03 · Institutional Intelligence

A century of institutional history — Convergence Studios, founded 1912, about 4,600 productions across ten industry eras, roughly 104 million fact rows seeded deterministically in two and a half minutes — and a multi-agent system that answers: what does our history tell us, and what happens if we change this assumption?

Not "chat with your database"

Investigations are multi-step analytical campaigns — baseline → cohort → comparison → trend → simulation — where every number comes from ClickHouse through the official mcp-clickhouse server, every statistic is computed in code, and every finding carries a verification state derived by rule, never asserted by the model.

Studio Head question
  → Analytical Planner          hypotheses per cognitive domain
  → Query Engineer              Gemini writes SQL, schema-grounded, SELECT-only enforced in code
  → official mcp-clickhouse     → ClickHouse (1912–2026 corpus, ~104M rows)
  → Statistical Verification    states computed in Python
  → Competing interpretations → Findings → Scenario simulation
  → Recommendation              confidence = verification coverage
  → Studio Head                 approve / reject / request deeper analysis   (Temporal signal)
  → VERIFIED knowledge promoted to DataHub with full query lineage

The verification states

StateMeaning
VERIFIEDeffect size holds against dispersion (scale-free, so 104M rows cannot buy significance) and holds across at least three eras
REGIMEa truth with an era boundary — real, bounded, promoted with its range
WEAKpresent but under the bar
CONTESTEDconflicting analyses, preserved side by side, never silently resolved
INSUFFICIENTnot enough to say

The corpus was engineered with known truths — overrun ordering indie < mid < tentpole, a sequel opening premium, seasonality, and one deliberately contested sequel pair where the raw cohort says premium and seven of nine within-franchise comparisons say fatigue — so the states can be checked against what was seeded.

Runtime proof

RequirementWhere
Official mcp-clickhouse MCP server at runtimeapp/tools/clickhouse_mcp/client.py; agents reach ClickHouse only through it, as a read-only user
Google Cloud AI at runtimeGemini on Vertex via google-genai
Durable human authoritythe Studio Head's approve / reject / request-deeper is a Temporal signal; a worker killed mid-pause, an approval landing with no worker alive, and a resumed, promoted decision were all demonstrated
Hosteda three-container Cloud Run pod: API, the official MCP server, and self-hosted ClickHouse with the corpus backup baked in and restored per instance behind a startup probe

Run it

cd ops && docker compose up -d --build && cd ..   # ClickHouse + mcp-clickhouse + PostgreSQL + NATS + Temporal + Redis
python3 -m venv .venv && .venv/bin/pip install -e ".[dev]"
.venv/bin/python -m seed.build                     # the deterministic 1912–2026 corpus