Compliance and audit export commands
Part of the NovaFabric CLI reference. Both nova and novafabric run the same binary.
Compliance evidence commands (v0.15.0, BQ-005)
These commands require pip install 'novafabric[compliance]' for full functionality.
OQ-01 status (corrected):
nova subject-proof(below) remains LEGAL-HOLD DRAFT MODE — it looks up an existing redaction proof; PII stays in alegal_hold/staging area, not sealed capsules. The DEK-based crypto-shredding strategy this note used to describe as "planned for v0.26.x" has since shipped (ADR-0069, v0.44.0) asnova pii erasebelow, which actually destroys the wrapping DEK for a data subject rather than only staging a proof. Corrected 2026-09-02: this sentence also namednova erasure request, which never destroyed anything — it printed a success message and did nothing, and now fails loudly.
nova export-annex-iv <capsule> --output-dir <dir> --deployment-id <id>
Export EU AI Act Annex IV technical documentation from a Run Capsule (Regulation (EU) 2024/1689 Art. 11). Produces a JSON-LD file covering all 15 mandatory Annex IV elements, with an optional PDF.
nova export-annex-iv .novafabric/runs/01HX.../ \
--output-dir ./compliance/ \
--deployment-id "prod-eu-classifier-v2"
nova export-annex-iv <capsule> \
--output-dir ./compliance/ \
--deployment-id "prod-eu-classifier-v2" \
--pdf # requires weasyprintOptions:
--output-dir, -o PATH(required) — directory to writeannex-iv-<deployment_id>.jsonld--deployment-id TEXT(required) — operator-assigned deployment identifier--pdf / --no-pdf— also render a PDF (requiresweasyprint, included innovafabric[compliance])
Output fields include: system description, intended purpose, training data characteristics, human oversight measures, robustness metrics, and post-market monitoring plan. Fields are marked complete, partial, or missing based on capsule content, and each element carries an evidence_source provenance marker (ADR-0197).
Exit codes: 0 (success), 1 (missing compliance extra or export error).
nova export-nis2 <capsule> --output <file> --incident-id <id>
Export a NIS2 incident report (Directive (EU) 2022/2555 Art. 23) from a Run Capsule.
nova export-nis2 .novafabric/runs/01HX.../ \
--output incident-report.json \
--incident-id "INC-2026-0042" \
--phase 1 # initial report (≤24h, default)
nova export-nis2 <capsule> --output report.json --incident-id INC-001 --phase 2 # ≤72h
nova export-nis2 <capsule> --output report.json --incident-id INC-001 --phase 3 # ≤1 monthOptions:
--output, -o PATH(required) — path to write the NIS2 incident report JSON--incident-id TEXT(required) — operator-assigned incident identifier--phase INT— reporting phase:1(initial, ≤24h),2(detailed, ≤72h),3(final, ≤1 month). Default:1
Fields requiring cross-session lineage (cap-006) are explicitly marked missing with reason requires_cap_006_cross_session_lineage.
Exit codes: 0 (success), 1 (missing compliance extra or export error).
nova export-ropa <capsule> --output <file>
Export a GDPR Art.30 Records of Processing Activities (RoPA) entry from a Run Capsule (cap-007).
nova export-ropa .novafabric/runs/01HX.../ --output ropa-entry.json
nova export-ropa .novafabric/runs/01HX.../ \
--output ropa.json \
--controller-name "Acme Corp" \
--controller-contact "dpo@acme.example"Options:
--output, -o PATH(required) — path to write the RoPA JSON-LD document--controller-name TEXT— GDPR Art.30(1)(a) controller name (operator-declared; marked missing if omitted)--controller-contact TEXT— controller contact / DPO email
Output is a JSON-LD document with gdpr: and nova: namespaces. Fields derivable from
capsule metadata are populated automatically; operator-declared fields are stubs with
missing_fields markers. Use --controller-name / --controller-contact to complete the entry.
Exit codes: 0 (success), 1 (export error).
nova export-nist-rmf <capsule> --output <file>
Export a NIST AI RMF 1.0 quantitative risk assessment report from a Run Capsule (cap-009).
nova export-nist-rmf .novafabric/runs/01HX.../ --output nist-rmf.jsonOptions:
--output, -o PATH(required) — path to write the NIST AI RMF JSON report
Derives scores for all four NIST AI RMF Core functions:
- GOVERN (GV-1.1 policy coverage, GV-1.2 evidence signing)
- MAP (MP-2.1 tool permission observability, MP-2.2 capture level adequacy).
MP-2.1 reads
tool-permission-events.jsonl, which this version does not capture (no NovaFabric producer writes it). When the file is absent the metric is reported withvalue: null,compliant: null,measurement_method: "not_captured"and listed undernot_captured_evidence— it is excluded from the scores and is never counted as a non-compliant metric or asmissing_evidence. A capsule that does carry the file (written by your own code throughCapsuleWriter.append_tool_permission_event) is scored as before. - MEASURE (MS-2.5 eval performance score, MS-2.6 regression detection)
- MANAGE (MG-2.2 PII redaction records — a count of
redaction_manifest.jsonentries, with no threshold or verdict, because no detection denominator is recorded; releases up to v0.107.0 reported a 1.0 "rate", compliant, from a single entry — and MG-4.1 lifecycle management score)
risk_level = low|medium|high|critical based on overall score and non-compliant metric count.
completeness = complete|partial based on whether all expected evidence files are present.
Exit codes: 0 (success), 1 (export error).
nova subject-proof <subject-id>
Look up PII redaction proof for a data subject (GDPR Art.17). HMACs the subject ID with NOVA_PII_PEPPER, queries the RedactionSubjectIndex, and returns a redaction_proof_report.json.
Requires:
NOVA_PII_PEPPERenvironment variable (the same pepper value used during capture). This command will exit 1 if the env var is not set.
NOVA_PII_PEPPER="<secret>" nova subject-proof "user@example.com"
NOVA_PII_PEPPER="<secret>" nova subject-proof "user@example.com" \
--output proof.json \
--key ~/.novafabric/keys/signing_key.pemOptions:
<subject-id>(positional, required) — data-subject identifier (email, phone, GDPR subject ID, etc.)--db PATH— path toredaction_subject_idx.db(default:$NOVAFABRIC_HOME/compliance/redaction_subject_idx.db)--key PATH— ed25519 private key for signing the proof report; if it cannot sign (another key type, a corrupt file) the command exits1and writes nothing — it never falls back to an unsigned proof--output, -o PATH— path to writeredaction_proof_report.json(default: stdout)
Exit codes: 0 (subject found), 1 (env var missing, compliance extra missing, DB error, or --key failed to sign).
Sector and transparency export commands (experimental)
Thirteen read-only renderers that project supplied references and sealed evidence into sector-specific disclosure and attestation artifacts. Shared conventions:
- each command takes one positional JSON document path and supports
--jsonfor a machine-readable artifact; - exit codes are
0(artifact rendered) and2(missing or malformed input) — rendering never mutates a capsule; - outputs are evidence artifacts, not judgments: register/database entries
are explicitly
DRAFT, claims are declared claims, and the scorecard is coverage — the commands never certify, score, or judge compliance.
| Command | Framework / audience | ADR |
|---|---|---|
nova export-compliance |
EU AI Act / ISO 42001 / NIST GenAI + CSA exporters | ADR-0107 |
nova export-model-risk |
SR 26-2 / SR 11-7 model-risk evidence | ADR-0159 |
nova export-model-independence |
Model-validation independence (ADR-0058 maker-checker) | ADR-0159 |
nova export-retention |
SEC 17a-4 / MiFID retention posture + RFC 3161 timestamp | ADR-0159 |
nova export-adverse-action |
ECOA / Reg B specific-reasons evidence (not a notice) | ADR-0159 |
nova export-cat |
CAT-style lifecycle agent-event trail (SEC Rule 613 pattern; not a CAT submission) | ADR-0159 |
nova export-part11 |
21 CFR Part 11 electronic records | ADR-0160 |
nova export-rai-scorecard |
Responsible-AI coverage scorecard | ADR-0158 |
nova export-public-annex-viii |
EU AI Act Annex VIII public DB (DRAFT) | ADR-0169 |
nova export-transparency-register |
Algorithm registers: ATRS / Amsterdam / Helsinki (DRAFT) | ADR-0169 |
nova export-public-disclosure |
Public-sector disclosure record (DRAFT) | ADR-0169 |
nova export-foia |
FOIA / public-records decision export (DRAFT) | ADR-0169 |
nova export-whistleblower |
Source-protecting whistleblower attestation | ADR-0169 |
nova export-citizen-explanation |
Subject-facing decision explanation | ADR-0169 |
nova export-public-incident |
Public-interest incident summary (DRAFT) | ADR-0169 |
nova export-election-disclosure |
Election / civic content-provenance disclosure | ADR-0169 |
nova export-accessibility-claim |
Declared accessibility-conformance claim | ADR-0169 |
nova export-control-attestation |
Governance-control attestation pack | ADR-0170 |
nova insurance |
Liability chain / SLA breach / coverage trigger evidence (experimental) | ADR-0170 |
nova export-model-risk <evidence>
Assemble an SR 26-2 / SR 11-7 model-risk evidence file from per-pillar refs.
nova export-model-risk evidence.json --json<evidence>— JSON:{model_id, development[], independent_validation[], ongoing_monitoring[], model_inventory[], partial[]}
nova export-model-independence --model <id>
Experimental (ADR-0159 D2 / NF-276). Unlike the document-driven renderers
above, this command reads the shipped ADR-0058 maker-checker records itself —
the registry promotion_proposals rows written by nova promote propose /
nova promote approve (opened read-only), plus, per --capsule-id, the NovaSeal
DSSE bundles written by nova seal propose / nova seal approve, whose outcome is
taken verbatim from the shipped verify_sod verifier. It renders whether the
validator (checker) identity differs from the developer (maker) identity.
It asserts only that independence was recorded — never that validation was
sufficient, and never a rating. A single-identity approval is missing with
reason single-identity approval; an open proposal, a bypass, or a record that
failed SoD verification is missing with its reason. So is a registry row written before ADR-0312 (no signoff_scheme): its identities were never checked against the signer trust list, so it cannot evidence independence (releases up to v0.107.0 rendered it complete). A --capsule-id record is
verified against that capsule's content (capsule.yaml, resolved like nova seal verify):
an approval of content changed since, or of a capsule that does not exist here, failed SoD
and is missing — releases up to v0.107.0 exported it as complete independence. The
single-sign-off nova approve table is not read (it records no developer identity).
nova export-model-independence --model credit-scorer@2.1
nova export-model-independence --model credit-scorer --capsule-id <capsule-id> --json--model— asset id,nameorname@version(required)--capsule-id— also read the NovaSeal maker-checker bundle for this capsule (repeatable)--db— registry database (default$NOVAFABRIC_DB_PATHor$NOVAFABRIC_HOME/registry.db)--data-dir,--policy-db— NovaSeal bundle directory and policy DB (defaults asnova seal)--json— machine-readable artifact
Exit codes: 0 (rendered, including missing fields), 2 (unreadable registry or malformed model id).
nova export-retention --bundle <zip>
Experimental (ADR-0159 D5 / NF-277). Attests the retention posture over one
or more Evidence Bundles — the registry retention-policy.yaml window and deletion
mode, the legal-hold state from the shared hold store (nova hold / /api/holds), each run's WORM lock (local adapter
worm.db, reported partial because it is dev/test only, or an S3 / Azure / GCS
WormReceipt supplied via --worm-receipts), the RFC 3161 trusted timestamp as
actually recorded in each bundle by nova export-evidence --timestamp (reported
complete when the TSR is present and bound to the manifest digest; missing
with the reason only when none was obtained), and the hash-chained audit-log
entries for the run. It attests posture, never compliance, and makes nothing
immutable. The TSR's TSA signature is not re-verified here — use
openssl ts -verify as the bundle README describes.
nova export-retention --bundle evidence.zip --registry prod
nova export-retention --bundle a.zip --bundle b.zip --regime mifid --json--bundle— Evidence Bundle ZIP (repeatable, required)--regime—17a-4(default;17 CFR 240.17a-4 (2022 amendment)) ormifid— a version-pinned tag, not a determination--registry— read the registry'sretention-policy.yaml,worm.dband legal holds from the locationsnova retentionandnova holduse:$NOVAFABRIC_HOME/registries/<name>/, then the legacy./.novafabric/registries/<name>/under the working directory (read, with a warning). With several policy copies the window is the first copy's anddeletion_modeisprohibitedif any copy says so (deletion refuses then too); with severalworm.dbfiles the latest lock per run wins--worm-db— read only this WORM DB instead of everyworm.dbof--registry--worm-receipts,--audit-log— supply cloud WORM receipts, override the audit log
Exit codes: 0 (rendered, including missing rows), 2 (unreadable/corrupt bundle, policy, hold ledger, WORM DB or receipts).
nova export-adverse-action --run-id <id>
Experimental (ADR-0159 D6 / NF-278). Reads one sealed Run Capsule (by run id,
resolved in $NOVAFABRIC_CAPSULE_DIR, or by path) and renders the evidence a
creditor's compliance team uses to author an ECOA / Reg B adverse-action notice:
the recorded model call(s) — model ref plus SHA-256 digests of the recorded
request messages and response choices (the prompt and response text are never
copied out); the capsule's inputs/ files, each checked against the capsule's
sealed evidence_digests; the recorded facets.feature_attribution principal
reasons in the order and with the rank recorded — never re-ranked, re-scored or
computed; and whether the capsule carries a NovaSeal envelope (presence only —
re-verify with nova verify). A capsule with no attribution facet reports
principal_reasons as missing with the reason; NovaFabric does not infer
reasons. Today that row is always missing on schema-valid capsules:
feature_attribution is not yet in the closed run-capsule facet registry
(ADR-0196 D2), and registering it is a schema change that needs its own ADR
before any producer can write it. Every recorded string the pack renders — reason
text, the attribution method / producer / model_call_id, and the model-call
refs / status / timestamps — is length-capped (1024 chars) and scanned against the
ADR-0009 secret rules; a match is suppressed (never rendered, listed in
suppressed_fields) and its row marked partial. A reason's rank must be an
integer and its contribution an integer or finite float; any other type (string,
boolean, list) is corrupt evidence (exit 2), never rendered. It is evidence of what was recorded — not an
adverse-action notice, not a credit decision, and not a legal verdict; the
output carries the CFPB line that "the model decided" or a generic reason does not
satisfy the specific-reasons duty. There is no notice text and no verdict field.
nova export-adverse-action --run-id 01KZ...
nova export-adverse-action --run-id 01KZ... --json
nova export-adverse-action --run-id 01KZ... --out aa-pack.json--run-id— run id or capsule directory path (required)--capsule-dir— capsule store to resolve the run id in--json— machine-readable pack on stdout--out— also write the pack JSON to a file (refused inside the capsule)
Exit codes: 0 (rendered, including missing rows), 2 (capsule not found; a
symlinked capsule.yaml / model-calls.jsonl; a file recorded in evidence_digests
that is absent, a symlink, or not a regular file; a sealed digest that does not match
the bytes on disk; malformed manifest, model-call line, or attribution facet; --out
inside the capsule or unwritable).
nova export-cat --run-id <id>
Experimental (ADR-0159 D6 / NF-280). Reads one sealed Run Capsule (by run id,
resolved in $NOVAFABRIC_CAPSULE_DIR, or by path) and renders a lifecycle-ordered,
self-contained agent-event trail — lifecycle stage, recorded actor identity ref and
recorded timestamps per event — as a firm-owned, self-hosted artifact modelled on
the SEC Rule 613 (17 CFR 242.613) Consolidated Audit Trail event lifecycle. The stage
mapping is explicit and nothing else is read:
| Stage | Capsule source | Actor identity ref |
|---|---|---|
origination |
capsule.yaml created_at |
none recorded at run level |
decision |
model-calls.jsonl |
response (else request) model ref |
authorization (optional) |
tool-permission-events.jsonl, human_approvals.jsonl |
authorising_identity (+ approval_principal) / approver_id |
action |
tool-calls.jsonl |
tool_name (+ agent_call_id link) |
disposition |
capsule.yaml finished_at + status |
none recorded at run level |
Events are ordered by UTC-normalised recorded timestamp, ties broken by stage, stream
and line; an event without a timestamp is kept (never given a guessed time), placed
last, and its stage marked partial. A missing required stage makes the trail
partial; authorization is reported but optional (a run whose policy needed no
decision records none). No event is ever synthesised; prompt/response text, tool
arguments/results and approval rationales never reach the trail. Every rendered
recorded string is capped (1024 chars) and scanned against the ADR-0009 secret
rules — a match is suppressed and the stage marked partial. It is not a CAT
submission: no CAT technical-specification format, reporter id or event-type code;
it never connects to the CAT central repository or a plan processor and transmits
nothing.
nova export-cat --run-id 01KZ...
nova export-cat --run-id 01KZ... --json
nova export-cat --run-id 01KZ... --out cat-events.json--run-id— run id or capsule directory path (required)--capsule-dir— capsule store to resolve the run id in--json— machine-readable trail on stdout--out— also write the trail JSON to a file (refused inside the capsule)
Exit codes: 0 (rendered, including missing / partial stages), 2 (capsule not
found; a symlinked capsule file; a stream recorded in evidence_digests that is
absent or not a regular file; a sealed digest that does not match the bytes on
disk; a malformed manifest or event line; an over-long field; an unparseable or
offset-less timestamp; --out inside the capsule or unwritable).
nova export-part11 <document>
Render a 21 CFR Part 11 electronic-records evidence artifact — the
records/signatures elements a run recorded (signer identity, §11.50 signing
intent, DSSE signature binding, record integrity, trusted timestamp, audit
trail), each complete / partial / missing. Facts, never a Part 11
call.
nova export-part11 part11.json --json<document>— JSON:{capsule_root, elements: {name: ref}, partial: {name: reason}}
nova export-rai-scorecard <document>
Render a Responsible-AI coverage scorecard — presence of evidence per
dimension (supported / partial / unsupported / not_applicable),
never a score: no threshold, no fair/unfair or pass/fail label.
nova export-rai-scorecard rai.json --json<document>— JSON:{evidence: {dimension: ...}, not_applicable[], partial[]}
nova export-public-annex-viii <document>
Render a DRAFT EU AI Act Annex VIII / Art. 71 public-database entry from
sealed evidence + operator declarations. Each field is either
capsule_evidence (a digest/ref into the sealed capsule — never the raw
value) or operator_declared (the operator's public declaration).
nova export-public-annex-viii entry.json --json<document>— JSON:{capsule_root, capsule_evidence{}, operator_declared{}}
nova export-transparency-register <document>
Render a DRAFT algorithm-register record from sealed evidence + operator declarations, crosswalked to a public register shape.
nova export-transparency-register reg.json --standard atrs
nova export-transparency-register reg.json --standard amsterdam --json<document>— JSON:{capsule_root, capsule_evidence{}, operator_declared{}}--standard TEXT— register shape:atrs|amsterdam|helsinki(default:atrs)
nova export-public-disclosure <document>
Render a DRAFT public-sector disclosure record from supplied references.
nova export-public-disclosure disclosure.json --json<document>— JSON:{authority_ref, agent_ref, decision_scope, human_oversight_ref, capsule_refs[], system_card_ref?}
nova export-foia <document>
Render a DRAFT FOIA/public-records decision export from a decision's ordered
record index + redaction claims (each redaction a salted digest plus the
claimed statutory exemption_ref).
nova export-foia foia.json --json<document>— JSON:{decision_ref, record_index[], redactions[{digest, exemption_ref}]}
nova export-whistleblower <document>
Render a source-protecting whistleblower attestation over a sealed bundle.
nova export-whistleblower wb.json --json<document>— JSON:{content_digest, authenticity_attestation, anonymity_set_ref?}
nova export-citizen-explanation <document>
Render a plain-language, subject-facing decision explanation.
nova export-citizen-explanation cit.json --json<document>— JSON:{decision_ref, factors[], human_involvement, contest_channel_ref?, logic_summary_ref?}
nova export-public-incident <document>
Render a DRAFT public-interest incident summary from a sealed incident.
nova export-public-incident incident.json --json<document>— JSON:{incident_ref, public_summary?, affected_scope?, remediation_ref?}
nova export-election-disclosure <document>
Render an election/democratic-process content-provenance disclosure — records, never judges. Binds a provenance receipt (e.g. NF-094 / C2PA / SynthID) by digest.
nova export-election-disclosure elec.json --json<document>— JSON:{content_ref, provenance_receipt_ref, disclosure_label, capsule_refs[]}
nova export-accessibility-claim <document>
Render a declared accessibility-conformance claim — a declared claim, never a guarantee.
nova export-accessibility-claim a11y.json
nova export-accessibility-claim a11y.json --standard en_301_549_v4_1_1 --json<document>— JSON:{declared_standard?, audit_digest?, export_format_check?}--standard TEXT— declared standard (wcag_2_2_aa|en_301_549_v4_1_1); overrides the document
nova export-control-attestation <document>
Render a governance-control attestation pack — presents evidence, never certifies.
nova export-control-attestation ctrl.json --json<document>— JSON:{capsule_root, catalog: [{control_id, evidence_kind?}], present_evidence{}, declared[]}
nova insurance
Risk-transfer evidence over facets.risk_transfer (ADR-0170 P2). experimental. Records
evidence — it never assigns fault, decides coverage, adjudicates a claim, or pays out. Every
output (rich or --json) carries the in-mission-boundary line: evidence supports, never
determines, an insurance or legal outcome. Read-only unless --write is passed, which merges the
recorded object into capsule.yaml after validating the whole facet (atomic replace; a
symlinked capsule.yaml or capsule directory is refused). A NovaSeal-sealed capsule (.seal/
present) is refused with exit 2 unless --force-unseal is also passed, which writes anyway
and warns that the existing seal no longer verifies and must be re-issued. The other sub-commands
ADR-0170 names (features, loss bind, subrogation, insurability) are planned — not yet
implemented.
nova insurance liability show --capsule <capsule>
NF-383 liability-attribution chain: ordered {role, party_ref, acted_as_ref, basis_ref, contribution_marker} — records who was attributed, never who is at fault. role ∈ principal |
deployer | operator | vendor | model_provider | integrator | agent | third_party;
contribution_marker ∈ recorded | disputed | none (recorded/disputed must name a
basis_ref). References are sha256: digests; acted_as_ref must name another chain member (no
dangling, self or cyclic edges); fault/verdict/payout-shaped keys are rejected.
nova insurance liability show --capsule 01HX... # read the recorded chain
nova insurance liability show --capsule 01HX... --chain chain.json --writeOptions: --chain PATH (declared chain JSON — a list or {liability_chain: [...]}), --write
(requires --chain), --json.
nova insurance sla verify --capsule <capsule> --sla <terms.json> --observed <decimal>
NF-384 SLA/warranty breach: compares the observed value to the declared commitment
{metric, operator (lt|lte|gt|gte|eq), threshold, window, unit?} and records breach (the
commitment did not hold) with both figures as exact decimals. sla_ref is the sha256 of the terms
file. Numeric condition only — no remedy, service credit or damages; distinct from ADR-0146 cost
showback.
nova insurance sla verify --capsule 01HX... --sla sla.json --observed 99.2 --evidence-ref sha256:<hex>Options: --evidence-ref (repeatable sha256: digest), --write, --json.
nova insurance coverage check --capsule <capsule> --exclusions <set.json> --facts <facts.json> --event-kind <kind>
NF-386 exclusion-aware coverage trigger: records covered_event_kind, trigger_facts (digests),
exclusion_set_ref (sha256 of the declared set, e.g. ISO CG 40 47), matched_exclusions (declared
exclusions whose declared fact markers are present among the observed facts — possibly empty), and
optional parametric_conditions (observed value vs declared parametric threshold; unobserved terms
are recorded as not observed). It never decides whether the policy responds.
nova insurance coverage check --capsule 01HX... --exclusions cg4047.json --facts facts.json \
--event-kind erroneous_output --json--exclusions: {exclusions: [{exclusion_id, fact_markers[]}], parametric_triggers?: [{metric, operator, threshold, unit?}]}. --facts: {trigger_facts: [{fact_ref, marker}], observations?: {metric: value}}.
Exit codes (all three): 0 recorded — a breach or a matched exclusion is a recorded fact, not a
failure; 2 input error — unknown capsule, unreadable/oversize (1 MiB cap)/malformed input, a non-finite
figure (NaN/inf), a non-digest reference, or a determination-shaped
field. JSON numbers in the input files are parsed as exact decimals.
Compliance commands — full reference
This section consolidates all compliance-related CLI surfaces. Commands are labeled per the docs honesty rule: works today, planned (implementation in progress or scheduled for v0.26.x), or future (ADR accepted, implementation not yet scheduled).
Already implemented — works today
| Command | Regulation | Version |
|---|---|---|
NovaSeal signing intent (DSSE intent field; Python API, no CLI flag — see below) |
FDA 21 CFR Part 11 §11.50 | v0.12.15+ |
nova export-annex-iv <capsule_id> |
EU AI Act Annex IV (Reg. 2024/1689 Art. 11) | v0.15.0 |
nova export-nis2 <capsule_id> |
NIS2 Directive Art. 23 (Phase 1/2/3) | v0.15.0 |
nova subject-proof <subject_id> |
GDPR Art.17 HMAC subject lookup | v0.15.0 |
nova audit coverage |
Multi-profile control coverage scoring | v0.16.0 |
nova audit bundle |
ZIP export of full audit report | v0.16.0 |
nova audit verify |
AuditReport schema check (the report is unsigned) | v0.16.0 |
nova classify run |
EU AI Act / NIST RMF / OMB M-24-10 risk tier | v0.16.0 |
nova verify <capsule> |
DSSE + RFC 3161 TSR + Merkle log verification | v0.10.0+ |
nova export-ropa <capsule_id> |
GDPR Art.30 Records of Processing Activities (cap-007) | v0.25.1 |
nova export-aibom <capsule_dir> |
CycloneDX 1.7 AI-SBOM / ML-BOM (cap-008); default output aibom.json |
v0.25.1 |
nova aibom status |
CRA SBOM compliance status: deadline 2026-09-11, format, per-capsule coverage | v0.35.0 |
nova aibom generate [--all] [--force] |
Per-deployment automation: generate/refresh aibom.json for one or all capsules |
v0.45.0 |
nova export-nist-rmf <capsule_id> |
NIST AI RMF 1.0 quantitative risk report (cap-009) | v0.25.1 |
nova assure <capsule_id> |
OWASP LLM Top 10 (2025) evidence checks — exit 1 on failure (E-10) | v0.25.1 |
nova mcp scan <manifest> |
OWASP LLM supply-chain risk scanner for MCP manifests (E-9) | v0.25.1 |
Signing intent (FDA 21 CFR Part 11 §11.50)
Every NovaSeal DSSE envelope can carry a declared signing intent per FDA 21 CFR
Part 11 §11.50. The intent is set through the Python API —
NovaSeal.seal(manifest, intent=SigningIntent.REVIEWED) (default: authored;
intent=None omits it); there is no --intent CLI flag. The CLI maker-checker
path is nova seal propose → nova seal approve:
nova seal propose <capsule-id> \
--key signer.pem \
--cert signer_cert.pem \
--justification "FDA §11.50 review approved — all eval gates green"Signing intents are expressed via the SigningIntent enum (src/novafabric/trust/novaseal/envelope.py):
| Intent | Meaning (FDA §11.50 mapping) |
|---|---|
approved |
Approval signature (§11.50(a)(1)) |
reviewed |
Review signature (§11.50(a)(2)) |
authored |
Authorship signature (§11.50(a)(3)) |
witnessed |
Witnessed signature |
verified |
Verification / QA signature |
Since ADR-0309 the intent is signed: it is written into the sealed payload under the
reserved _novaseal.signing_intent key (so the DSSE signature covers it) and repeated
as the signature entry's intent label. nova verify reports the intent only when the
signed copy is present and agrees with the label (signing_intent_status: signed). A seal
made before ADR-0309 carries only the unsigned label and reports label-only with no
intent; a label that contradicts the signed value means it was edited after signing, and
verification fails.
The five-check SoD verifier (exit codes 3–7) enforces the separation-of-duties
required by FDA §11.50(b).
nova seal sign --backend sigstore
works today — Keyless Sigstore signing (ADR-0071, v0.44.0). Produces a Sigstore
Bundle v0.3 using an ephemeral OIDC-bound certificate and Rekor v2 transparency log
inclusion. Requires pip install novafabric[sigstore].
nova seal sign manifest.json --backend sigstore --capsule-id 01HX...
nova seal sign manifest.json --backend sigstore --capsule-id 01HX... --home /data/novaOptions:
--backend [local|sigstore]— signing backend (default:local;sigstorerequiresnovafabric[sigstore]).localis not implemented innova seal sign: it exits2and signs nothing — seal withnova capture(whennovaseal.yamlexists) ornova seal propose. Releases up to v0.107.0 exited0there, so anova seal sign … && publishscript published an unsigned manifest.<capsule_manifest>(required) — path to a capsule manifest JSON file, or an inline JSON string--capsule-id TEXT— capsule ID for bundle storage (default: SHA-256 of the manifest)--home PATH—NOVAFABRIC_HOMEoverride for bundle storage path
Bundles stored at $NOVAFABRIC_HOME/sigstore/<capsule_id>.bundle.json.
Verify with nova verify --backend sigstore --capsule-id <id>.
Note: The sigstore SDK performs OIDC browser-based or ambient identity
(GitHub Actions, GCP, AWS) credential lookup at sign time. Not suitable for
air-gapped HPC environments (use --backend local with RFC 3161 instead).
nova pii erase <subject_id>
experimental (ADR-0069, ADR-0314) — Destroys the AES-256-GCM DEK for a data subject.
PII captured with nova capture --pii-subject <id> is encrypted under
that DEK, so after erasure it is permanently unreadable and readers render it as
[erased]. PII in capsules captured without --pii-subject is not encrypted and is
not affected. Writes an ErasureReceipt (immediate) or
ErasureDeferredReceipt (Art.17(3)(b) retention window still active).
Legal holds win (GDPR Art.17(3)(e)). While any legal hold is active
(nova hold create or POST /api/holds, in any registry; a damaged hold
ledger counts as a hold), the DEK is not destroyed: the command writes an
ErasureHeldReceipt (outcome: held, dek_retained: true, the active holds, and each
associated capsule under held_capsules with its hold ids) and exits 3. One DEK protects
all of a subject's capsules and holds are registry-global today, so nothing is erased while a
hold is active — repeat the request after release. There is no override flag. Every outcome
(completed, deferred, held) is appended to the audit log as an erasure.request event carrying
the subject's SHA-256 only.
nova pii erase "user@example.com"
nova pii erase "user@example.com" --output receipt.json
nova pii erase "user@example.com" --retention-months 6Options:
--output, -o PATH— path to write theErasureReceiptJSON (default: stdout; status lines go to stderr, sonova pii erase X > receipt.jsonis a valid receipt)--capsule-dir PATH— search for redaction manifests here (default: wherenova capturewrites —$NOVAFABRIC_CAPSULE_DIR, else$NOVAFABRIC_HOME/capsules/; before 2026-10-23 the env var was ignored, so an erasure in the Docker layout scanned an empty directory)--retention-months INT— Art.17(3)(b) retention floor in months (default:6; provider mode:120)
Exit codes: 0 (erased or deferred; receipt written), 1 (subject not found or error),
3 (held by a legal hold — nothing erased; ErasureHeldReceipt written).
capsule_ids_affected (and a held receipt's held_capsules) lists the capsules whose redaction
manifest references the subject, matched by the subject's HMAC under NOVA_PII_PEPPER. Without the
pepper the subject's capsules cannot be identified and the list is empty (a warning says so); the
erasure itself is unaffected. Before this release every capsule with a redaction manifest was listed,
other subjects' included.
Scope: single data subject. Requires DEK store at $NOVAFABRIC_HOME/dek.db.
nova pii status <capsule_id>
works today — Show PII encryption status for a capsule — which fields are encrypted,
which subjects have active DEKs, and which have been erased (crypto-shredded, ADR-0069).
Read-only and local-first: correlates the capsule's redaction_manifest.json against
the DEK store at $NOVAFABRIC_HOME/dek.db without creating or modifying either.
Subjects appear only as HMACs — no key material or plaintext PII is ever shown.
nova pii status 01HXAMPLECAPSULEID # resolve by capsule ID
nova pii status .novafabric/runs/01HX.../ # or by capsule directory path
nova pii status 01HXAMPLECAPSULEID --json # machine-readable PIIStatusReport
NOVA_PII_PEPPER=secret nova pii status 01HX... # pepper enables DEK correlationOptions:
--capsule-dir PATH— directory scanned for the capsule'sredaction_manifest.jsonwhen a bare capsule ID is given (default: wherenova capturewrites —$NOVAFABRIC_CAPSULE_DIR, else$NOVAFABRIC_HOME/capsules/; before 2026-10-23 the env var was ignored)--json— emit the status report as JSON instead of human-readable text
Per-subject dek_state:
active— a live DEK exists; ciphertext is still recoverable by the key holdererased— no DEK exists (crypto-shredded vianova pii erase, ordek.dbabsent)unknown—dek.dbis present butNOVA_PII_PEPPERis not set, so manifest HMACs cannot be correlated; ordek.dbholds live DEKs and none of them matches any subject of this capsule — the pepper may be wrong (rotated, mistyped, another deployment's), so absence of a match is not taken as erasure. Releases up to v0.107.0 reportederasedhere, an erasure confirmation for data that was still decryptable.
Exit codes: 0 (report produced — including capsules with no redaction manifest), 1 (capsule not found or manifest unreadable).
Scope: single capsule. Works without a DEK store and without NOVA_PII_PEPPER (degrades honestly as above).
nova export-rocrate <capsule_dir>
Export a Run Capsule as a FAIR-compliant RO-Crate v1.1 research object (ZIP archive).
works today — implemented in v0.32.0.
nova export-rocrate .novafabric/runs/01HX.../
nova export-rocrate .novafabric/runs/01HX.../ --output ./out.rocrate.zipOptions:
--output, -o PATH— output ZIP file path (default:<capsule_dir>.rocrate.zipnext to the capsule dir)
A capsule holding a symbolic link anywhere is refused (exit 1, no archive written); the export never follows a link. See Concepts.
nova export-rocrate-science --capsule <capsule>
Export a science capsule as a FAIR Workflow-Run-RO-Crate science profile (ADR-0164 NF-324).
experimental — ADR-0164 P2. Composes over the nova export-rocrate carrier (NF-040): the
carrier's RO-Crate 1.1 file set is kept, and the Workflow Run Crate 0.5 entities are added — a
W3C-PROV-aligned CreateAction, the reproducibility-receipt digests and seeds as
PropertyValues, the hypothesis→claim DAG (with isBasedOn parent edges), and the sealed
science root as an identifier. Byte-deterministic when the capsule carries a timestamp.
nova export-rocrate-science --capsule .novafabric/runs/01HX.../ --out ./crates
nova export-rocrate-science --capsule 01HX... --orcid 0000-0002-1825-0097 --ror 03yrm5c26 --jsonWrites <run_id>.science.rocrate.zip and <run_id>.fair-binding.json into --out (a
run_id that is not one plain file name, or a link at either output name, is refused with exit
1 and nothing written; each file is written to a new file and renamed into place) (the NF-324
fair_binding record: profile URIs, rocrate_digest, prov_alignment: "w3c-prov",
sealed_root, receipt_root, persistent identifiers, unbound, verdict: null).
Options:
--capsule TEXT— capsule directory or run id (required)--out PATH— output directory (default: the capsule's parent directory)--profile process-run-crate|workflow-run-crate— default:workflow-run-cratewhen the receipt declares aworkflow_digest, elseprocess-run-crate.provenance-run-crateis planned.--doi TEXT,--orcid TEXT(repeatable),--ror TEXT(repeatable) — persistent-identifier references (ORCIDs becomecontributor, neverauthor)--json— print thefair_bindingrecord
Exit codes: 0 exported; 1 refused — no science_provenance facet, a tampered receipt, a
workflow profile with no declared workflow, or a malformed DAG; 2 usage error.
Every output carries the in-mission-boundary line: NovaFabric records and exports provenance;
it never runs experiments or adjudicates scientific validity.
nova science receipt build|verify --capsule <capsule>
Computational-reproducibility receipt (ADR-0164 NF-323). experimental — ADR-0164 P2.
build binds the environment digest, seed(s), input-data digest, code digest and optional
workflow digest — plus the declared determinism_class and, when the science facet records
one, the sealed capsule root — under one bound_root. Components not supplied are named in
receipt_incomplete, never fabricated. verify recomputes the root offline and checks the
declared incompleteness. Record-only: nothing is re-executed and reproducible_in_fact is
always null.
nova science receipt build --capsule 01HX... --env sha256:<hex> --data sha256:<hex> \
--code sha256:<hex> --seed 1337 --seed 42 --determinism statistical --write
nova science receipt verify --capsule 01HX... [--strict] [--json]build options: --env, --data, --code, --workflow (each sha256:<64 hex>), --seed INT
(repeatable, ordered), --determinism bitwise|statistical|nondeterministic|undeclared,
--write (persist into capsule.yaml under facets.science_provenance.reproducibility_receipt;
atomic replace, symlinked capsule.yaml refused), --force-unseal (with --write: rewrite a
NovaSeal-sealed capsule — refused with exit 2 otherwise — and warn that its seal no longer
verifies and must be re-issued), --json.
Exit codes: 0 ok; 1 (verify) root mismatch, misreported incompleteness, malformed or
absent receipt — or --strict with an incomplete receipt; 2 usage error (unknown capsule,
malformed digest or seed).
nova science lab show|verify --capsule <capsule>
Declared lab-experiment provenance (ADR-0164 NF-322). experimental — ADR-0164 P3.
nova science lab show prints facets.science_provenance.lab_experiment — lab_kind
(self_driving | cloud_lab | manual | simulation), protocol_ref, instrument_refs, run_id,
sim_to_real (sim | real | hybrid), outcome_digest, optional started_at / node_ref — and
how many instrument records the capsule carries. nova science lab verify checks offline, and
fails closed, that: every instrument_ref resolves to an instrument record; no referenced
instrument was calibrated after the experiment (experiment time is started_at, else the
capsule's created_at; unknown fails); the experiment and instrument digests re-derive; a
simulation is not declared real; and an optional node_ref names an experiment_design /
experiment_run node of the NF-321 lineage. Declarations only — NovaFabric never dispatches to
or controls a lab and never reads instrument telemetry (controls_lab: false,
reads_telemetry: false, verdict: null).
nova science lab show --capsule 01HX... [--json]
nova science lab verify --capsule 01HX... [--json]The blocks are written by the library API (novafabric.science.lab.build_lab_experiment,
build_instrument_record, attach_lab); there is no build subcommand.
Exit codes: 0 ok; 1 (verify) a failed check or an absent block — and for every
subcommand a malformed block (unknown lab_kind, malformed digest, a telemetry / raw-data /
credential-shaped field); 2 usage error (unknown capsule).
nova science instrument show --capsule <capsule>
Declared instrument / calibration provenance (ADR-0164 NF-329). experimental — ADR-0164 P3.
Prints each facets.science_provenance.instrument_provenance[] record — instrument_id,
instrument_class, firmware_digest, calibration_ref (digest of the calibration record,
never the record), calibration_timestamp, manufacturer_ref, and the record_digest a lab
experiment's instrument_refs point at. Add --json for machine output. Exit 0 shown (or
none recorded); 1 malformed block; 2 unknown capsule.
nova lineage export-prov <capsule_dir>
Export a W3C PROV provenance graph from a capsule's lineage.jsonl, as PROV-JSON
(default) or PROV-N text — both serializations of the same graph.
works today — PROV-JSON since v0.32.0; PROV-N added (ADR-0176).
nova lineage export-prov .novafabric/runs/01HX.../
nova lineage export-prov .novafabric/runs/01HX.../ --output prov.json
nova lineage export-prov .novafabric/runs/01HX.../ --format prov-n -o prov.provnOptions:
--format, -f {prov-json,prov-n}— serialization (defaultprov-json)--output, -o PATH— output path (default:<capsule_dir>/prov.jsonorprov.provn)
nova export-hipaa-proof <capsule_dir>
works today — Export a HIPAA Safe Harbor (§164.514(b)) de-identification
proof artifact for a capsule, documenting which of the 18 identifier categories
were redacted (a redaction manifest records it) and which are not assessable.
The exporter reads the capsule's redaction-evidence files, never its captured content, so
it never states that an identifier is absent: not_present (and present_unredacted) are
reserved for a scan of the content (planned). Releases up to v0.107.0 reported every
category without redaction evidence as not_present — including for a capsule whose
prompts held a name, an SSN and an email.
DISCLAIMER: This is a technical evidence artifact only. It is NOT legal advice, NOT a compliance certification, and does NOT guarantee Safe Harbor legal sufficiency.
nova export-hipaa-proof .novafabric/runs/01HX.../ # writes hipaa-proof.json in capsule dir
nova export-hipaa-proof .novafabric/runs/01HX.../ --output hipaa-proof.jsonOptions:
--output, -o PATH— path to write the proof JSON (default:<capsule_dir>/hipaa-proof.json)
nova export-aibom <capsule_dir>
Export an AI Bill of Materials (AIBOM) using CycloneDX ML-BOM 1.7 (ECMA-424 2nd Edition). Required for EU Cyber Resilience Act (CRA) compliance. Deadline: 2026-09-11 (ADR-0073).
works today — implemented in v0.25.1; --output default + nova aibom status added v0.35.0; nova aibom generate (per-deployment automation) added 2026-06-04.
nova export-aibom .novafabric/runs/01HX.../ # writes aibom.json in the capsule dir
nova export-aibom .novafabric/runs/01HX.../ --output /tmp/aibom.cdx.json
nova aibom status # show deadline + per-capsule coverage
nova aibom status --capsules-dir ~/novafabric-data/capsules
nova aibom generate .novafabric/runs/01HX.../ # generate one (skips if present)
nova aibom generate --all # batch: every capsule missing an aibom.json
nova aibom generate --all --force # batch: refresh all (overwrite)
# NF-056 CycloneDX 1.7 extensions (opt-in; default output unchanged):
nova aibom generate <cap> --citations --force # bind components to capsule/evidence digests
nova aibom generate <cap> --tlp TLP:AMBER --force # distribution marker in metadata.properties
nova aibom generate <cap> --model-card auto --force # model-card externalReference per model
nova aibom generate <cap> --no-include-datasets --force # suppress type:data components
nova aibom validate <cap>/aibom.json # structural + binding check (exit 1 on error)Options (nova export-aibom):
--output, -o PATH— path to write the AIBOM JSON (default:<capsule_dir>/aibom.json)
Options (nova aibom status):
--capsules-dir PATH— root directory to scan for capsules (default:$NOVAFABRIC_CAPSULE_DIRor$NOVAFABRIC_HOME/capsules)
Options (nova aibom generate) — per-deployment automation (CRA):
<capsule_dir>— optional positional: generate for a single capsule--all— batch-generate across--capsules-dir, skipping capsules that already have anaibom.json--capsules-dir PATH— capsule store for--all(default:$NOVAFABRIC_CAPSULE_DIRor$NOVAFABRIC_HOME/capsules)--force— regenerate even when anaibom.jsonalready exists--citations— (NF-056, ADR-0105) bind each model/dataset component to its source capsule/evidence digest (+ ADR-0097 inclusion proof when the manifest records one). Default off (byte-stable output).--tlp TLP:CLEAR|GREEN|AMBER|AMBER+STRICT|RED— TLP 2.0 distribution marker inmetadata.properties(novafabric:tlp). Default: none.--model-card auto|<path>— add amodel-cardexternalReference to each model component (autoderives aregistry://URI). Default: none.--include-datasets / --no-include-datasets— emittype:datadataset components from lineage (default on).
Options (nova aibom validate <bom.json>) — NF-056: structural CycloneDX 1.7 + NovaFabric-binding check (specVersion/bomFormat/serialNumber, TLP marker validity, per-component name/bom-ref/hash-alg/citation digest). Exit 0 valid, 1 on validation errors, 2 on a missing/malformed file. --json emits {valid, errors}.
nova dataset provenance-card <asset> (experimental)
Experimental — NF-058, ADR-0105. Emit (and optionally sign) a dataset provenance card recording a dataset's source, version, content hash, and transform history. Feeds the AI-BOM (NF-056) and the SLSA-for-ML attestation (NF-057). Op digests only — never raw values, prompts, or cell contents.
nova dataset provenance-card dataset:gaia@2026-05 \
--source oci://reg/gaia:2026-05 --version 2026-05 --hash sha256:<64-hex-digest> \
--license CC-BY-4.0 --tlp TLP:CLEAR --registry-digest sha256:<64-hex-digest> --sign
# pull the transform history from a capsule's lineage derivation edges:
nova dataset provenance-card dataset:x@1 --source oci://x --version 1 --hash <sha256> \
--from-capsule ./my-capsule --sign --out card.json--source,--version,--hash(required) — the dataset's source URI, version label, and content SHA-256.--hash(and--registry-digest) must be the full digest — 64 hex digits, bare or prefixedsha256:(case-insensitive; stored lowercase). Anything else — a truncatedb17a, another algorithm — exits 2 and nothing is built or signed. (Before the L32 fix, 2026-10-23, a truncated hash was signed into the card.)--from-capsule <dir>— derivetransformHistory[]from the capsule'slineage.jsonlderivation edges (eachsignedOpDigestis the content hash of the recorded edge).--license,--tlp,--registry-digest,--retrieved-at— optional card fields.--sign [--identity <id>]— Ed25519-sign the card with the keyring key (reuses the NovaSeal/Evidence-Bundle path). The signature is taken over the canonical-JSON body excluding the signature block. An unsigned card is schema-invalid — a signed card is evidence.--out, -o <path>— write the card JSON (default: stdout).
nova export-c2pa <capsule_dir>
Export a C2PA v2.3-compatible provenance manifest for AI-generated content (ADR-0074). Required by EU AI Act Art.50. Deadline: 2026-08-02.
works today — implemented in v0.33.0.
nova export-c2pa .novafabric/runs/01HX.../
nova export-c2pa .novafabric/runs/01HX.../ --output manifest.json
nova export-c2pa .novafabric/runs/01HX.../ --training-mining # adds notAllowed assertionOptions:
--output, -o PATH— output JSON file (default:<capsule_dir>/c2pa-manifest.json)--training-mining— includec2pa.training-mining: notAllowedassertion (opt-in)
The output is never written through a link (works today). Inside the capsule (the
default, or any --output under the capsule directory) the file is replaced by an atomic
rename after an O_NOFOLLOW walk, so a symlink or hard link planted at
c2pa-manifest.json is replaced and its target is untouched; a linked directory on the
way (say outputs -> /elsewhere) is refused. An explicit --output outside the capsule
is written in place, but a symlink, FIFO or device there is refused. A refusal prints one
x line and exits 1; nothing is written.
Note: the manifest is structurally valid for TSP signing. Hard-binding C2PA verification requires operator enrollment with a C2PA-certified Trust Service Provider (TSP signing path deferred to a future release).
nova export-compliance <subcommand>
Export EU AI Act / ISO / NIST compliance evidence from the ADR-0107 exporters. Each
subcommand reads a capsule or a JSON input and writes a report as JSON (to --out, else
stdout). The CLI only parses and serialises — the exporter logic lives in
novafabric.compliance.export.
works today — implemented in v0.89.0 (experimental).
# NIST GenAI + CSA Agentic profile, built from a capsule's NIST-RMF report (NF-097)
nova export-compliance genai-profile .novafabric/runs/01HX.../ --evidence tool_permissions,eval_gate --out profile.json
# ISO/IEC 42001 control-evidence mapping from a declared catalog (NF-095)
nova export-compliance iso42001 --catalog controls.json --evidence eval_gate --capsule-id run-123 --out iso.json
# First sealed revision of a GPAI Art. 53 documentation form (NF-093)
nova export-compliance gpai53 --model my-gpai --fields art53.json --out form.json
# Art. 72 post-market-monitoring report; serious findings refer Art. 73 incidents (NF-091)
nova export-compliance pmm --system triage --findings findings.json --occurred-at 2026-07-01 --out pmm.jsonSubcommands and key options:
genai-profile <capsule>—--evidence a,b(governance-evidence kinds present),--out. A focus area is evidenced from the base report only by a positive metric read from a recorded artifact (audit coverage, eval results, a regression report, a redaction manifest); default-derived scores and mere file presence do not count. Releases up to v0.107.0 marked governance, content provenance and incident disclosure evidenced for a capsule holding onlyrun_id:.iso42001 --catalog c.json—--evidence a,b,--capsule-id ID(supplies the re-performable reference evidenced controls require; without it, evidenced controls honestly degrade tonot_evidenced),--out.gpai53 --model NAME --fields f.json—--out.fields.jsonis a{field: value}object. Each revision'scontent_digest(digest_scheme: v2) covers the whole revision — fields,created_at,retention_until, revision number, predecessor link — and the model name, so a silent edit to any of them fails verification. Revisions sealed beforev2(nodigest_scheme) coveredfieldsonly, and still verify only those.pmm --system NAME --findings f.json --occurred-at ISO—--period-start,--period-end,--out. A serious finding (severitycritical/high) must carry anincident_classification, or the command fails closed.
Note: NF-090 (Art. 12) is served by nova euaiact; NF-092 (Annex IV) by the AIBOM/Annex-IV
exporters; NF-094 (Art. 50 dual-layer receipt) composes with nova export-c2pa.
nova export-system-card <capsule_dir>
Generate and seal an auto-generated system/audit card (ADR-0085). The card is
assembled from capsule + eval + lineage facts (capsule.yaml, eval_result.json,
lineage.jsonl) and sealed by reusing the existing DSSE/Ed25519 signing path — no
new crypto. It is generated, never hand-written, so it cannot drift from the
capsule it describes. The sealed output is a DSSE envelope (predicate type
https://novafabric.io/system-card/v0) that verifies with the same verifier used
for Evidence Bundles.
works today — implemented in v0.48.0.
nova export-system-card .novafabric/runs/01HX.../
nova export-system-card .novafabric/runs/01HX.../ -o card.intoto.json --key ed25519.pemOptions:
--output, -o PATH— output DSSE JSON (default:<capsule_dir>/system-card.intoto.json)--key PATH— PEM-encoded ed25519 private key to seal with. If omitted, a local key under$NOVAFABRIC_HOME/keys/is created/loaded.
Related: eval results now also pin the asset version and an optional
dataset version (nova eval-driven run_evals, ADR-0085) so a stored eval
result can be tied back to exactly what was measured.
nova export-blob --dest <uri>
Export a batch of capsules to blob storage with one signed, verifiable
completeness manifest (ADR-0141,
spec batch-blob-export-v0). Members are
selected explicitly (--capsule, repeatable) or by scanning the capsule directory
with an optional created_at time range (--since/--until, inclusive); the
selection is frozen at export start (a batch is a snapshot). Each member is packed
deterministically and written content-addressed under objects/<sha256> at the
destination — already-present blobs are skipped, so re-runs are idempotent and an
interrupted export resumes. The Ed25519/DSSE-signed export-manifest.json
(validates against schemas/export-manifest.schema.json) is written last as the
atomic completion marker: it lists every member's capsule_id, content_hash, and
size, plus a batch_digest over the sorted member list, so a recipient can verify
integrity and completeness offline. Source capsules are never modified.
A capsule holding a symbolic link is refused (exit 2, "Export refused", nothing written):
packing used to follow links and store their targets. See Concepts.
A member whose redaction-proof.json records unsafe_skips refuses the whole batch, as
nova export-evidence does — and so does a member with no usable proof: none at all
(redaction never ran), one that is not valid JSON (a torn write), or one that is not a JSON
object. --allow-unsafe-skips overrides both. Releases up to v0.107.0 read a missing or
truncated proof as "0 unsafe skips" and exported the capsule.
experimental — implemented; API may change.
# Local directory (always works, fully offline)
nova export-blob --dest ./exports/audit-q3 -c 01HX... -c 01HY...
# Time-range scan of the capsule store
nova export-blob --dest ./exports/w27 --since 2026-06-29 --until 2026-07-05
# Query filter (ADR-0129 DSL) — composes with the time window
nova export-blob --dest ./exports/gpt4o --query "model = 'gpt-4o'"
nova export-blob --dest ./exports/failures \
--query "status = 'error'" --since 2026-07-01
# Any S3-compatible endpoint (existing optional boto3 adapter; private endpoints OK)
NOVA_S3_ENDPOINT_URL=https://s3.internal nova export-blob \
--dest s3://audit-bucket/exports/2026-07/ --worm --worm-retention-days 2555Options:
--dest URI— local directory ors3://bucket[/prefix](any S3-compatible endpoint viaNOVA_S3_ENDPOINT_URL).azure:///gcs://are planned (ADR-0141 P2) and report a clear error.--capsule, -c ID|PATH— explicit member (repeatable); recordsquery: null.--capsule-dir PATH— scan root (default: wherenova capturewrites —$NOVAFABRIC_CAPSULE_DIR, else$NOVAFABRIC_HOME/capsules).--since / --until RFC3339— inclusivecreated_atwindow for the scan; recorded as the manifestqueryfor reproducibility.--query EXPR— select members with an ADR-0129wherefilter, e.g."model = 'gpt-4o'"or"status = 'error'". Composes with--since/--until(both must hold), and is mutually exclusive with--capsule. The parsed, canonical predicates are recorded in the manifestqueryfield alongside the time window — so a manifest reader sees what was asked, not free text they must re-parse. Filters on the query'swhereclause only:run_idis not a DSL dimension, and grouping by it would invite the high-cardinality blow-up the DSL's cardinality cap exists to prevent.--key PATH— PEM ed25519 signing key (default: the NovaFabric keyring key, created if absent).--worm/--worm-retention-days N— write members and manifest under WORM retention (S3 Object Lock COMPLIANCE via the existing ADR-0031 adapter; local dirs get best-effort read-only files); intent recorded in the manifest.--out PATH— also save a local copy of the manifest.--public-key-out PATH— save the signer's public key PEM for recipients.
Verify offline (no NovaFabric service; extends nova verify):
nova verify export-manifest.json --public-key export.pub.pem
nova verify manifest.json --public-key pub.pem --dest s3://audit-bucket/exports/2026-07/Verdicts: VALID (signature + batch digest + every member's bytes at the
destination check out), INCOMPLETE (a member missing, or size/hash mismatch —
e.g. a deleted or tampered blob), INVALID (bad signature, count, or batch
digest — e.g. a member quietly dropped from the manifest). Exit 0 only on VALID.
nova import <source>
Import a batch export (nova export-blob)
into the local capsule store — the verified inverse of the export
(ADR-0207, spec
batch-import-v0). <source> is either a
directory containing export-manifest.json + objects/, or a single
.tar/.tar.gz archive of that layout (the air-gap courier format). An archive is
streamed, never read whole into memory, and bounded with the server-ingest defaults: at
most 10 000 entries, 2 GiB uncompressed, and a 100:1 compression ratio (past 1 MiB) — a
larger one fails to extract (INVALID) before its bytes are written. Releases up to
v0.107.0 had no bounds, so a small compression bomb filled the disk or killed the CLI.
Verification-first, fail-closed: before any byte enters the store the
manifest must verify (VALID — DSSE signature via the supplied public key,
count/batch_digest consistency, and every member's bytes re-hashed).
INVALID or INCOMPLETE refuses the whole import. Each member's
capsule_id becomes a directory name in the store, so it must be one path
segment: an id that is absolute, contains /, \, .., a NUL or other control
character, or names a symbolic link already in the store makes the whole manifest
INVALID (exit 3) before any byte is written, in signed and --allow-unsigned
mode alike. Before v0.107.0 such an id unpacked a capsule outside the store
(review storage-backup-01). Unpacking is hardened
against hostile archives (absolute paths, .., links, devices — refused) and
staged: members extract to a scratch directory and are atomically renamed into
place, so a crash never leaves a half-written capsule. After unpack, lineage
(lineage.jsonl → lineage store) and the dashboard runs_cache are reindexed
(the offline query index self-scans and needs no rebuild). Every run —
including refusals and dry runs — writes an **import receipt** JSON under
$NOVAFABRIC_HOME/import-receipts/<import_id>.json and one hash-chained audit
entry (capsule.import).
Idempotent by content address: a member whose run_id already exists
locally with byte-identical content (same deterministic re-pack hash) is
skipped, so re-running an import is a no-op and an interrupted import resumes.
Collisions are never overwritten: same run_id with different local
content is reported per-member with both hashes (exit 5); delete the local
capsule with existing retention tooling and re-import to replace it — there is
deliberately no --force.
experimental — implemented (ADR-0207 P1); API may change.
# Verified import (the normal path; key from `nova export-blob --public-key-out`)
nova import ./exports/audit-q3 --public-key export.pub.pem
# DR drill: verify + classify with zero writes, same exit codes
nova import ./exports/audit-q3 --public-key export.pub.pem --dry-run
# Air-gap courier archive
nova import batch.tar.gz --public-key export.pub.pem
# Machine-readable report
nova import ./exports/audit-q3 --public-key export.pub.pem --jsonOptions:
--public-key PATH— PEM ed25519 public key of the export signer. Required unless--allow-unsigned.--allow-unsigned— skip the DSSE signature check only (loud warning; recorded permanently asverification.mode: "unsigned"in the receipt and audit log). All content-hash/size/consistency checks still run — tamper still refuses.--capsule-dir PATH— target store root (default: wherenova capturewrites —$NOVAFABRIC_CAPSULE_DIR, else$NOVAFABRIC_HOME/capsules).--dry-run— verify + classify (would-import / would-skip / collision), write nothing to the store or indexes; the receipt is still written.--no-reindex— unpack only; skip lineage + runs-cache updates.--json— print the import receipt JSON to stdout.
Exit codes: 0 success (imported / already present), 2 usage error, 3
verification INVALID (structure, signature, digest), 4 verification
INCOMPLETE (member missing / hash or size mismatch), 5 collision(s), 6
member(s) failed during unpack. Nothing is imported on 3/4; non-colliding
members still import on 5.
nova euaiact export
Export structured EU AI Act Art.12 log events for authority access requests (Art.74). Binding for high-risk AI systems: 2026-08-02 (ADR-0076).
works today — implemented in v0.33.0.
nova euaiact export --from 2026-01-01 --to 2026-06-30 --output euaiact-report.json
NOVA_EUAIACT_HIGH_RISK=true nova euaiact export --from 2026-01-01
nova euaiact export --pretty # Rich table output to terminal
nova euaiact status # show active configurationOptions:
--from DATE— start date ISO 8601 (default: no lower bound)--to DATE— end date ISO 8601 (default: now)--capsules-dir PATH— capsule root (default:$NOVAFABRIC_CAPSULE_DIRor$NOVAFABRIC_HOME/capsules)--output, -o PATH— output file (default: stdout as JSON)--pretty— Rich table output instead of JSON
Environment:
NOVA_EUAIACT_HIGH_RISK=true— declares this is a high-risk AI system under EU AI Act Annex IIINOVA_EUAIACT_PROVIDER=true— enables provider mode (120-month / Art.18 retention floor instead of 6-month deployer floor)
Governance commands (v0.16)
nova classify run
Classify an AI system's risk tier against EU AI Act Annex III, NIST AI RMF, and OMB M-24-10, from a YAML description or from flags.
nova classify run --input system.yaml
nova classify run --input system.yaml --json
nova classify run --name "Loan scorer" --domain finance --context credit_scoring --affects-rightsOptions:
--input PATH/-i— YAML file describing the AI system (AISystemRecordfields below).--name,--domain,--contextand--descriptiongiven alongside it override the file's values.--name TEXT,--domain TEXT,--context TEXT— required together when--inputis not given;--description TEXTis optional--biometrics/--affects-rights/--general-purpose— risk-relevant properties, each with a--no-…form (default: off)--role [provider|deployer]— operator role (default:deployer)--subject-count N— estimated number of persons affected--json— emit the raw JSON result
The vocabulary versions in force are listed by nova classify list-vocabularies.
Keyword matching. Vocabulary keywords match whole words, case-insensitively (exam does not
match "example", credit does not match "accredited"), and a keyword-only prohibited practice needs
co-occurring signals: subliminal manipulation fires on subliminal or below threshold of consciousness alone, or on two of manipulate/manipulative/dark pattern; vulnerability exploitation
needs two of vulnerable/exploit…/disability/elderly/children together with
affects_fundamental_rights. A vocabulary can set strong_keywords and min_keyword_matches the same way.
The result is a rules-based screening aid, not a legal determination. Releases up to v0.107.0 matched raw
substrings and tiered a parcel-handling robot arm "prohibited".
Exit codes: 0 = classified; 1 = prohibited tier detected; 2 = bad input (file not found, YAML not a mapping, missing required flags).
AISystemRecord fields: name, description, use_case_domain, deployment_context, uses_biometrics (bool), affects_fundamental_rights (bool), is_general_purpose (bool), operator_role (provider | deployer), subject_count_estimate (int, optional).
nova classify list-vocabularies
List all available classification vocabularies.
nova classify list-vocabulariesnova classify from-capsule
Infer an AI system record from a captured run capsule's metadata and classify it. Reads
<capsule-id>/capsule.yaml from the store nova capture writes ($NOVAFABRIC_CAPSULE_DIR,
else $NOVAFABRIC_HOME/capsules), or from <data-dir>/capsules/ with --data-dir. A path
to a capsule directory works too. (Before 2026-10-23 the default was ~/novafabric-data, a
directory capture never writes, so a default run exited 2 "Capsule directory not found".)
nova classify from-capsule 01HXAY7M5JZ8R7K4P9DPBYK2WX
nova classify from-capsule 01HXAY7M5JZ8R7K4P9DPBYK2WX --data-dir /srv/nova-data --json
nova classify from-capsule ./path/to/capsule/Reference: src/novafabric/governance/, src/novafabric/cli/classify.py, ADR-0056.
Compliance audit commands (v0.16)
All five commands work over a set of capsules and take --profile / -p (default
nist-ai-rmf; a profile ID or the path to a custom profile YAML) — except
nova audit verify, which checks one report file. Without --data-dir they read the
capsules nova capture writes: $NOVAFABRIC_CAPSULE_DIR, else $NOVAFABRIC_HOME/capsules.
--data-dir / -d ROOT reads ROOT/capsules/ instead. (Earlier versions defaulted to
~/novafabric-data, a directory capture never writes, so a default audit scanned no
capsules.)
Available profiles: nist-ai-rmf, eu-ai-act-high-risk, gdpr, soc2-type2, iso42001, scientific-reproducibility.
An unknown profile ID (and no such file) exits 1. A custom profile file that is not valid
YAML, or lacks a required field (id, name, framework, version; per control id,
title, description; per evidence requirement evidence_type, description), exits 2
with one line naming the file and the field — before the L25 fix it raised a bare
KeyError traceback, and nova serve's audit endpoints answered 500 instead of 422.
nova audit map
Map capsule evidence to a profile's regulatory controls and report coverage gaps. Exits 1
if any required control has no evidence (with or without --json; releases up to v0.107.0
always exited 0). The seal_signature evidence type, and the summary's Seals: verified, mean
a capsule's NovaSeal seal verifies — signature, content binding and capsule id, the signer
self-asserted; a file merely named seal-bundle.json no longer counts, and a genuine .seal/
now does.
The pii_scan evidence type is the PII redaction manifest redaction_manifest.json that
nova capture --pii-subject writes (it must parse); every capture's redaction-proof.json is
the secret scanner's proof and no longer makes GDPR Art.17 / Art.5 / Art.25 "covered".
Not captured by this version. The tool_permissions and policy_events evidence types
read tool-permission-events.jsonl, which no NovaFabric producer writes today. When no scanned
capsule carries that file, those requirements are listed under evidence_not_captured instead
of evidence_missing: a control made only of them is reported not_applicable and excluded from
overall_score (counted in not_captured_controls); a control that also has captured evidence is
at most partial, never covered. Once any capsule carries the file, it is evaluated normally.
nova audit map --profile nist-ai-rmf
nova audit map --profile eu-ai-act-high-risk --capsule 01HXAY7M5JZ8R7K4P9DPBYK2WX
nova audit map --profile gdpr --jsonOptions: --capsule ID — audit a single capsule; --json — machine-readable output instead of a table.
nova audit report
Generate a full compliance audit report (JSON-LD by default).
nova audit report --profile nist-ai-rmf
nova audit report --profile gdpr --output report.json
nova audit report --format textOptions:
--output PATH/-o— write the report to a file (default: stdout)--format [json-ld|text]— output format (default:json-ld)
nova audit verify
Check that an audit report is well-formed JSON-LD with the report's schema. The report is not
signed, so this proves nothing about its contents: the score, the coverage and the seal flag are
what the report claims, and the output says so (Seals: … claimed by the report, not re-checked here). To recompute them, run nova audit map against the capsules. Releases up to v0.107.0
printed Report … is valid and Seals: verified from the file's own field.
nova audit report --profile nist-ai-rmf --output report.json
nova audit verify report.jsonFor a coverage threshold gate, use nova audit coverage --threshold.
nova audit bundle
Export the audit report and every contributing capsule evidence file as a ZIP archive.
nova audit bundle --profile eu-ai-act-high-risk --output audit.zipOptions: --output PATH / -o — output ZIP (default: audit-bundle.zip).
A contributing capsule that holds a symbolic link refuses the bundle (exit 1, nothing written); the bundle never follows a link.
nova audit coverage
Show regulatory control coverage for the data directory and exit 1 if the score is below the threshold — a CI gate for promotion.
nova audit coverage
nova audit coverage --profile eu-ai-act-high-risk --threshold 0.9Options: --threshold FLOAT — minimum required coverage score, 0.0–1.0 (default: 0.8).
Reference: src/novafabric/compliance/audit/, src/novafabric/cli/audit.py.
Examiner exporter commands (v0.16)
nova export-examiner bagit
Export a capsule as an RFC 8493 BagIt archive with SHA-256 checksums.
nova export-examiner bagit 01HXAY7M5JZ8R7K4P9DPBYK2WX \
--data-dir .novafabric/runs/01HXAY7M5JZ8R7K4P9DPBYK2WX/ --output-dir ./exportsOptions:
<capsule-id>(required) — used as the bag name and External-Identifier--data-dir DIR/-d(required) — capsule directory containingcapsule.yamland the payload files--output-dir DIR/-o(required) — directory that receives<capsule-id>-bag.zip--contact-email TEXT— written tobag-info.txt--no-audit-report— leaveaudit-report.jsonout of the payload even if present
A capsule directory holding a symbolic link is refused (exit 1, no bag written).
Output: BagIt ZIP containing bagit.txt, bag-info.txt, manifest-sha256.txt, tagmanifest-sha256.txt, and the capsule payload under data/. The payload is every regular file of the capsule — the streams, inputs/, outputs/ and .seal/ included — streamed and checksummed; a special file (FIFO, socket, device) is named in bag-info.txt as Payload-Omitted, and a capsule holding a symbolic link is refused. Releases up to v0.107.0 bagged only capsule.yaml, top-level ZIPs and audit-report.json, yet the checksums verified clean.
nova export-examiner pccp
Export an FDA Predetermined Change Control Plan (PCCP) document — per FDA guidance for AI/ML-based software as a medical device — comparing a baseline capsule with a proposed one.
nova export-examiner pccp --baseline 01HXBASE... --proposed 01HXPROP... \
--data-dir .novafabric/runs/ --output pccp.jsonOptions (all required): --baseline ID / -b (the current, approved version),
--proposed ID / -p (the modification to evaluate), --data-dir DIR / -d (directory of
capsule sub-directories, or the capsule directory itself), --output PATH / -o (the PCCP JSON file).
nova export-examiner iso42001
Export an ISO/IEC 42001 AI Management System evidence package for an audit period.
nova export-examiner iso42001 --data-dir .novafabric/runs/ \
--period 2026-01-01/2026-06-30 --output iso42001.zip
nova export-examiner iso42001 --data-dir .novafabric/runs/ \
--period 2026-01-01/2026-06-30 --output iso42001.zip --control 6.1 --control 8.1Options: --data-dir DIR / -d, --period START/END (ISO 8601 interval) and
--output PATH / -o are required; --control / -c selects ISO 42001 clauses (repeatable;
default: every supported clause).
Output: ZIP containing iso42001-evidence.json, control-narratives.json, capsule-index.json, and README.txt.
Reference: src/novafabric/compliance/export/examiner.py, src/novafabric/cli/export_examiner.py.