Second-Order Causal Decision Framework

NOMOS Decision Audit Hub · Second-Order Causal Decision Framework

NOMOS is a deterministic, auditable decision kernel with built-in LLM guardrails. It does not replace your ERP / CRM / risk / compliance systems — it sits behind them: upstream systems submit structured decision requests, and NOMOS returns candidate sets, counterfactuals, a hash-chained audit trail, and a formal convergence state. The final call always rests with a human signatory.

deterministic hash-chained audit IMDA AI Verify 95/100 human final say nomos v0.4.0
GitHub → Gitee mirror →
✦ About

What is NOMOS

A second-order causal decision framework. Every round runs three causal operators over a bounded, hash-linked, human-gated reconstruction — never fabricating missing facts, weights, thresholds, evidence, or responsibility.

NOMOS upgrades one-shot causal reconstruction into an iterative, hash-chained, human-gated process. Second-order effects (synergy, redundancy, amplification) must be explicitly declared by the responsible party — the engine never estimates or learns them on its own. LLMs are confined to a three-tier permission zone and cannot flip states, adjudicate, or rewrite weights.

✦ Positioning

What it is — and what it is not

✅ It is❌ It is not
A deterministic decision kernel behind ERP / CRM / risk / compliance systemsA replacement for those business systems
A framework that never invents missing facts, weights, thresholds, or responsibilityA probabilistic model that estimates interaction strengths on its own
LLM-guarded: models confined to a 3-tier permission zoneAn auto-decider — the final call always rests with a human
IMDA AI Verify 95/100 · 402 tests, 0 failuresA black-box recommender with no audit trail
✦ Core Capabilities

Four pillars

1 · RECONSTRUCTION

Three-Layer Causal ReconstructionForward / Backward / Delta

Forward invalidation propagation, backward root-cause tracing (reverse BFS), and delta reconstruction on a copied request — re-running deterministic evaluation to check convergence.

forward propagationroot-cause BFSdelta rebuild
2 · INTERACTION

Assumption Interaction Layer4 Invariants

Handles second-order interactions (synergy / redundancy / amplification) under four invariants: non-fabrication, order conservation, effect boundedness, and monotonicity.

I-1 non-fabricationI-2 order conservationI-3 boundednessI-4 monotonicity
3 · CONVERGENCE

Formal Convergence Engine5 Explicit States

Replaces heuristic stop checks with five states — FIXED_POINT, NO_GAIN, BUDGET_EXHAUSTED, DIVERGED, BLOCKED — backed by termination / fixed-point / boundedness propositions.

FIXED_POINTBUDGET_EXHAUSTEDBLOCKED
4 · GUARDRAILS

LLM Compliance LayerT1 / T2 / T3

Three permission tiers — T1 Annotation, T2 Proposal, T3 Narrative — with zone isolation, no adjudication, and strength stripping. LLM output never touches the decision graph directly.

T1 annotationT2 proposalT3 narrative
✦ Quick Start

Run it in one command

# Install (Python 3.8+), then run the full test suite
pip install -e ".[dev]"
pytest                    # 402 tests, 0 failures

# Core decision-engine demo
nomos-demo
nomos-demo --decision examples/baidu_org.json --json

# Hub demo with two stress scenarios
nomos-hub-demo

# Start the API service (dev)
export SP_ENV=development
uvicorn second_perspective.api.main:app --reload

Key endpoints: POST /v1/hub/analyze, POST /v1/decisions/evaluate, GET /v1/decisions/{id}/history, POST /v1/decisions/{id}/approval. Optional PostgreSQL persistence via SP_DATABASE_DSN and OIDC via SP_OIDC_ISSUER.

✦ License & Authorization

Dual-track licensing · not open source

Free for personal, non-commercial research; government / enterprise commercial use requires a paid commercial license.

AudienceTerms
Personal researchersFree for non-commercial research, not for any commercial use
Government / enterprisePrior written authorization required
How to applyGlobal: ai@nohnlins.com · China: lin@as.nohnlins.com
Apply (Global) Apply (China)