Standards and specifications

What NovaFabric implements, how to verify each claim yourself, and — kept deliberately prominent — what is not claimed.

This page exists because "supports EU AI Act" is not a checkable statement. A standards engineer, an auditor, or a reviewer needs to know which artefact, produced by which command, against which version of which specification, and how mature that implementation is. Everything below is stated in those terms.

The distinction this whole page rests on. NovaFabric produces evidence and conformance receipts. It does not issue verdicts, certificates, or attestations of compliance, and no software can. A receipt records that a stated check ran against stated inputs and what it observed. Deciding whether that satisfies an obligation is a judgement made by a qualified human — an auditor, a notified body, a regulator, a DPO.

The codebase enforces this distinction rather than merely promising it; see Assurance cases §4, "Conformance receipts — never verdicts".

Maturity labels are the project's standard ones and are not rounded up: works today (in main, tests pass) · experimental (implemented, interface may change) · planned · future design. See Project status.

1. Evidence, signing, and provenance formats

These are the load-bearing ones: they define the artefact a third party verifies, and they are the part of the system that must keep working offline in five years.

Specification Version / profile What NovaFabric does Verify it yourself Maturity
in-toto Attestation / DSSE DSSE envelope, ed25519 Signs Evidence Bundles and capsule digests as DSSE envelopes nova seal verify <run-id> · nova verify-envelope <file> works today
RFC 3161 Time-Stamp Protocol Anchors each seal with a trusted timestamp; CRL caching so verification works air-gapped nova seal verify <run-id> works today
RFC 8785 JSON Canonicalization Scheme Canonical JSON for digest computation, so a digest is reproducible across implementations inspect proposal_digest in a sealed capsule works today
JSON Schema draft 2020-12 schemas/run-capsule.schema.json and siblings define every artefact the system emits nova validate <run-id> — or validate the schema with any 2020-12 validator works today
Merkle inclusion proofs RFC 6962-style tree Evidence Provenance tree over capsule hashes nova merkle-tree <run-id> works today
SLSA provenance for released artefacts Build provenance and signed release artefacts in the publish pipeline release attestations on the GitHub release works today
CycloneDX 1.7, AI/ML profile AI Bill of Materials generated from capsule facts nova export-aibom <run-id> experimental
C2PA content credentials Provenance manifest for generated media nova export-c2pa <run-id> experimental

Why this section matters most. An Evidence Bundle is designed to be verifiable with sha256sum and an ed25519 verifier — no NovaFabric installation required. If you can only check our evidence with our tool, it is not evidence. That constraint is what forced the choices above toward boring, widely-implemented formats.

2. Telemetry and lineage interoperability

NovaFabric records against open conventions rather than a proprietary shape, so capsule data can leave the system.

Specification Version / profile What NovaFabric does Verify it yourself Maturity
OpenTelemetry traces, GenAI semantic conventions Model calls and spans are recorded against the GenAI semconv; trace.jsonl is OTel-shaped inspect model-calls.jsonl / trace.jsonl in any capsule works today
OpenLineage run/job/dataset events Emits lineage as OpenLineage events for ingestion by Marquez and others nova lineage emit-openlineage works today
W3C PROV PROV-O (RDF) Exports the provenance graph as PROV-O triples nova lineage export-prov experimental
SHACL shape validation Validates provenance triples on ingest into the knowledge graph novafabric[spkg] extra experimental
openCypher query Graph queries over the capsule knowledge graph nova kg query experimental
Model Context Protocol MCP tool exchanges Transparent MCP proxy records tool calls; a conformance suite runs in CI nova mcp conformance · .github/workflows/mcp-conformance.yml experimental

3. Regulatory and management-system frameworks

Read the caveat before the table. Everything in this section produces a document or receipt derived from captured run facts. It is an input to a compliance process, never an output of one. NovaFabric cannot and does not determine whether you comply with any of these instruments.

Framework What NovaFabric produces Command Maturity
EU AI Act Annex IV technical-documentation export; risk-tier classification nova export-annex-iv · nova classify run experimental
NIST AI RMF 1.0 Risk report mapped from capsule facts nova export-nist-rmf experimental
ISO/IEC 42001 AI management-system evidence export; examiner package nova export-compliance iso42001 · nova export-examiner iso42001 experimental
GDPR Art. 30 Records of Processing; Art. 17 erasure via DEK crypto-shredding; DSAR assembly nova export-ropa · nova pii erase · nova dsar assemble experimental
NIS2 Incident report export nova export-nis2 experimental
HIPAA Safe Harbor de-identification proof nova export-hipaa-proof experimental
OMB M-24-10 Risk-tier vocabulary in classification nova classify list-vocabularies experimental

Records-retention (WORM) conformance

A separate, installable behavioural test suite that verifies an object-storage backend meets write-once-read-many requirements. It tests the storage, not your compliance.

Ten mandatory cases covering immutability, retention locks, legal holds, and lifecycle exemptions, referenced against SEC 17a-4, MiFID II, CFTC 1.31 and FINRA 4370. Backends: S3-compatible, Azure Blob, GCS.

pip install "git+https://github.com/MSKazemi/novafabric#subdirectory=packages/nova_worm_conformance"
nova-worm-conformance run --backend s3 --bucket my-capsule-store

Also reachable in-tree as nova storage validate. Maturity: experimental.

4. Identity and access (server mode only)

Specification What NovaFabric does Maturity
OIDC Token verification for the multi-user API experimental
SAML 2.0 SSO assertion verification, XML-DSIG with XML-Signature-Wrapping resistance (novafabric[saml]) experimental
SCIM 2.0 User/group provisioning events experimental

None of these are required for local mode, which has no accounts at all.

What is explicitly not claimed

Stated plainly, because a standards page that only lists strengths is not useful to anyone evaluating the project.

  1. No certification, of anything, by anyone. NovaFabric is not certified, accredited, or endorsed under any framework listed here. Nothing on this page has been assessed by a notified body, an accredited certification body, or a regulator.

  2. Producing an Annex IV export is not EU AI Act compliance. The export is a document assembled from what a run actually did. Whether your system is high-risk, whether the documentation is adequate, and whether your obligations are met are determinations for qualified people. The same applies to every row in §3.

  3. Conformance receipts are observations, not verdicts. A receipt says "this check ran, against these inputs, and observed this". It does not say "you pass". This is a deliberate architectural constraint, not a disclaimer bolted on afterwards.

  4. No byte-exact replay of remote LLM calls. exact mode requires a deterministic environment and a per-call seed — realistic for a local or on-prem model, not for a hosted endpoint that can change under you. For drifting remote models, semantic mode scores similarity of meaning on a 0.0–1.0 scale. Any claim of "deterministic replay" for hosted models should be read carefully, including ours.

  5. Capsule and Evidence Bundle formats are not frozen. They change until the v1.0 freeze — additively, with old capsules staying readable, but they move. If you are building a long-lived integration against the schema, that is the milestone to watch.

  6. Most of §3 is experimental. Implemented and tested, interface may change. Do not build a submission pipeline on an experimental exporter without pinning a version.

  7. Standards coverage is not uniform. Implementing a format (DSSE, CycloneDX) is a much stronger claim than mapping to a framework (ISO 42001). §1 is where the rigour is; §3 is best understood as structured evidence assembly.

For standards bodies and working groups

If you maintain or contribute to any specification above, three things may be useful, and all three are things we would rather hear about than guess at:

1. Implementation feedback. NovaFabric is a real implementation of DSSE + RFC 3161 + in-toto in an AI-workload context, of the OTel GenAI conventions, and of OpenLineage. Where a specification was ambiguous, under-specified for AI workloads, or expensive to implement correctly, that is recorded in the ADR ledger — those are exactly the findings an editor usually has to extract from implementers by hand.

2. Conformance suites you can run. nova-worm-conformance (object-storage WORM) and nova mcp conformance are executable, not prose. If your group wants an independent implementation to test a draft against, these are available under Apache-2.0.

3. Corrections. If any row above misstates your specification — wrong version, wrong clause, an overstated claim, a profile we do not actually meet — that is a bug, and one we want reported. Open an issue with the documentation label, or start a Discussion. A correction from a specification's own maintainers is the most valuable kind of contribution this page can receive, and it will be credited in the release notes.

Contact: open a Discussion or see SUPPORT.md. For anything security-sensitive use the private reporting path in SECURITY.md.

See also