Runtime governance environment enforcing the CoSyn Governance Stack (CGS) v15.1.0 on every AI interaction.
The runtime wrapper places AI models inside a deterministic governance pipeline that enforces reasoning discipline before and after model execution.
Unlike prompt guardrails, policy engines, or tool-call security layers, CoSyn governs the reasoning process itself.
The Runtime Wrapper distribution includes the following constitutional artifacts:
| Artifact | Version |
|---|---|
| CoSyn Constitution | v15.1.0 |
| Persona Governor | v3.15.0 |
| Stack Architect | v3.15.0 |
These artifacts define the governance rules enforced by the runtime wrapper.
CoSyn is a runtime governance operating environment for AI systems.
Every interaction passes through a structured governance pipeline:
User ↓ Ingress Governance Gates ↓ LLM Reasoning ↓ Egress Governance Gates ↓ User Response + Telemetry Ledger
The model does not operate freely.
It operates inside the CoSyn governance pipeline.
The wrapper enforces:
• explicit interpretation of user intent • identification of hidden assumptions • bias and optimization frame declaration • reasoning outcome scoring • structured response compliance • full telemetry audit trail
AI safety approaches typically operate at one of three layers:
| Approach | Enforcement Layer |
|---|---|
| Prompt guardrails | Input / output text |
| Policy engines | Model output evaluation |
| Tool contracts | Tool invocation boundary |
These approaches leave the reasoning pipeline itself ungoverned.
CoSyn addresses this gap by enforcing governance before and after reasoning occurs.
| System | Governance Layer |
|---|---|
| SmythOS | Agent workflow orchestration |
| SAFi | Policy enforcement on model outputs |
| CORE | Constitutional governance for AI-assisted development workflows |
| Edictum | Deterministic contracts on tool calls |
| CoSyn | Governance of the AI reasoning pipeline |
CoSyn governs the cognitive structure of AI interactions, not just the actions or outputs.
The runtime wrapper acts as governance middleware between the user and the AI model.
User ↓ FastAPI Runtime Wrapper ↓ Ingress Governance Gates ↓ Model Adapter ↓ Egress Governance Gates ↓ Telemetry Store ↓ Response
Components:
• FastAPI runtime service • Redis session state store • ingress governance gates • egress governance gates • model adapter layer • telemetry audit system
The pipeline contains two governance stages.
These gates execute before the model runs.
| Gate | Purpose | Implementation |
|---|---|---|
| ICC | Interpretive Commitment Control — parse intent and detect constraint conflicts | Stub (model-based) |
| ASTG | Assumption Stress Test — enumerate hidden premises | Stub (model-based) |
| BSG | Bias Selection Gate — require explicit optimization frame | Partial rules |
| EDH | Echo Detection Heuristic — detect repetition patterns | Stub (embedding) |
| SPM | Session Pattern Monitor — detect emerging interaction signals | Rules |
| PRAP | Pre-Reasoning Assurance Protocol — aggregate gate outcomes | Rules |
Ingress halts immediately if any gate fails.
These gates execute after the model returns.
| Gate | Purpose | Implementation |
|---|---|---|
| OSCL | Outcome-Scored Self-Critique Loop | Stub |
| FINALIZATION | Structured output validation | Rules |
| TELEMETRY | Audit logging | Full |
The FINALIZATION gate enforces:
• structured response format • scope compliance • lexical compliance • option labeling discipline • governance stack rules
Each interaction generates a TelemetryTurnRecord containing:
• gate outcomes • model latency • reasoning metadata • session state • compliance results
Telemetry can be rendered using:
POST /telemetry/render
Supported levels:
minimal standard full
The runtime includes special command handlers:
| Signal | Behavior |
|---|---|
| NFAR / no further action required | Output: Standing by. |
| EOS | Session snapshot + reset |
| Trace This | Emit audit ledger |
The following components are placeholders pending training corpus completion.
| Component | File | Stub Behavior |
|---|---|---|
| ICC classifier | classifier/icc_classifier.py | generic intent classification |
| ASTG classifier | classifier/astg_classifier.py | no assumptions detected |
| EDH similarity | classifier/edh_similarity.py | similarity score 0 |
| OSCL scorer | egress/gates/oscl.py | fixed score 0.85 |
Rules-based governance gates are fully implemented.
| Requirement | Version |
|---|---|
| Python | 3.11+ |
| Redis | 6+ |
| pip | 23+ |
cd
python -m venv .venv source .venv/bin/activate
pip install -r requirements.txt pip install -e .
cp .env.example .env
redis-server --daemonize yes
uvicorn cgs_runtime_wrapper.api.main:app --reload --port 8000
Server runs at:
cp .env.example .env
docker compose up --build
Services:
| Service | Port |
|---|---|
| Redis | 6379 |
| Runtime Wrapper | 8000 |
Shutdown:
docker compose down
All requests must include an API key.
Header:
X-API-Key
Example:
curl -X POST http://localhost:8000/turn -H "Content-Type: application/json" -H "X-API-Key: your-api-key"
Model adapters implement the ModelAdapter interface:
class ModelAdapter(ABC): async def call_model(self, prompt: str, session_id: str) -> tuple[str, int]
Supported adapters:
• stub • Claude • OpenAI • custom implementations
This repository is distributed under the Source Available License v4.
• Non-commercial use permitted.
• Commercial licensing available under LICENSE-COMMERCIAL.
Contact the project maintainer for enterprise licensing.
Certain architectural concepts implemented in the CoSyn Runtime Wrapper may qualify for patent protection or defensive publication.
The project maintains a documented patent and defensive publication strategy.
See:
PATENTS.md
for details regarding:
- defensive publication strategy
- patent candidate categories
- contributor notice regarding patent inclusion
- commercial patent licensing boundaries
CoSyn is a governance framework and runtime architecture created by the project author. The name and associated marks may be protected under trademark law.
FastAPI + Redis governance middleware implementing the CoSyn runtime governance pipeline.
Deterministic ingress and egress governance gates on every LLM interaction.
Audit-traceable. Runtime-enforced. Governance-bound.