Zilli v3.0 - MOM Meta-Intelligence System

Overview

Zilli v3.0 is a self-evolving engineering platform purpose-built for AI to autonomously design, develop, test, optimize, and deploy MOM (Model of Models) meta-intelligence systems. Its core philosophy fuses swarm intelligence orchestration (MOM technology) with dual-model collaborative engineering (Zilli's original architecture), delivering high-performance, low-cost, data-sovereign private MOM deployments for data-sensitive enterprises.

In Zilli v3.0, MOM is not just the final product — its key techniques (task DAG decomposition, meta-evaluation models, multi-objective optimization) are applied back into Zilli's own development pipeline, creating a "building MOM with MOM" closed loop.

Zilli is maintained by Ethercoin, running on the Ethercoin decentralized compute network (12,847+ nodes / 356 PFLOPS) with ZK+TEE+PoRW verification for trusted training and inference.

Dual-Layer Architecture

Zilli v3.0 employs a dual-layer architecture: the top meta-orchestration layer (MOM kernel) handles swarm intelligence coordination; the bottom development pipeline layer (Zilli engine) handles automated tool production.

Top Layer: MOM Meta-Orchestration (Runtime)

An operating system above models: decomposes complex requests into parallelizable task DAGs, dynamically selects optimal model ensembles based on capability profiles, predicts performance via meta-evaluation models, and achieves Pareto optimality across quality, cost, latency, privacy, and throughput.

Bottom Layer: Zilli Pipeline (Build-time)

Dual-model collaboration (SOTA Planner + cost-effective Executor) executes the Plan → Generate → Verify → Reflect → Evolve loop. The Planner handles deep reasoning and reflection (<5% of calls), while the Executor handles 95% of generation work, continuously evolving through distillation and RL.

Five-Phase Execution Pipeline

Plan — SOTA-driven

Decomposes high-level requirements into task DAGs with I/O schemas, acceptance criteria, and permitted tools. Generates orchestration files marking key nodes requiring Planner assistance.

Generate — Executor-driven

Executor runs leaf tasks in parallel; Planner correction mode activates when confidence drops below threshold, performing single-pass Critic-Edit. Produces code, configs, tests, and documentation.

Verify — Layered automation

Static analysis → sandbox testing → behavioral consistency checks → Planner review (high-risk tasks). Auto-retry on failure (up to 3 times), injecting error context each round.

Reflect — Deep-dive by Planner

Analyzes the full execution trajectory, generates root cause classification (planning quality / generation quality / environment / requirement drift), and extracts success patterns into the trajectory memory store.

Evolve — Closed-loop learning

Three parallel paths: instant strategy updates, training data accumulation, and periodic distillation with RL training. The system improves with every cycle.

MOM Meta-Orchestration Core

The MOM kernel is the intelligence center of Zilli v3.0, responsible for optimal multi-model swarm intelligence scheduling at runtime.

Task Decomposer

Recursively breaks complex requests into DAG sub-tasks with topological sorting, critical-path analysis, and parallel-group scheduling. Exports Mermaid diagrams for visualization.

GPS-MOM Predictive Routing

Four-step decision pipeline: PPM predictor (task-family classification + difficulty scoring + LRU cache) → 3-tier strategy selection (ECONOMY / STANDARD / ENHANCED) → profile filtering → optimal model binding. Prediction latency <10ms, cache hit rate >60%.

Model Capability Profiler

Tracks a 5-dimension capability vector per model with ELO ratings, Bayesian EMA online updates, and softmax Thompson sampling. Atomic JSON persistence.

Bayesian Meta-Evaluator

Gaussian conjugate prior updates estimate the true error distribution — robust even on small samples. Drift detection, feature importance, and calibrated confidence intervals built in.

Multi-Objective Optimization Engine

NSGA-II solves for Pareto-optimal model-to-task bindings across five dimensions: quality, cost, latency, privacy, and throughput. Supports weight-based preferences and constraint solving.

Dynamic Cost Control

Three-tier budget management: monthly budget + hourly quota + emergency mode. SOTA model calls are hard-capped below 5% (max_sota_ratio enforced) and 10% of total cost.

Enterprise Privacy Governance

Zilli's privacy module provides end-to-end data governance for sensitive enterprise workloads. Five data classification levels drive automated policy enforcement across the entire Agent lifecycle.

Five-Level Data Classification

PUBLIC / INTERNAL / CONFIDENTIAL / RESTRICTED / REGULATED. Automatic PII/PHI detection elevates data containing personal identifiers to at least CONFIDENTIAL.

PrivacyGatekeeper

Makes local/cloud/deny decisions based on data classification and tenant policy. RESTRICTED and REGULATED data is forced to local execution.

Compliance Reports

Generates GDPR, HIPAA, SOC2 compliance reports from existing audit trails (JSONL format) out of the box. No additional storage infrastructure required.

Training & Distillation

The Executor model evolves continuously through SFT + RL (CISPO/GRPO) + distillation, achieving "cost-effective AI writes AI".

Training Data Pipeline

Automatically collects successful trajectories as positive samples; failed tasks are corrected by Planner reflection; human feedback boosts sample weights.

Layered Experience Replay

Golden trajectories (reward > 0.8) are stored directly; low-reward trajectories are corrected by Planner and stored in the failure reflection pool. Mixed sampling for training.

Executor-only Evaluation

Planner is completely disabled to verify the Executor's standalone capability. Pass criteria: core task success rate ≥85%, cost < 5% of SOTA.

Loop Engine & Self-Improvement

A generic Retry → Verify → Correct loop framework, paired with the Self-Harness meta-loop for system-level self-diagnosis and self-repair.

Loop Engine

Generic LoopRunner + MetaLoopRunner two-level meta-loop. 5 verifiers (TestSuite / Predicate / ExternalModel / Skill / Composite) × 3 triggers (fixed-interval / event / dynamic-interval).

Self-Harness Meta-Loop

Three stages: WeaknessMiner failure clustering → bounded HarnessEdit proposals → held-in / held-out split validation. Improvements are auto-accepted at ≥5% pass-rate gain with no regressions.

Skill Evolution Engine

4 evolution strategies (prompt optimization / error handling / boundary refinement / tool augmentation) + code-fingerprint diversity gating (6-dim weighted Jaccard + n-gram) to prevent population collapse.

Unknowns Discovery (Fable Method)

Blind-spot passes surface unknown unknowns → interview question generation → implementation notes tracking deviations → quizzes verifying understanding. "Finding your unknowns" as a reusable engineering capability.

SWE-bench Repair Agent

Fully autonomous bug-fix loop: reproduce → explore → diagnose → fix → verify, with optional Docker sandbox isolation and iterative convergence.

L6 Swarm Intelligence (RFC-006)

Multi-agent collaborative orchestration: task decomposition (DAG longest-chain depth validation) → parallel execution (deterministic sink convergence) → routing (exact-match priority + capability profiling) → consensus (with human escalation channel). One-command orchestration via the zilli swarm CLI.

Engineering Maturity

Current release: v1.1.0 — production-grade engineering quality.

Tests & Quality

1,380 tests passing with 0 warnings; 92% coverage; ruff clean at 0 errors; pyright clean at 0 errors. GitHub Actions CI covers lint + typecheck + multi-version testing.

End-to-End Loop

Full Evolve → Train → Deploy → Monitor pipeline (EvolveToTrainPipeline) with checkpoint resume and a Champion-Challenger arena using statistical significance testing.

Feedback Loop

FeedbackCollector async queue + JSONL batch persistence (100-record trigger), LLM-as-Judge quality scoring with heuristic fallback, and online PPM weight training.

Tech Stack

Languages

Python 3.12+ (Rust helper kernel zilli_hotpath PyO3 binding v0.3.0: PPM prediction at 7.7µs/call ≈ 20.2× faster than pure Python, 100% functional parity)

Workflow Engine

Celery + Redis (persistent DAG execution, task retries, result callbacks)

RL Algorithms

CISPO, GRPO (Monte-Carlo + GAE advantage estimation)

Vector Store

ChromaDB (semantic retrieval, metadata filtering, collection management)

Serving & Console

FastAPI (OpenAI-compatible endpoints, streaming, API-key auth) + Streamlit console (audit browsing, cost monitoring, PPM stats)

Sandbox

Docker SDK dynamic containers (isolated test execution)

Repository

github.com/ethercoinai/Zilli →