing in R2. Gemini can only extract factual candidates backed by exact verbatim substring quotes from the snapshot. Unmatched candidates are automatically rejected.

3. Ikegami Philosophy & Static Linting

L001: Fact segments must cite verified claims.
L004: Uncited numbers in script trigger warnings or errors.
L005: Absolute assertions (100%, absolutely) flagged.
L006: Zero impersonation of real public figures (hard error).
L007: Sensitive geopolitical/religious terms routed to risk reviewer.
L008: Superiority framing prohibited.
L011: Fabricated first-person reporting strictly blocked.

4. Hypit Container Rendering

Based on SVML markup, headline cards and captions are timed to audio waveforms, rasterized via Chrome Headless, and encoded to MP4 with FFmpeg inside Cloudflare on-demand containers.