The time machine
for AI systems.
Capture, replay, and audit AI-agent executions. Evidence infrastructure for reproducible AI. Local-first, cryptographically signed.
01 / THE SYSTEM
One fabric. Six domains of evidence.
NovaFabric is not a single tool — it's a coordinated system of ~425,000 lines across six domains: capture, an evidence model, analysis, cryptographic trust, governance, and serving. Every piece composes.
Wrap any subprocess. Record everything.
- Framework Adapters
- Multi-Target Runners
- Capture Orchestrator
- Wire-Level Hooks
- Secret Scanner
The portable, self-contained unit of record.
- Run Capsule
- Event Envelope v1
- Evidence Fabric
- Capsule Schemas
Replay, diff, trace lineage, evaluate.
- Replay Engine
- Diff Engine
- Lineage Graph
- Eval Suites
- Knowledge Graph
Sign, timestamp, and verify evidence.
- NovaSeal Signing
- Evidence Bundle
- Merkle Log
Policy gates, approvals, compliance.
- Promotion (Maker-Checker)
- Policy Engine (OPA)
- Compliance & Export
- Governance & Judge
- Asset Registry
Multi-user API, dashboard, topology.
- Storage Layer
- Server (REST API)
- Dashboard
- Live Topology
What the lab
is building.
Six active research and engineering areas, each labeled with its current maturity. Labels come from the NovaFabric source-of-truth docs and are updated only when evidence changes.
no capability uses stable until v1.0 schema freeze
Evidence + replay layer
Local-first run capsules for capture, replay, diff, validation, and audit workflows. CLI capture, SDK decorator, API/MCP proxy, and OTel GenAI semconv all captured.
NovaSeal trust layer
Tamper-evident DSSE/RFC 3161 sealing in local profile. Cryptographic SoD enforced. Enterprise HSM/X.509 PKI not claimed.
Lineage and provenance
SQLite lineage is experimental (10M-edge KuzuDB benchmark done). KuzuDB backend is prototype. Lineage federation is experimental with legal review pending.
Real runs.
Real capsules.
Each entry is a real project using novafabric to capture and replay AI workflows. The capsule snippet shows what a typical run produces.
submit your capsule →Open a PR on GitHub — add your capsule snippet to the showcase.
Standalone tools
in the ecosystem.
Behavioral test suite verifying object storage backends meet WORM requirements. 10 mandatory test cases covering immutability, retention locks, legal holds, and lifecycle exemptions.
Cluster-scale OpenTelemetry collector for capsule events, spans, and metrics. Deployable to Kubernetes via DaemonSet or SLURM via prolog/epilog. Handles millions of events per node.
Web UI for capsule inspection, lineage visualization, and asset registry browsing. Ships embedded in nova serve. 13 tabs covering all major primitives.
One lab.
The full stack of agentic AI.
The next decade of computing will be defined by AI agents operating at the scale of HPC clusters — running Claude and frontier models on thousands of nodes, coordinating across scientific workflows, and making consequential decisions autonomously.
We're building the infrastructure layer for that world: tools for capturing executions, monitoring fleets, governing behavior, and making every agent run reproducible, auditable, and trustworthy.
Open research. Apache-2.0. One engineer. No VC. No data collection.
Portable by design
Capsules are self-contained directories. No proprietary formats, no vendor lock-in. A capsule you create today will be readable in a decade.
Non-invasive capture
Patch SDK call sites and HTTP transports at runtime. No code changes, no decorators required. Works with any Python AI framework.
Evidence-grade integrity
NovaSeal uses DSSE + RFC 3161 timestamps. Verification is always offline and deterministic. Meets financial and clinical compliance requirements.
Scale from laptop to cluster
SQLite for local development. Postgres + RBAC for multi-user server mode. Go collector for SLURM/Kubernetes cluster deployments.