dof-sdk 0.5.1

Deterministic Observability Framework — formal governance, privacy benchmarks, and adversarial testing for multi-agent LLM systems

DOF-MESH -- Deterministic Observability Framework

Verify. Prove. Attest. · Mathematics, not promises.

DOF-MESH v0.5.1 | 8 Chains | Enigma Group · Medellín, Colombia

PyPI | Getting Started | Documentation | On-Chain Proof | @Cyber_paisa

---

The Problem

Your AI agent just made a decision. You have logs. Logs the agent wrote about itself.

That is not proof. That is testimony.

The industry's fix is AI watching AI — until your watchdog hallucinates that everything is fine while an attacker manipulates your agent. We tested 10 frontier models. None could reliably govern themselves. Every one improved 6% to 50% only when a deterministic layer enforced the rules from outside.

Regulators are already asking: "How do you prove your AI made that decision correctly?" Nobody has a standard answer yet.

Our Solution

A deterministic verification layer for autonomous AI agents.

Every decision is translated into formal constraints, verified mathematically in under 30ms, and anchored on-chain as a cryptographic receipt. Not logs. Not monitoring. A proof.

Identity → Task → LLM → Governance → Z3 Proof → On-Chain → Supervisor

Z3 is the theorem prover used in safety-critical systems where failure is not an option. It doesn't estimate compliance. It proves it.

Verified by formal constraints. Auditable by anyone. Immutable once recorded.

DOF doesn't monitor. DOF proves.

---

What DOF-MESH Is — and Is Not

DOF-MESH is not an agent framework. You don't replace your agents with ours.

It is a verification layer. Your agent runs. DOF intercepts the output, checks it against deterministic rules and Z3 formal proofs, and writes a cryptographic receipt on-chain before the result is accepted. One layer on top of whatever you already have.

LangChain. CrewAI. AutoGen. A custom agent. It doesn't matter. DOF governs it.

The framework is what you install. Everything else in this repo is validation that it works.

---

Quick Install

pip install dof-sdk

from dof import DOFVerifier

verifier = DOFVerifier()
result = verifier.verify_action(
    agent_id="apex-1687",
    action="transfer",
    params={"amount": 500, "token": "USDC"}
)
# → verdict: "APPROVED"
# → z3_proof: "4/4 VERIFIED [GCR_INVARIANT:VERIFIED | SS_FORMULA:VERIFIED | ...]"
# → latency_ms: 8.2
# → attestation: "0x44b45cd026916c..."

Or run the full mesh:

git clone https://github.com/Cyberpaisa/DOF-MESH.git && cd DOF-MESH
cp .env.example .env
docker-compose up -d --build

---

Architecture

+=====================================================================+
|                        DOF-MESH  v0.5.0                             |
|                                                                     |
|  +---------------------------------------------------------------+  |
|  |                      INTERFACE LAYER                          |  |
|  |    CLI | A2A Server | Telegram | Streamlit Dashboard | Voice  |  |
|  +---------------------------------------------------------------+  |
|                                  |                                  |
|  +---------------------------------------------------------------+  |
|  |                     EXPERIMENT LAYER                          |  |
|  |   ExperimentDataset | BatchRunner | Schema | Parametric Sweep |  |
|  +---------------------------------------------------------------+  |
|                                  |                                  |
|  +---------------------------------------------------------------+  |
|  |                    OBSERVABILITY LAYER                        |  |
|  |    RunTrace | StepTrace | 5 Derived Metrics | JSONL Audit     |  |
|  +---------------------------------------------------------------+  |
|                                  |                                  |
|  +-------------------------------+-------------------------------+  |
|  |       GOVERNANCE CORE         |      VERIFICATION CORE        |  |
|  |                               |                               |  |
|  |  ConstitutionEnforcer         |  Z3Verifier (4/4 PROVEN)      |  |
|  |  HARD rules: block            |  Formal invariant proofs      |  |
|  |  SOFT rules: warn             |  keccak256 proof hashes       |  |
|  |  Zero LLM in governance       |  ASTVerifier + Z3Gate         |  |
|  +-------------------------------+-------------------------------+  |
|                                  |                                  |
|  +---------------------------------------------------------------+  |
|  |                    CORE INFRASTRUCTURE                        |  |
|  |                                                               |  |
|  |  crew_runner.py       Orchestration, retry x3, crew_factory   |  |
|  |  providers.py         TTL backoff (5/10/20m), provider chains |  |
|  |  supervisor.py        MetaSupervisor weighted scoring         |  |
|  |  memory_manager.py    ChromaDB + HuggingFace embeddings       |  |
|  |  autonomous_daemon.py Perceive, Decide, Execute, Evaluate     |  |
|  |  node_mesh.py         NodeRegistry + MessageBus + MeshDaemon  |  |
|  +---------------------------------------------------------------+  |
|                                  |                                  |
|  +-------------------------------+-------------------------------+  |
|  |      AGENT RUNTIME LAYER      |       ON-CHAIN LAYER          |  |
|  |      (any framework)          |                               |  |
|  |                               |  DOFProofRegistry             |  |
|  |      LangChain · CrewAI       |  Avalanche · Base · Celo      |  |
|  |      AutoGen · Custom         |  Polygon · SKALE · Conflux    |  |
|  |                               |  8 chains · 30+ attestations  |  |
|  +-------------------------------+-------------------------------+  |
+=====================================================================+

---

Core Components

Component	What It Does
ConstitutionEnforcer	Deterministic governance -- HARD rules block, SOFT rules warn. Zero LLM involvement. ~50 token constitution injected per agent.
Z3Verifier	4 mathematical theorems formally PROVEN every cycle. Generates keccak256 proof hashes for on-chain recording.
Z3Gate	Neurosymbolic gate -- LLM proposes, Z3 verifies. APPROVED / REJECTED / TIMEOUT / FALLBACK.
MetaSupervisor	Weighted quality scoring: Q(0.40) + A(0.25) + C(0.20) + F(0.15). Outputs ACCEPT, RETRY, or ESCALATE.
DOFProofRegistry	Multi-chain attestation engine. Writes proof receipts to 8 chains. Verifiable by any third party.
MeshDaemon	29 nodes + threshold consensus. Byzantine guard, CRDT memory, constitution hash beacon.
ProviderManager	LiteLLM router across 7+ LLMs. TTL backoff, automatic failover, Thompson Sampling.

---

The Numbers

We wrote 4,157 tests before calling it a framework. Most projects write them after something breaks.

The proofs aren't in our database. They're on 8 public blockchains. Go check — nobody has to take our word for it.

We tested 10 of the best AI models in the world. None could govern themselves. All improved 6% to 50% the moment a deterministic layer enforced the rules.

The same verification engine used to certify commercial aircraft and nuclear reactors now certifies your agents. The one technology where "probably correct" was never an option.

No model decides if your agent broke a rule. A deterministic function does. Every time. Same answer.

Metric	Value
Unit tests	4,191
Autonomous cycles	238+
On-chain attestations	30+
Chains (mainnets)	7
Core modules	142
Lines of code	57,000+
Z3 theorems	4/4 PROVEN
Hierarchy patterns (Z3)	42 PROVEN
LLM providers	7+ (Cerebras, Groq, DeepSeek, Gemini, NVIDIA, SambaNova, Zhipu)
Governance mode	100% deterministic, 0% LLM

---

Pre-Execution Governance — The Architectural Leap

Most AI governance frameworks check what happened. DOF checks what's about to happen.

Every tool call passes through a 3-layer pipeline before execution is permitted:

Agent proposes action
        │
        ▼
┌─────────────────────┐
│  Layer 1            │   ConstitutionEnforcer
│  HARD BLOCK         │   Blocked tools rejected instantly.
│                     │   No LLM. No tokens. No exceptions.
└─────────┬───────────┘
          │ PASS
          ▼
┌─────────────────────┐
│  Layer 2            │   Z3Gate (neurosymbolic)
│  FORMAL VERIFY      │   4 invariants proven mathematically.
│                     │   keccak256 proof hash generated.
└─────────┬───────────┘
          │ APPROVED
          ▼
┌─────────────────────┐
│  Layer 3            │   PostToolUse
│  ATTEST             │   Attestation written to JSONL + chain.
│                     │   Immutable. Auditable. Verifiable.
└─────────────────────┘

from core.tool_hooks import ToolHookPipeline, GovernanceViolation

hook = ToolHookPipeline()

# Before any tool executes:
pre = hook.pre_tool_use("transfer", "amount=500 token=USDC", "apex-1687")
if not pre.allowed:
    raise GovernanceViolation(pre.reason)

# After execution:
post = hook.post_tool_use("transfer", result, "apex-1687", pre_result=pre)
print(post.attestation_hash)  # 0x44b45cd026916c...

34 tests. 0 failures. Every tool call governed before it runs.

---

Tech Stack

Layer	Technology
Core Framework	Python 3.11+
Formal Verification	Z3 Theorem Prover -- 4/4 invariants PROVEN
Blockchain	web3.py · Avalanche · Base · Celo · Polygon · SKALE · Conflux
LLM Routing	LiteLLM Router (7+ providers, TTL backoff, Thompson Sampling)
SDK	dof-sdk on PyPI -- pip install dof-sdk
Vector Memory	ChromaDB + HuggingFace embeddings (all-MiniLM-L6-v2)
Persistence	JSONL audit logs -- zero external telemetry dependencies
Protocols	A2A · MCP · ERC-8004 · ERC-8183 · x402
Container	Docker Citadel (OrbStack) -- read-only agent sandbox

---

On-Chain Attestation

DOFProofRegistry deployed on 8 chains (3 mainnet + 5 testnet). Verified on-chain 03 Apr 2026.

Chain	Contract	Status
Avalanche C-Chain (43114)	0x154a3F49a9d28FeCC1f6Db7573303F4D809A26F6	✅ mainnet
Base Mainnet (8453)	0x4e54634d0E12f2Fa585B6523fB21C7d8AaFC881D	✅ mainnet
Celo Mainnet (42220)	0x35B320A06DaBe2D83B8D39D242F10c6455cd809E	✅ mainnet
Avalanche Fuji (43113)	0x0b65d10FEcE517c3B6c6339CdE30fF4A8363751c	✅ testnet
Base Sepolia (84532)	0x7e0f0D0bC09D14Fa6C1F79ab7C0EF05b5e4F1f59	✅ testnet
Conflux Testnet (71)	0x554cCa8ceBE30dF95CeeFfFBB9ede5bA7C7A9B83	✅ testnet
Polygon Amoy (80002)	0x0b65d10FEcE517c3B6c6339CdE30fF4A8363751c	✅ testnet
SKALE Base Sepolia (324705682)	0x4e54634d0E12f2Fa585B6523fB21C7d8AaFC881D	✅ testnet · zero gas

Roadmap: Polygon mainnet · Conflux eSpace · SKALE Base mainnet

ERC-8004 Agent:    #1687 (Apex) · #1686 (AvaBuilder)
ERC-8183:          DOFEvaluator.sol → complete() / reject()
Proof hash:        keccak256(Z3 proof transcript) -- verifiable via verifyProof()
Gas cost:          $0.01/tx · SKALE chains: zero gas · Merkle batch: 10K attestations = 1 tx

---

How The Pipeline Works

1. IDENTITY        Agent authenticates via ERC-8004 identity
                   |
2. TASK            Discovery loop finds next task (or via A2A / Telegram)
                   |
3. LLM INFERENCE   LiteLLM routes to best available provider
                   Fallback chain: Cerebras → Groq → DeepSeek → Gemini → NVIDIA
                   |
4. GOVERNANCE      ConstitutionEnforcer evaluates output
                   HARD rules: block on violation | SOFT rules: warn and log
                   ZERO LLM in this step -- purely deterministic
                   |
5. Z3 PROOF        Z3Verifier generates formal mathematical proof
                   4 invariants checked, proof hash = keccak256(proof)
                   |
6. ON-CHAIN        DOFProofRegistry writes attestation to 8 chains
                   ERC-8004 receipt with agent ID + proof hash
                   |
7. SUPERVISOR      MetaSupervisor scores: Q(0.40)+A(0.25)+C(0.20)+F(0.15)
                   Decision: ACCEPT → next cycle | RETRY → re-execute | ESCALATE → human

---

Winston Experiment — Multi-Model Validation

10 frontier models evaluated with and without the Winston framework. Deterministic scorer, 0 LLMs. Externally validated on Adaline (200M+ API calls/day).

Model               BLUE(Winston)  RED(baseline)  Delta
────────────────────────────────────────────────────────
DeepSeek-V3           88.7          38.7         +50.0
GLM-4.5               90.0          42.7         +47.3
Mistral-Large         78.7          41.3         +37.4
Claude Sonnet         90.0          56.0         +34.0
ChatGPT-4o            88.7          63.0         +25.7
Gemini-2.5Pro         84.7          71.3         +13.4
────────────────────────────────────────────────────────
Average                                          +26.1

Full data: experiments/winston_vs_baseline/

---

Documentation

Document	Description
Getting Started	Installation, first run, environment setup
System Architecture	Full governance pipeline with latency per layer
Case Study — Apex #1687	238 autonomous cycles · 0 incidents · 30+ attestations
Winston Experiment	Raw data: +26.1% average across 10 frontier models
Multichain Deployment	Deploy DOFProofRegistry to 8 chains
SKALE Integration	Zero-gas chains · x402 · BITE · IMA bridge
Attestations	On-chain proof records — publicly verifiable
Full Index	All public documentation

---

Repository Structure

equipo-de-agentes/
  core/                     # 142 modules -- the framework engine
    governance.py            # ConstitutionEnforcer, HARD/SOFT rules
    z3_verifier.py           # 4 theorems formally PROVEN
    z3_gate.py               # Neurosymbolic gate (APPROVED/REJECTED/TIMEOUT)
    supervisor.py            # MetaSupervisor weighted scoring
    providers.py             # LiteLLM router, TTL backoff
    autonomous_daemon.py     # 4 phases: Perceive→Decide→Execute→Evaluate
    node_mesh.py             # NodeRegistry + MessageBus + MeshDaemon
    claude_commander.py      # 5 modes: SDK, Spawn, Team, Debate, Peers
    ...
  agents/                   # agents running on DOF (validation, not the product)
  contracts/                # DOFProofRegistry.sol (deployed on 8 chains — 3 mainnet + 5 testnet)
  integrations/             # CrewAI, AgentKit, Virtuals, SKALE, Tempo
  tests/                    # 4,157 unit tests
  docs/                     # public documentation
  logs/                     # JSONL audit trails (append-only)
  experiments/              # Winston vs baseline raw data

---

Built By

@Cyber_paisa · Telegram

DOF-MESH -- The production laboratory where deterministic AI governance is built. 4,157 tests. 8 chains (3 mainnet + 5 testnet). 142 modules. Mathematics, not promises.

---

DOF-MESH -- Deterministic Observability Framework

Verify. Prove. Attest. · Mathematics, not promises.

Enigma Group · Medellín, Colombia