open source · Apache-2.0 · self-hosted
Replayable execution and evidence infrastructure.
Open-source, self-hosted replay and evidence infrastructure for AI agents and agentic systems. It captures agent executions as portable Run Capsules with no application-code changes, then adds replay, behavioral/structural diff, lineage, cryptographic provenance, assurance and audit. The same capture layer can wrap general commands and supports local, Docker, Kubernetes and SLURM/HPC environments. Observability tells you what happened; NovaFabric focuses on whether a past execution can be replayed, compared, and proven.
// the problem
A modern AI system isn't a single model — it's a drifting graph of prompts, models, tools (HTTP APIs, MCP servers, shell commands), memory/RAG context, datasets, environments, and execution targets (laptop, container, K8s pod, SLURM job). Each layer drifts at its own rate, and when something misbehaves you can't tell which layer moved.
The ecosystem has excellent telemetry for what happened — and almost nothing for what would happen if I ran this again, under the same conditions, today. Reproducibility isn't a nice-to-have: it's the precondition for debugging, regression detection, governance, and trust.
// what it does
The identity layer — name@version for models, agents, prompts, tools, datasets, evals, and deployments.
A portable per-run snapshot — inputs, outputs, model calls, tool calls, a full environment lock, and the trace — written on success and failure. A folder you own, not a row in someone's cloud.
Reconstruct a past run in four honest modes — exact, mocked, semantic, and forensic — each making a specific, falsifiable promise.
Causation — which assets and prior runs contributed to this run, and what depends on it.
A signed, audit-ready export — capsule + lineage + attestations + redaction proof — that a third party can verify without NovaFabric installed.
nova capture python my_agent.py — per-SDK hooks, wire-level hooks (httpx/requests/aiohttp/urllib3), and transparent HTTP/MCP proxies record any command. No application changes; all patches removed after the run.
// the honest contract
Deterministic re-execution against the same local model, container, data, and environment — bit-identical where deterministic.
Re-run the command with every model/tool call served from the capsule cache — no model called, no tool invoked, fully deterministic.
Re-execute and compare meaning (via a judge model), not tokens — for remote closed-weight LLMs whose weights and routing change without notice.
Inspect the captured run without re-executing anything — read-only, no network, no side effects.
The taxonomy is deliberately honest: NovaFabric does not claim deterministic exact replay of remote closed-weight models. Replay modes make the promise explicit rather than pretending a changing provider is deterministic.
// trust & governance
Every sealed capsule binds a DSSE signature (ECDSA P-256 / Ed25519), an RFC 3161 trusted timestamp (eIDAS-qualified path for the EU), and an append-only Merkle log with inclusion proofs.
Local keys, AWS KMS, Azure Key Vault, GCP KMS, or keyless Sigstore (Fulcio + Rekor). nova verify re-checks signature, timestamp, and log integrity.
An OPA/Rego policy engine (allow / deny / require-human / require-batch) and dual-approval with cryptographic separation of duties — proposer ≠ approver, enforced by key fingerprint.
WORM retention (S3 Object Lock, Azure immutable blob, GCS Bucket Lock) with legal holds, plus exporters for EU AI Act Annex IV, NIST AI RMF, GDPR RoPA, and CycloneDX AI-BOM.
// standards-based, not reinvented
Only the Run Capsule and Evidence Bundle are new on-disk formats — everything else adopts an existing open standard, so NovaFabric feeds your observability stack rather than replacing it.
Design slogan: local-first now, distributed-ready always, cluster-scale later — a one-node run is the smallest case of a distributed run. Built in Python 3.12+ with a Go collector tier; server mode, collector, object store, and dashboard are marked experimental, and cross-cluster federation is design-intent, not shipped.