NeuralMind Compliance Summary

One-page reference for procurement, security review, and compliance teams. Consolidates the NIST AI RMF, SOC 2, CMMC 2.0, and GDPR claims already documented across SECURITY-GUIDE.md and ENTERPRISE.md into a single reviewable surface.

Honest scope: NeuralMind itself is not certified to any compliance framework. The architecture supports certification of your deployment — every required control is either built in or available to switch on. The evidence below is auditable; the certifications are yours to obtain.

At a glance

NIST AI RMF (AI Risk Management Framework) — full coverage

The four NIST AI RMF functions and the evidence NeuralMind provides for each:

GOVERN — oversight, accountability, policies

• Per-tool permission policy for MCP calls (admin, builder, reader, configurable in neuralmind-backend.yaml). By default callers declare their own role, unauthenticated; with security.identity: os (v4.7.0+) the role is bound to the OS account — see SECURITY-GUIDE.md §Access Control
• Access audit trail at .neuralmind/audit_events.jsonl — every query, search, build, and MCP call logged with timestamp, action, status, and details (the actor is the value the MCP caller declares, or locally NEURALMIND_ACTOR or the OS login)
• Query provenance — every retrieval result is traceable to the specific code nodes that produced it (no black-box “trust us”)
• Audit export for your assessor: neuralmind audit export . --format jsonl (or --format cef). NeuralMind has no report generator, so build the report from the export

MAP — impact assessment, context

• Per-query explanation: which code nodes were retrieved, similarity scores, layer-by-layer disclosure (L0→L3)
• Replay overlay (v0.6.0+) shows the L3 hits a previous agent query received, on the live graph view
• Confidence signals — search results carry similarity scores; community detection labels nodes by code subsystem

MEASURE — performance, quality

• Token reduction measured per query — the 12-50× real-repo range comes from field reports (neuralmind benchmark); CI verifies a conservative floor on a fixture at every commit (tests/benchmark/), and the public benchmark reproduces on demand (python -m evals.public.run)
• Index quality metrics — top-k retrieval hit rate, escalation rate, faithfulness eval framework scaffolded
• Context tokens recorded for every query in the audit log

MANAGE — risk controls

• Secret scanning before indexing: neuralmind scan-for-secrets is a separate step, and build --redact-secrets scrubs indexed text as a backstop. Neither runs at query time
• Rate limiting enforceable at the MCP server boundary
• No anomaly detection or alerting. Denied calls are in the audit log; feed neuralmind audit export to your SIEM to alert on them

SOC 2 — Trust Services Criteria evidence

A SOC 2 report is a CPA firm’s attestation about a service organization’s controls, measured against the AICPA Trust Services Criteria. Security (the Common Criteria, CC) is in every report; Availability (A), Processing Integrity (PI), Confidentiality (C), and Privacy (P) are optional. NeuralMind runs inside your environment and transmits no repository content to us, so in your audit it is software within your system boundary, not a subservice organization.

The policies the project itself runs under (access control, change management, incident response, and others) are in compliance/, each tagged with the criteria it addresses.

CMMC 2.0 — practice evidence

CMMC assesses a defense contractor’s environment, not a tool. Level 1 covers the 15 FAR 52.204-21 requirements for Federal Contract Information (FCI); Level 2 covers the 110 NIST SP 800-171 Rev. 2 requirements for Controlled Unclassified Information (CUI); Level 3 adds selected NIST SP 800-172 requirements.

If NeuralMind indexes source code that is CUI, the index, synapse store, and audit log derived from it are CUI too, and NeuralMind becomes an asset inside your assessment scope. The table shows what NeuralMind gives your assessor and what stays your environment’s job. Practice IDs are Level 2.

The AI agent is the boundary that matters. NeuralMind hands code slices to your coding agent, and the agent sends them to its model provider. If that code is CUI, the provider is a cloud service handling covered defense information and must meet DFARS 252.204-7012 (FedRAMP Moderate or equivalent). NeuralMind reduces how much code the agent sends; it does not make a provider eligible.

To point an assessor at where a practice is implemented in your own code, neuralmind compliance . lists annotations such as // AC.L2-3.1.1: Authorized access control. Finding an annotation shows where a practice is claimed, not that it works.

GDPR considerations

NeuralMind processes code. Code can contain comments referencing names, emails, or other PII. The relevant GDPR posture:

• Lawful basis — operator-controlled (NeuralMind doesn’t make the decision; you do)
• Data minimisation — NeuralMind retrieves only the ~800 tokens of context needed per query, not the full codebase
• Purpose limitation — retrieval is per-query; no profile-building, no cross-query aggregation by default
• Storage limitation — the synapse store decays unused edges automatically (configurable via NEURALMIND_SYNAPSE_DECAY_HALF_LIFE)
• Right to erasure — synapse store and audit log are local files; rm -rf .neuralmind/ is a complete erasure path
• Data residency — entirely under operator control; no cross-border transfers initiated by NeuralMind
• Pseudonymisation — the audit log actor is NEURALMIND_ACTOR or the OS login on local queries, and caller-declared at the MCP boundary; set NEURALMIND_ACTOR to a hashed token rather than a username
• Breach notification — local files, no external surface area by default

NeuralMind does not act as a data processor in the GDPR sense in its default configuration — there is no external entity to which data is transferred, and the operator is the sole controller. The one exception is the opt-in NEURALMIND_LLM_SEED=1 documentation-synapse-seeding path, which makes Anthropic a subprocessor for README/architecture-doc prose only, if and only if an operator explicitly enables it. See compliance/THIRD_PARTY_LLM_DISCLOSURE.md for the full disclosure, including how to verify it stays disabled.

Software Bill of Materials (SBOM)

Every tagged release from v0.9.0 onward ships a CycloneDX JSON SBOM attached as a GitHub Release asset:

• File: neuralmind-vX.Y.Z.sbom.json on the release page
• Format: CycloneDX 1.x JSON — compatible with Grype, Trivy, Dependency-Track, FOSSology, and most enterprise SCA scanners
• Scope: NeuralMind package + every transitive runtime dependency with versions and licenses
• Generator: Anchore’s syft via the official anchore/sbom-action (workflow source)

The SBOM is regenerated on every tag push; you can pin a specific release’s SBOM by URL: https://github.com/dfrostar/neuralmind/releases/download/vX.Y.Z/neuralmind-vX.Y.Z.sbom.json.

Container image provenance

Container builds from v0.9.0 onward are auto-published to GHCR (workflow source):

• Registry: ghcr.io/dfrostar/neuralmind:vX.Y.Z and :latest
• Platforms: linux/amd64 + linux/arm64
• Base image: python:3.12-slim (Debian slim, official Python upstream)
• User: non-root (neuralmind UID)
• Network: no outbound calls at build time (all transitive wheels pre-downloaded in the builder stage; runtime install uses --no-index)
• OCI labels: org.opencontainers.image.source, version, licenses=MIT
• Signature and provenance (v4.14.0+): each image is signed keylessly with cosign by the release workflow and carries a GitHub build-provenance attestation. The workflow verifies both before moving :latest. To check them yourself, see compliance/SSDF.md

Deployment postures (strict → permissive)

Choose the strictest your operational needs allow — all four use the same NeuralMind binary; the difference is which network paths you cut.

Verification — every claim is verifiable

Contact

• Security disclosures: see SECURITY.md
• Compliance questions for your specific environment: open a GitHub Discussion tagged “compliance”
• Procurement / commercial questions: email hello@neuralmind.uk. The software itself is MIT-licensed and free; commercial support for deployment, integration, and evaluation is available. Evaluating teams can request a free AI-spend assessment — we benchmark NeuralMind against one of your repos and report your actual token-reduction ratio.

Document scope

This is the summary view. For depth:

• SECURITY-GUIDE.md — threat model, encryption, secrets management, SOC 2 and CMMC 2.0 summaries
• compliance/NIST.md — which NIST documents apply to NeuralMind, where each stands, and what’s next
• compliance/SSDF.md — NIST SP 800-218 practice-by-practice evidence, release verification commands, and the VEX document
• compliance/THIRD_PARTY_LLM_DISCLOSURE.md — precise, code-cited answer to “does any client content (including media files) ever reach a third-party LLM”
• use-cases/air-gapped.md — strictest deployment posture, step-by-step
• use-cases/offline-regulated.md — broader regulated-industry walkthrough
• SECURITY.md — security policy, vulnerability disclosure