1.6.1
Local Assembly Civic Data Lab — Hypotheses and Overview: Reading Assembly Speeches as Observations
A research lab built on a cross-searchable record of speeches from all 1,788 Japanese local assemblies, designed to read the data as observations rather than verdicts. The subject is structural — distribution, propagation, and silence across municipalities — not individual condemnation. Differentiation rests on data quality assurance, verification scripts, and transparent editorial judgment.
This note sets out the working hypotheses, observation policy, and editorial premises that all articles in the Local Assembly Civic Data Lab build upon. It serves as a map before any individual piece.
What problem we are working on
Japanese local assembly minutes have long been published — but in fragmented systems, one municipality at a time, with little capacity for cross-cutting search. Questions raised by councillors and responses from mayors and officials accumulate daily as the language that precedes policy, before it is fixed in ordinances or budget documents.
KOUSOU 0 (kousou0.isvd.or.jp) is the infrastructure that arranges the minutes of all 1,788 local assemblies in a form that allows cross-cutting citation. On top of that base, this lab takes one stance: read the data as observations. Observations here are inputs for structural reading, not grounds for ranking or judgment.
Rather than "Municipality X is the worst at deferring decisions", the subject of analysis is "How is the use of deferral phrasing distributed across the country", "Since when has the same wording appeared in different assemblies", or "On which topics do how many assemblies remain silent". This distinction is operationalised as a three-tier rule (see below).
Outline of the lab
| Aspect | Content |
|---|---|
| Data base | KOUSOU 0 (241 million speech records, aggregated 2026-08-04) |
| Subject of publication | Structural analysis, national distribution, time-series patterns, topic-level silence counts |
| Not published | Evaluative descriptions of individual councillors, worst-of-the-worst naming of individual assemblies |
| Operator | Institute for Social Vision Design (ISVD) |
| Review | Editorial committee (Phase 1 in formation) + DA review |
Editorial judgment — three tiers for councillor-level data
To balance research utility and neutrality, councillor-level data is published at three distinct granularities.
| Granularity | How it is presented |
|---|---|
| Municipality-level aggregates | "Across the country, the share of deferral phrasing ranges from 0% at the minimum to 21% at the maximum" |
| Party-group-level aggregates | "On a given topic, group A speaks roughly twice as often as group B" |
| Individual verbatim quotation | Always with a direct URL to the original assembly record, and without interpretive or evaluative framing |
Individual evaluative descriptions ("councillor A is unmotivated", and similar) are not produced in this lab. Individual scrutiny belongs to journalism as a separate layer; the lab's contribution stops at structural analysis.
Judgment criteria, the complaints-handling protocol, and the procedure for retractions are documented and published. Summaries of editorial committee discussions will also be made public.
Distinguishing what we push and what is pulled
The lab distinguishes between push and pull in handling findings.
| Mode | Form | Expected flow |
|---|---|---|
| pull | A structural analysis published as a Local Assembly Civic Data Lab article | Journalists, researchers, and citizens cite it voluntarily |
| push | Handing findings directly to a chosen newsroom ("This municipality is the problem") | Not done |
Pushing findings to specific media would mean ISVD setting the agenda of journalism from outside, which runs into journalistic-ethics tension. The lab's role stops at placing verifiable structures in the public layer. Where a joint project shares methodology, primary data, and verification scripts with a partner, the verbal articulation of findings is not part of the deliverable.
The core of differentiation — data quality, verification scripts, transparent editorial judgment
If the National Diet Library, the Ministry of Internal Affairs, the Digital Agency, and others advance official open-data publication of assembly proceedings, the differentiator of "owning the data" may structurally erode. What the lab can keep as a long-term distinction lies in the process of preparing the data into a trustworthy form and in making that judgment externally verifiable.
Three investments anchor this position:
| Area of investment | Content |
|---|---|
| Data quality assurance | Public quality gates for deduplication, entity resolution, and municipality-name verification |
| Verification scripts | Aggregation queries and reproducible code published on GitHub so that third parties can reach the same conclusions |
| Transparent editorial judgment | Public documentation of the criteria for avoiding individual condemnation, the retraction protocol, and the editorial committee's operations |
This direction connects with an earlier observation that "the real substance is in proof systems, quality, and grasp of reality".
Scope of themes
Published and forthcoming articles point in the following directions:
| Theme | Content |
|---|---|
| Time-series analysis | When a given word first appears in assemblies and how it spreads (first article: "When did 'AI' and 'Generative AI' first appear in Japanese local assembly speeches") |
| Structural analysis | Reading national distributions of response phrasing and the allocation of debate (flagship: distribution of response phrasing, gaps in childcare debate) |
| Structuring silence | Making visible the topics on which debate fails to take place |
| Policy propagation | How the same policy framing travels between municipalities |
| Cross-lab citation | Connection with other labs (agnotology, public-asset-ppp, and others) |
Each article follows the MUST principle — "the subject is structural; do not condemn the individual" — and passes editorial committee review (Phase 2 onwards) and DA review before publication.
Limits — what is not yet covered
- Coverage is read as breadth and depth, which are separate measures. Breadth (assembly coverage) is 1,742 of 1,788 assemblies (97% covered) with at least one speech row. The remaining 46 assemblies hold no rows at all, and we have not tallied why each one is absent. Depth (period, session and speech layers) is still being measured, and we publish no single ratio for it
- The old depth metric was withdrawn on 2026-08-11. It counted assemblies for which full retrieval had been verified, but those ratios ranged from 113% to 2,822%. Numerator and denominator counted different units, so "100%" only meant holding more items than the denominator. We no longer force a common unit; items are matched by a shared identifier per source system. Assemblies for which no ratio can be produced are reported by reason. The new definition will be published once the measurement settles
- Ingestion progress is mixed into the time series. Older years are less covered, and we have not yet separated that effect from year-on-year trends
- Semantic analysis has not been applied. Many analyses sit at the surface aggregation stage; polarity, support/opposition, and topic clustering are being extended in stages
- Individual attribution is out of scope. Evaluation of individual councillors is not undertaken here, and individual scrutiny by journalism is respected as a separate layer
These limits are stated in the methodology section of each article. The lab's operating policy, the editorial committee's bylaws, and the corrections contact will be presented on a separate operations page.
Related Research Notes
This section organizes 10 research notes by type.
Time Series Trend Analysis (trends-)
- Seven Years of Assembly Seat Reduction Debates — The Time Structure of 'Candidate Shortage' and Seat Apportionment Ordinance Revisions in Japanese Local Assemblies
- Seven Years of 'Decarbonization' Discourse — Discontinuous Diffusion of Climate Language in Japanese Local Assemblies (2018–2024)
- Seven Years of 'Fiscal Consolidation' Discourse — The Temporal Structure of Fiscal-Crisis Language in Japanese Local Assemblies (2018–2024)
- Seven Years of LGBTQ and Partnership Ordinance Debate — Regional Diffusion of Sexual Minority Discourse in Japanese Local Assemblies
- Seven Years of Young-Carer Mentions in Japanese Local Assemblies — The Time Structure of 'Children Who Care for Family Members' in Assembly Discourse
Case Studies (case-)
Literature Maps (literature-map-)
Regional Distribution Analysis
- Regional Distribution of 'Vacant Houses' and 'Public Facilities' Debate in Japanese Local Assemblies — A Seven-Year Structural Analysis Across 1,320 Municipalities
- Seven Years of 'Care' Mentions in Japanese Local Assemblies — The Time Structure of Long-Term Care Insurance Act Revisions and Policy Lag
- Regional Distribution of 'Foreign Workforce' Debate in Japanese Local Assemblies — A Seven-Year Structural Analysis Across 870 Municipalities
Statistics cited in this article
- 1KOUSOU 0 collection status（Aggregated 2026-08-04） Open source
References
- 1.machikarte — Nationwide Local Assembly Speech Search Platform (Beta)(external site, opens in a new tab)
Institute for Social Vision Design (ISVD) (2026)
ISVD
- 2.machikarte (GitHub) — schema, aggregation queries, licenses (MIT + CC BY 4.0)(external site, opens in a new tab)
Institute for Social Vision Design (ISVD) (2026)
GitHub
- 3.When Politicians Talk AI: Issue-Frames in Parliamentary Debates Before and After ChatGPT(external site, opens in a new tab)
Suter, V. et al. (2025)
Policy & Internet
Behind a subscription paywall.
Previous
National Distribution of Deferral Phrasing in Assembly Responses — A Structural Analysis of 18.97 Million Records from 870 Municipalities
Next
