Deterministic cultural-intelligence orchestration

Status: executable policy and measurement control; agent execution and public

publication remain external, human-gated actions.

Run a candidate-assessment batch with:

pnpm connections:orchestration -- --input=tests/fixtures/cultural-intelligence-orchestration.json

The command supports two explicit modes. `assessments` mode calibrates the pure

policy with supplied evidence counts and ratios. Production-oriented `report`

mode accepts a review-only `CollectionIntelligenceReport`, candidate-specific

`rightsEvidence`, and optional bounded `requests`. In report mode the service

derives evidence counts, source independence, contradiction penalties,

completeness, provenance quality, and consequential/sensitive review flags from

detector output; callers cannot author their own score.

Missing rights evidence becomes `unknown` and blocks reuse. Duplicate or invalid

rights rows fail the entire run. Every detector candidate becomes a traceable

review-queue record containing copied source evidence, claim boundary, pending

approval slots, agent authorization state, and a locked publication state.

The routing contract is fail-closed:

bounded agent pass; and

estimate above a positive cost ceiling can authorize two independent passes;

duplicate executors are refused and both typed results are required; and

subject-matter approvals are recorded where applicable.

• incompatible or unknown rights block reuse;
• weak candidates are archived without an agent call;
• adequately supported factual aggregations enter deterministic human review;
• only strong ambiguous candidates receive a structured packet authorizing one
• only high-scoring consequential conflicts with an explicit expected-value
• every route remains non-publishable until its editorial, rights, novelty, and

An evidence packet contains the assessment, bounded task, constraints, and a

required typed response covering claims, citations, uncertainty, alternatives,

missing evidence, and refusal. It never authorizes publication or a novelty

claim. The pure service performs no network or LLM call; an external orchestrator

must separately execute an allowed packet and preserve the result.

The run artifact measures candidate-level no-LLM coverage. A batch meets the

initial target only at 80% or greater. The retained fixture proves the policy can

route four of five representative candidates without an LLM while allowing one

bounded literature/interpretation pass and authorizing zero publication actions.

This is production-like control evidence, not evidence of live traffic volume or

historical truth.

The artifact also exposes a stable descending score ranking with candidate-ID

tie-breaking, preserving each route beside its score for reproducible triage.

Non-LLM similarity and graph results enter this pipeline only through the

source-backed candidate adapter documented in

`cultural-intelligence-probabilistic.md`(cultural-intelligence-probabilistic.md).