Subagent Exploration Report: Causal + Signal Intelligence + Knowledge Graphs
Date: 2026-03-17
Scope: Rapid repository reconnaissance with implementation mapping for causal KG decision intelligence.
Mission
A lightweight “subagent-style” pass was executed to:
- map major platform surfaces in this monorepo,
- identify where causal + signal intelligence capabilities already exist,
- align current implementation opportunities with the latest causal KG research directions.
Repo Recon Summary
Monorepo characteristics
- Multi-language monorepo with Python, TypeScript/Node, infra/deployment artifacts, and service-level Cloud Run manifests.
- Root workspace includes services, cells, SDK packages, deployment automation, and large extension footprints.
- Strong emphasis on decision intelligence, orchestration, and KG-linked pipelines.
High-signal top-level areas
services/— Cloud Run microservices (KG, ingestion, instrumentation, replay/simulation, journey systems).miz-oki-adk-agents/andsrc/— agent runtime, orchestration logic, and cell-level implementations.plugins/— pluginized decision components (including ads/marketing decision workflows).architecture/anddocs/— architecture references and historical implementation reports.scripts/anddeployment/— deployment + verification automation for production operations.
Existing Causal/KG Alignment Already Present
The codebase documentation indicates this platform already includes meaningful foundations aligned with causal intelligence patterns:
- closed-loop decision systems (SRPVDAL/SRPVG style loops),
- causal DAG-oriented reasoning modules,
- KG and event-sourced integration patterns,
- marketing journey KG architecture with ingestion → normalization → semantic mapping → graph operations.
This means the repo appears structurally ready for deeper Causal Knowledge Graph (CKG) integration rather than requiring a greenfield build.
Mapping to 2026 Research Directions
Below is a practical mapping between research trends and probable implementation loci in this repo.
| Research trend | Repo-aligned implementation direction | Candidate areas |
|---|---|---|
| Causal Knowledge Graph architecture | Introduce explicit causal edge types + SCM metadata in KG schemas and ingestion normalization layers | services/kg-*, data-pipelines/, config/ |
| Causal discovery-driven KG fusion | Add causal-consistency checks to graph merge/fusion jobs and schema conflict resolution scripts | scripts/graphrag/, services/graph_writer/, mcp/ |
| Graph-based causal uplift for marketing | Add CATE/uplift outputs into campaign optimization and budget reallocation services | services/budget-reallocation-service/, plugins/ads-decision/ |
| GraphRAG + causal reasoning | Extend retrieval contracts with causal path retrieval and counterfactual query operators | services/customer-journey-system/, src/, miz-oki-command-center-ui/ |
| Counterfactual simulation agents | Integrate intervention simulation APIs into decision loops and post-decision evaluation | services/replay-sim-ope/, miz-oki-adk-agents/ |
| Neuro-symbolic KG reasoning | Add symbolic policy constraints and explainability traces on top of neural scorers | policies/, contracts/, mcp/ |
| Causal benchmarks | Add benchmark harness wrappers for CausalBench/CARL-GT style evals in consolidated tests | consolidated_tests/, tests/, scripts/tests/ |
Recommended Execution Plan (Incremental)
Phase 1: Causal Schema Hardening
- Extend KG node/edge schema to include:
causal_relation_type,effect_sign,confidence,intervention_target,counterfactual_support.- Add schema validation gates during ingestion.
Phase 2: Causal Retrieval + Decision API
- Add retrieval primitives for:
- confounder neighborhood lookup,
- minimal adjustment set hints,
- intervention target path scoring.
- Expose as MCP/HTTP tools for agent consumption.
Phase 3: Uplift + OPE Integration
- Couple campaign uplift scoring (CATE-like outputs) with existing replay/OPE service to validate policy changes before rollout.
Phase 4: Benchmark Harness
- Stand up repeatable benchmark runner in
consolidated_tests/for causal reasoning tasks and intervention quality metrics.
Risks and Controls
-
Risk: correlation leakage masquerading as causal lift.
Control: mandatory intervention simulation checks + refutation tests before promotion. -
Risk: heterogeneous graph fusion introducing inconsistent edges.
Control: causal-consistency-first merge policy. -
Risk: opaque decisioning in production.
Control: attach structured causal explanation traces to every action recommendation.
Bottom Line
This repository is mature enough to adopt a first-class Causal Knowledge Graph operating model quickly. The highest leverage path is to harden causal schema semantics in existing KG/decision flows, then layer retrieval-augmented counterfactual tooling and benchmark-driven validation on top.
Implementation Follow-up (Code Added)
To ensure the methodology is not only documented, the replay/OPE service now includes an executable causal gating contract:
POST /gatenow acceptscausal_evidencein the request body.- Gate decisions now run causal evidence validation and hard-block promotions when evidence is missing or below threshold.
GET /methodologyexposes the active causal methodology contract (required fields, confidence threshold, allowed relation/effect values).GET /statusnow includescausal_min_confidencein runtime config.
This moves the work from research-report-only into live service logic.