For platform and engineering teams

The page for the person who has to say yes. You have been asked to evaluate NovaFabric for a team, and someone will ask you whether it is safe to depend on. This page answers that in the order those questions actually get asked, links to the evidence rather than asserting, and tells you plainly where the answer is "not yet".

If you want the product story instead, start at the README. If you are evaluating for a paper, For researchers is the better page. If you are evaluating for an audit or compliance function, read Assurance cases — it explains why NovaFabric emits receipts and never verdicts.

The 60-second answer

Question	Answer
License	Apache-2.0. No CLA required to use; a CLA applies to contributions (CLA.md).
Maturity	Beta, v0.101.0. Local capture/replay/diff/lineage are stable and used daily. Server mode, the collector, the dashboard, and the at-scale lineage backends are `experimental`.
Format stability	On-disk capsule and evidence-bundle formats are not frozen until the v1.0 schema freeze. Do not build a long-lived external contract on capsule internals yet.
Runtime dependency on us	None. No accounts, no telemetry, no license server, no phone-home, no update check. If this project disappeared tomorrow, your capsules keep working — they are folders.
Where data lives	Your infrastructure only. Local mode writes to `~/.novafabric/`; server mode writes to your own Postgres.
Network requirement	Core features work fully offline and air-gapped.
Bus factor	One maintainer today. This is the honest headline risk — see below.

Does it fit our stack?

Capture requires no application code changes. nova capture <command> wraps any command. That is the whole integration for most teams.

You run	How NovaFabric attaches
Python agents / scripts	Auto-hooked SDKs: OpenAI, Anthropic, MCP, httpx, requests, aiohttp, urllib3, Bedrock
Non-Python clients	`nova api-proxy` / `nova mcp-proxy` — wire-level, language-agnostic
Frameworks	Adapters for LangGraph, CrewAI, AutoGen, DSPy, Google ADK, OpenAI Agents, Bedrock AgentCore, MLflow, Langfuse, A2A
CI	The GitHub Action — three lines of YAML
Kubernetes	Helm chart and manifests in `deploy/`
HPC / Slurm	`deploy/hpc/`
Air-gapped sites	`make airgap-bundle` — one signed tar, verifiable with zero network

Python 3.12+. That is a real constraint; check it before anything else.

Security and supply chain

Start with SECURITY.md (disclosure process) and Trust surfaces (the signing, attestation, and timestamping layers, and what each one actually proves).

What is in place today:

Secret redaction on the capture path, with a redaction proof written into every capsule — the capsule records that redaction happened, so a reviewer can check it.
Signed container images — keyless cosign signing plus SLSA build provenance attestation and an SBOM attached as OCI artifacts (publish-image.yml).
Pinned deploy images — deploy/IMAGE_PINS.md.
OpenSSF Scorecard runs in CI; the badge on the README is live, not decorative.
Dependency licence policy — default-deny on AGPL, SSPL, BSL, source-available, GPL, and Elastic, enforced by a CI job (license-policy.yml; the policy itself is recorded as a decision, see decisions). If your legal team asks whether a copyleft dependency can appear in a future version: the gate blocks it in CI.
pip-audit on a schedule, with an explicit waiver file under review.

What is not claimed:

No SOC 2, ISO 27001, HIPAA, or FedRAMP certification, and none is planned during v0.x. This is a deliberate decision, not an oversight — see Assurance cases.
Capture is defence-in-depth, not a sandbox. NovaFabric does not claim to contain a hostile workload.

Operating it

Operator guide — deployment, configuration, backups.
SLO catalog — the service levels the server mode is written against.
Support policy — version support windows and what "supported" means for a v0.x project.
Benchmarks — capture overhead, measured, with the command and hardware stated so you can reproduce it rather than trust it.

Failure posture: by design, a NovaFabric component failing must not block your workload. If capture breaks, your job still runs. Verify this yourself — it is the single most important property for anyone putting this in a production path.

The risks we would raise if we were you

We would rather you hear these from us than find them in week three.

Bus factor of one. There is one maintainer. The mitigations are real but partial: Apache-2.0, no hosted dependency, capsules are plain folders readable without this tool, and the path to becoming a maintainer is written down and open (maintainer criteria). If your organization needs a second maintainer to exist before adopting, that is a reasonable position — and funding or staffing one is the most useful thing a company can do for this project.
v0.x formats are not frozen. If you need a stable on-disk contract today, wait for v1.0 or pin a version and plan a migration.
Most subsystems are experimental. The README labels each one. Take those labels literally; they are maintained deliberately and the project has a standing rule against blurring them.
Small community. Few external users means fewer people have hit the bug you are about to hit. The flip side is that your issue gets a real answer quickly.

What adoption gets you from the project

A stated response target — see CONTRIBUTING.md and SUPPORT.md. It is an honest target, not an SLA.
Influence on the roadmap that is disproportionate to your size right now. A team that lists itself in ADOPTERS.md with a concrete use case is, today, the strongest input to ROADMAP.md that exists.
A deprecation window that considers you. The support policy is written against what adopters actually run — which only works if you tell us.

How to run a real evaluation in an afternoon

Do not evaluate this from the README. Run it.

pip install novafabric

# 1. Capture something of your own. No code changes, no keys needed.
nova capture <your command>

# 2. Read the capsule. It is a folder — open it, inspect it, be suspicious of it.
nova validate ~/.novafabric/capsules/<ulid>
ls -R ~/.novafabric/capsules/<ulid>

# 3. Prove the redaction did what it says.
#    Put a fake secret in the environment first, then grep the capsule for it.

# 4. Replay it offline, with the network off.
nova replay <ulid> --mode forensic

# 5. Check the failure posture: break NovaFabric deliberately and confirm
#    your workload still completes.

Step 3 and step 5 are the ones that matter. Everything else is documentation you could have read.

Found a gap? Open an issue — an evaluation that failed is more useful to this project than one that passed, and it will get a direct answer. If it is security-sensitive, follow SECURITY.md instead of opening a public issue.