Public, but Unread — Tokyo Insight: Making the Tokyo Metropolitan Assembly's Live Deliberation Record Askable (an On-Demand, Citation-Grounded, Abstention-First Civic-RAG Engine)

The most revealing record of how Tokyo is actually governed has been public for a quarter century, and almost nobody reads it. The stenographic records (速記録) of the Tokyo Metropolitan Assembly — 7,855 records since 1999 at the shipped snapshot — capture government as a live, adversarial, attributed exchange: assembly members pressing, the executive answering, commitments made and qualified on the record. This is a different genre of civic information from the polished official reference (the press release, the plan summary, the FAQ), which states conclusions but not the contest that produced them. The record is published; what was never built is the path (動線): the official site is organized for archival deposit — committee, year, sequence number — so using it requires already knowing where to look, and the primary record goes unread in favor of secondary summaries. Tokyo Insight closes that gap: ask a question in natural language, and a local-first engine routes over a shipped pack of per-record facts and derived vectors (15,683 agenda-section embeddings; no minutes text), fetches only the few relevant records from the official site on demand (robots-respecting, politely delayed, hard-capped), and answers with chunk-level hybrid retrieval (multilingual-e5 + BM25) under a governance prompt that preserves exactly what makes deliberation valuable — who asked, who answered, in which meeting, on which agenda item — with every claim cited back to the official record, and a two-gate abstention rule instead of guessing. The engine ships; the content does not: no corpus is built, bundled, or hosted, and the official record remains the only archive. The paper documents the civic gap, the architecture, the distribution-and-law posture that makes the design shippable (Articles 40, 30, 30-4, 32 mapped to concrete mechanisms), and the deliberate limits (no answer-quality benchmark has been run; the civic-impact claim is the motivating hypothesis, not yet a measured result). Completes the Möbius program's per-product systems-paper series.

AI co-observer: Claude Fable 5 (Anthropic), working method only; the registered author is the human author alone.