MoreSalamander / projects / Genesis OS / 01

01 · Signal Intelligence

External intelligence for an AI-native film studio. Watch its intelligence bureau research the outside world: agents plan a mission, retrieve real evidence through Parallel's Search API, verify claims while preserving disagreement, and answer the one question the Studio Head asked.

Source on GitHub ↗ Hosted console ↗ The federation Next · 02 Operational

The question it answers

"What is happening outside the studio that should change what we do?"

The loop

STUDIO HEAD objective
  → SIGNAL INTELLIGENCE EXECUTIVE     owns the objective, not the searches
  → RESEARCH PLANNER                  Gemini decomposes it dynamically
  → MARKET ∥ TALENT ∥ INDUSTRY ∥ STRATEGIC     four locked cognitive domains
  → PARALLEL SEARCH API               external observation, at runtime
  → EVIDENCE → VERIFICATION           VERIFIED / UNVERIFIED / CONFLICTED, conflicts kept
       ↑ PROVENANCE AUDIT             re-opens a claim's pages and quotes what they say NOW
  → KNOWLEDGE GRAPH                   Source → Evidence → Claim → Entity
  → STRATEGIC COGNITION → RECOMMENDATION     confidence + evidence trail
  → COVERAGE AUDIT                    which findings rest on nothing verified?
  → NESTED QUESTIONS                  each shortfall becomes a question of its own (≤4, never recursive)
  → STUDIO HEAD                       approve / reject / request more research

What makes it not a search chatbot

Verification states are computed, not asserted. A claim with no corroboration is UNVERIFIED on the record; a contradiction is CONFLICTED and shown, never resolved by the model.
The provenance audit puts no model between page and panel. A claim's recorded pages are re-opened on demand and quoted verbatim, or honestly reported UNREACHABLE.
The answer is read back against its own graph. The coverage audit finds which findings rest on nothing verified, and turns each shortfall into a nested question that is researched like one you typed — bounded to four, and a raised question never raises more.
The human authority boundary is a workflow signal. Missions run in a Temporal worker; approval is a durable signal the Studio Head sends.

Runtime proof

Requirement	Where
Parallel Search API at runtime	verified live with 25 real sources retrieved in the verification mission
Google Cloud AI at runtime	Gemini via `google-genai`
Honest degradation	with no keys it boots in MOCK mode; every substrate chip reads MOCK rather than pretending

Run it

uv venv .venv --python 3.12 && uv pip install -p .venv -e ".[dev]"
./ops/dev.sh          # API :8000 + Temporal worker + console :3000, one process group