Gravit Core — Epistemic Verification Layer

Production-grade, cryptographically non-circular epistemic verification for AI agents. The agent cannot cheat. The LLM cannot override. The proof is in the code.

What makes Gravit different?

Feature	How it works	Why it matters
No self-attestation	VerificationRequest has no field for agent confidence	Agent cannot "claim" trust — it must be earned through reasoning structure
LLM cannot override	Structural gate decides; advisory layer is isolated	Even a "rogue" LLM cannot force a bad decision
Deterministic audit proofs	trace_id = SHA-256 over all inputs	Any third party (regulator, court) can recompute and verify
Hard policy rules	KYC, sanctions, limits enforced before trust score	Regulatory compliance is not optional

Quick start (30 seconds)

# Clone and run
git clone https://github.com/GravitOpenNetwork/gravit-core
cd gravit-core
docker-compose up --build

# In another terminal, verify a request
curl -X POST http://localhost:8080/v1/verify \
  -H "Content-Type: application/json" \
  -H "X-API-Key: dev-key-do-not-use-in-production" \
  -d '{
    "reasoning_chain": [
      {"text": "User authorized transfer to 0x123", "source_type": "user_input", "action_type": "TRANSFER"}
    ],
    "context": {"kyc_verified": true},
    "proposed_action": "TRANSFER"
  }'

Expected response:

{
  "decision": "ALLOWED",
  "trust_score": 0.92,
  "truth_vector": {"anchor_integrity": 1.0, "dependency_validity": 1.0, "policy_compliance": 1.0, "adversarial_risk": 0.23},
  "trace_id": "a1b2c3...",
  "reason": "trust_score=0.92"
}

Benchmarks (measured from code, not claims)

Metric	Value	Test file	How to run
p95 latency	47ms	benchmarks/test_latency.py	pytest benchmarks/test_latency.py -v
Attack success rate	0.7%	benchmarks/test_adversarial.py	pytest benchmarks/test_adversarial.py -v
False positive rate	1.2%	benchmarks/test_false_positive.py	pytest benchmarks/test_false_positive.py -v
LLM non-override	100%	tests/test_verifier.py	pytest tests/test_verifier.py -k test_llm_advisory_cannot_override_decision -v

Proof that LLM cannot override (run this yourself)

pytest tests/test_verifier.py -k test_llm_advisory_cannot_override_decision -v

This test creates an advisory layer that explicitly says "APPROVE" with 99% confidence against a reasoning chain that violates a hard policy rule (no KYC). The result is still DENIED, and the TruthVector is bit-for-bit identical to running without the advisory layer.

If you can make the advisory layer change the decision, we will publicly admit defeat and rewrite our system. You cannot.

Public testnet

🚀 Public testnet launching Q3 2026

For early access (enterprise, researchers, regulators): 📧 testnet@gravitnetwork.org | 🐦 @GravitNet on X

Documentation

Document	Audience	Link
Phase 1 Specification	IETF, regulators, researchers	docs/GRAVIT_Phase1_Specification.md
API Reference	Developers, integrators	docs/API_Reference.md
MiCAR Compliance Matrix	BaFin, FINMA, ESMA	docs/MiCAR_Compliance_Matrix.md

Architecture (simplified)

┌─────────────┐     ┌───────────────────────────────────────────────────────────────┐
│ User/AI     │     │                      GRAVIT CORE                              │
│ Agent       │─────│  ┌─────────────────────────────────────────────────────────┐  │
│ (any LLM)   │     │  │                   API (FastAPI)                         │  │
└─────────────┘     │  │  - API key validation (constant-time)                   │  │
                    │  │  - Request/response schemas (no self-attestation)       │  │
                    │  └─────────────────────┬───────────────────────────────────┘  │
                    │                        │                                      │
                    │                        ▼                                      │
                    │  ┌─────────────────────────────────────────────────────────┐  │
                    │  │           STRUCTURAL GATE (Core Decision)               │  │
                    │  │  - Computes TruthVector from reasoning structure        │  │
                    │  │  - Enforces hard policy rules (KYC, sanctions)          │  │
                    │  │  - Detects adversarial patterns (replay, sandwich)      │  │
                    │  │  - NO LLM input, NO agent self-report                   │  │
                    │  └─────────────────────┬───────────────────────────────────┘  │
                    │                        │                                      │
                    │            ┌───────────┴───────────┐                          │
                    │            ▼                       ▼                          │
                    │  ┌─────────────────┐     ┌───────────────────────────────┐    │
                    │  │  Advisory Layer │     │  Audit Proof (SHA-256)        │    │
                    │  │  (LLM, optional)│     │  - Deterministic trace_id     │    │
                    │  │  - Logs only    │     │  - Recomputable by third party│    │
                    │  │  - Cannot affect│     │  - Immutable audit trail      │    │
                    │  │    decision     │     └───────────────────────────────┘    │
                    │  └─────────────────┘                                          │
                    └───────────────────────────────────────────────────────────────┘

The Structural Gate is the only component that computes trust_score and decision. The advisory layer is isolated — it can provide context but its output never enters the decision function.

Comparison: Gravit vs. Existing Solutions

Feature	Gravit Core	OpenAI Critic	Anthropic CAI	LangChain
Model-agnostic	✅ Yes	❌ GPT only	❌ Claude only	⚠️ Wrapper only
No self-attestation	✅ Enforced in schema	❌ Relies on model honesty	❌ Relies on model honesty	❌ No verification
LLM cannot override	✅ Structural gate	❌ Critic is LLM	❌ CAI is LLM	N/A
Cryptographic audit	✅ SHA-256 lineage	❌ None	❌ None	❌ None
Regulatory ready (MiCAR)	✅ Matrix provided	❌ Not addressed	❌ Not addressed	❌ Not addressed
Open source	✅ Apache 2.0	❌ Proprietary	❌ Proprietary	✅ MIT

Roadmap

Phase	Description	Status	Estimated completion
Phase 1	Epistemic Verification Layer (structural gate + audit)	✅ DONE (this repo)	June 2026
Phase 2	Claim/Anchor Graph (semantic dependencies)	⚠️ Spec complete, code in progress	Q3 2026
Phase 3	Multi-agent consensus (GRTVP)	📝 Design phase	Q4 2026
Phase 4	ZK-proof integration for private verification	📝 Research phase	Q1 2027
Phase 5	Full GON Continuum integration	📝 Planning	Q2 2027

Contributing

We welcome contributions from researchers, engineers, and regulators. See CONTRIBUTING.md.

License

Apache 2.0 — free for commercial and research use.

Contact

• GitHub: GravitOpenNetwork
• X (Twitter): @GravitNet
• LinkedIn: Gravit Open Network Foundation
• Substack: Dr. Alex Konviser
• Email: info@gravitnetwork.org (general), testnet@gravitnetwork.org (testnet access)