Methodology and verification policy

What gets listed

• The tool or resource is real and publicly reachable at an official URL.
• A public page from the vendor or publisher describes it as supporting AI assurance, verification, safety, governance or compliance.
• We can cite that page, and the listing quotes or closely paraphrases what it says.

The directory is not exhaustive. It lists a small number of general governance platforms as a baseline and concentrates on tools with a stated link to aerospace, defense, space, safety-critical use, or one of the frameworks we cover.

How tags are assigned

• Framework tags appear only when the cited page names the framework in connection with the product. A framework mentioned only as a topic, a blog category or a marketing keyword does not count. We never infer that a tool "supports" a framework.
• Industry tags appear only when the cited page names the industry, a customer story in it, or an offering aimed at it. Everything else is tagged cross-industry.
• Who it's for and Deployment (on-prem, air-gapped, or SaaS) appear only when the cited vendor page states them in those terms. If the page does not say, the fields stay blank. We do not infer SaaS from a marketing homepage.
• Open source is a badge and a filter, not the same as the Type field. The badge appears only when the official repository LICENSE or COPYING file is an OSI-approved license (the SPDX id is recorded on the listing). Public source under CC BY 4.0, Community Specification License 1.0, or no license file does not get the badge. Marketing use of the words "open source" is not enough.
• Listings show the evidence text next to its source link so readers can check it.

What we do not claim

• We do not say a tool is certified, qualified, accredited or approved by any authority unless the cited page says so, and we say who said it.
• We do not rank, score or recommend tools.
• A listing is not an endorsement. A framework tag is the vendor's statement, not our evaluation.

Coverage cells are a separate layer. Cited mapping versus analyst assessment explains when a source names a requirement and when we add an editorial judgment from a documented capability.

Sources for explainers and articles

• Framework explainers cite the issuing body or official legal text (for example EUR-Lex, eCFR, FAA, EASA, NIST, ISO, DoD CIO).
• Every factual statement in an article carries a link. Sentences that are our analysis are labeled "Analysis" or "Our reading".
• Where a standard is sold, we describe it from the publisher's own public page and do not reproduce it.
• Statuses that change (CMMC phases, EU AI Act dates, ED-324 / ARP6983) carry an "as of" date. Where two sources disagree, we show both.
• We state when we searched for something and found nothing, instead of asserting a negative.
• U.S. federal terminology. Where this site speaks in its own voice about current U.S. federal executive-branch policy, it uses Super Intelligence (SI) and notes that those systems were formerly called artificial intelligence (AI). Official titles of cited documents, standards, frameworks, and sources stay exactly as published. Official names of statutes, regulations, historical documents, and non-federal standards stay exactly as published. The cited explainer is Executive Order 14434.

What we reviewed and did not list

Absence from the directory means we could not verify a qualifying claim, not that a tool is unsuitable.

• Conventional DO-178C verification tools. VectorCAST stays out: the pages we reviewed describe verification of conventionally written software and do not claim assurance of machine-learning models. LDRA's AI and machine learning capability page is listed separately because it names EASA guidance for Level 1 and 2 machine-learning applications. Rapita Verification Suite is listed from the product page's DO-178C qualification-kit sentences (more than 20 DO-178B and C projects up to DAL A); that listing is not an ML-model verifier.
• F5 AI Guardrails (the page at calypsoai.com). The "defense" wording on the page we reviewed refers to defending AI systems against threats such as prompt injection, not to the defense industry.
• Daedalean. Its former web address now serves the Destinus site, and we found no current product page documenting ML certification tooling.
• Thales AI Toolkit. The official catalogue page exists, but our retrieval of it was blocked, so we could not quote evidence from the page itself.
• Galois AI + formal methods. The public page describes a practice, not a named product we could list as a tool.
• Ansys SCADE Vision and Applied Intuition. SCADE Suite is listed because the product page names KCG as qualified for DO-178C/DO-330 at TQL-1 and names Avionics and Flight Controls. A webinar titled “AI-Driven Solutions for Embedded Software Design and V&V” is not treated as a product claim that SCADE verifies machine-learning models. SCADE Vision stays out. Applied Intuition pages we reviewed describe automotive or physical-AI verification without a cited aerospace or defense offering that met the listing rule. Ansys medini analyze remains a separate listing because the safety-analysis pages name SOTIF hazards from perception-algorithm limitations.
• Commercial tools naming JSP 936, NATO PRUAI, UL 4600, ISO/IEC 23894 or ISO/IEC TR 5469. Official product pages we fetched still do not name those five. The JSP 936 filter now lists the official MOD AI Practitioner's Handbook. The other four stay Resources-only.
• optro.ai homepage versus Optro AI Governance. Optro AI Governance is listed from the product page at optro.ai/product/ai-governance (fairnow.ai currently redirects there). The optro.ai homepage is a separate GRC Intelligence surface (formerly AuditBoard) and is not mixed into that listing.
• CISA AI landing and UK Government AI Playbook HTML. cisa.gov/ai timed out again this pass, so it is not listed. The GOV.UK AI Playbook listing page exists; its HTML sibling 404ed, so it is not listed. Canada’s Algorithmic Impact Assessment is listed this round as a government-resource questionnaire (65 risk and 41 mitigation questions; MIT-licensed prototype).
• Eclipse OpenCert. The Eclipse project page still names DO178C, ISO 26262 and EN 50128/50126/50129 for standards knowledge management, but the latest release is 1.0.0 (Neon 3) dated 2019-06-30 and the GitLab project note we fetched is 16 June 2021, so it was not listed. CAISAR and PyRAT are the neural-network safety-engineering listings this round (PyRAT is the CEA AISER verifier; CAISAR orchestrates it among other provers).
• Llama Prompt Guard 2 versus Llama Guard 4. Llama Prompt Guard 2 is a BERT prompt-attack classifier (benign/malicious). Llama Guard 4 is a separate 12B multimodal content-safety classifier. Both are under the Llama 4 Community License, so neither has an OSI badge. Prompt Guard is not on the aircraft path.
• Saidot ISO 27001. Saidot is tagged EU AI Act, ISO/IEC 42001 and NIST AI RMF from the product and homepage policy sentences. The homepage ISO/IEC 27001:2022 badge is Saidot’s own ISMS certificate, not a customer AIMS template, so it is not tagged. Azure Marketplace is not the word SaaS.
• Portkey versus Prisma AIRS Network Intercept. Portkey Guardrails is listed from the Gateway Guardrails docs (input/output checks, HTTP 246/446). The homepage now says Portkey is PRISMA AIRS AI Gateway; that rebrand is not mixed into the separate Prisma AIRS Network Intercept listing. New fully air-gapped Portkey deployments are documented as no longer offered, so air-gapped is not tagged.
• Arthur Policy Templates versus the ISO 42001 column. Arthur’s docs name four built-in Policy Templates without EU AI Act / NIST AI RMF / ISO 42001 titles. Those three frameworks are tagged from the official column that maps Arthur traces, evals and guardrail telemetry to ISO 42001 auditor evidence and says the same evidence base supports an EU AI Act risk assessment and a NIST AI RMF self-attestation.
• Collibra’s own ISO 42001 certificate. Collibra AI Command Center is tagged EU AI Act and NIST AI RMF from the product and docs pages. Collibra’s January 2025 ISO 42001 certification (vendor AIMS) is not treated as a customer 42001 template, so that framework is not tagged.
• D-RisQ Modelworks. Official pages name DO-178C/DO-333 and autonomous vehicles in air. The product is Simulink/Stateflow formal-methods design verification, not a neural-network toolbox, so it was not listed. NNDK and CAISAR are the neural-network safety-engineering listings.
• CISA SBOM-for-AI HTML. The May 2026 software-bill-materials-ai-minimum-elements page timed out on this pass. The provenance article keeps that URL from the existing citation and uses the live 29 July 2026 general SBOM HTML page for the new “may require additional elements” sentence.
• IMDA AI Verify portal. aiverify.imda.gov.sg returned HTTP 500 this pass. The AI Verify Testing Framework listing uses the AI Verify Foundation pages instead.
• α,β-CROWN. The official repository is a VNN-COMP winner under BSD-3-Clause. The pages we fetched do not name aerospace or defense, so it is not listed. It is cited in the formal-verification article as a tool we read and left out.
• TensorFlow Model Card Toolkit. Listed from the GitHub README (Apache-2.0). The repository is archived. The TensorFlow.org guide currently redirects unauthenticated clients to Google sign-in, so the listing does not cite that guide. Official pages do not name aerospace, defense, NIST AI RMF, ISO/IEC 42001 or DoD Responsible AI, so those tags are omitted.
• NASA IV&V Program page. The official IV&V services page we fetched does not name AI, so it is not listed as an AI-assurance tool.
• Commercial eVTOL or powered-lift AI-assurance products. Official product pages we fetched do not name eVTOL or powered-lift together with AI assurance, so none are listed.
• NASA ProgPy OSI badge. ProgPy is listed. The GitHub license metadata is Other / NOASSERTION, so it does not get the OSI badge.
• AdaCore QGen. Official pages name DO-330 / TQL-1 for a Simulink/Stateflow code generator. They do not name machine-learning model assurance, so it stays with other conventional DO-178C verification tools we left out.
• MAITE as a separate listing. JATIC documents name the maite library. The GitHub LICENSE file was not fetchable on this pass (404), so we did not list MAITE as its own tool or award an OSI badge. The CDAO JATIC program listing covers that ecosystem.
• Protect AI LLM Guard. The official GitHub repository we fetched is archived, so we did not list that runtime product. Check Point AI Guardrails (the Lakera Guard API) is listed separately from Check Point docs.
• Argevide NOR-STA. Official pages describe a second Argevide product. PREMIS is already listed for Core GSN; we did not add a second listing from the same vendor on this pass.
• Frameworks with no listed tools: FAA AI Safety Assurance Roadmap, ISO/IEC 23894, ISO/IEC TR 5469, NATO PRUAI, DO-200B, NIST SP 800-218A, EU MDR + AI Act. We found no tool whose public page names these as supported, so we list none. Those filter pages still list official documents and working-group pages and stay indexable once that content is present.

Verification log

Independent fact-check of the directory, explainers and articles. Review date: 15 October 2026.

Each tool was checked for product name, what the cited page says it does, framework tags and industry tags. Vendor rename, acquisition and shutdown were searched on official or primary pages. Article and explainer claims prioritized dates, version numbers, CMMC phase status, EU AI Act dates, ED-324 status, and “must” / “required” language. Where a cited URL returned 403, an official mirror was used (EUR-Lex HTML for Regulation (EU) 2026/1744; DoD CIO for CMMC; EUROCAE WG-114 for ED-324; NIST publication pages for AI 600-1 and SP 800-171).

Corrections and remaining risks are recorded on the updates page for this pass.

The 30-claim pass prioritized “must” / “required” / “usable” / “paused” language, CMMC Phase I vs Phase II wording, HMT “requirements,” ACAS Xu as if certified, EASA “first-class” data language, RTA ancestry, the Hugging Face “defense” industry tag, and matrix cells attributed to official catalogs. Before/after wording is on the updates page for this pass.

Before/after wording for the three corrections is on the updates page for this pass.

Before/after wording for the Issue 03 correction is on the updates page for this pass. Round 11 already covered thirty claims from rounds 8–10; this pass sampled remaining dated, “must”, and taxonomy claims from rounds 1–7 plus FAA/EASA wording those earlier tables did not re-open.

Round 11 already covered thirty claims from rounds 8–10, Round 13 the space/eVTOL/MRO/EO 14434 articles, and Round 16 the DO-330 / G-34 / CDAO T&E / 600-1 / CMMC scoping / JSP 936 supplier articles. This pass sampled remaining dated and “must” claims from the FedRAMP, NATO, Allied-policy, ISO engineering, IAQG/UKAS, DFARS, EASA NPA, DAF, Japan MOD and Part 108 pages those tables did not re-open. Zero corrections.

Rounds 19 and 20 already covered thirty claims from rounds 1–18. This pass sampled dated catalog, entry-into-application, and listing claims from the Round 21 UL 4600 and ISO/IEC 42005 articles, the Round 22 ISO/IEC 24029 and GPAI articles, and three Round 21–22 listings (nnenum, Dioptra, UK MOD AIPH / DAIC). Zero corrections. The MathWorks ACAS Xu “45 different networks” sentence was also live on the cited example page this pass; it is not a sixteenth table row.

Rounds 19, 20 and 23 already covered forty-five claims from rounds 1–22. This pass sampled dated, audience, deployment and “MUST” claims from the Round 23 Trustible / Saimple / QuantPi / LYNX listings and the Round 24 ART / AIA Toolkit / SPDX / PyRIT / Promptfoo / Dataset Cards listings. Zero corrections. Zero live re-fetches blocked this pass. The Saimple homepage Article 15 sentence was not re-fetched here; claim 3 is the product-page 24029 / SaaS / on-premises wording only.

Rounds 19, 20, 23 and 25 already covered sixty claims from rounds 1–24. This pass sampled dated, version, language and deployment claims from the Round 25 HiddenLayer / Resaro / MathWorks / garak / NeMo / PREMIS listings and the Round 26 nnenum / HEART / Guardrails AI / Bedrock Guardrails listings. Zero corrections. Zero live re-fetches blocked this pass. The MathWorks DO Qualification Kit page was live on this editorial fetch (it has been 403-blocked on some automated checker runs). HEART contract number W519TC-23-9-2037 was confirmed on the IBM/heart-library README in the Round 26 listing evidence; this pass re-fetched pyproject.toml for the 0.7.0 / Python range and kept the contract sentence from that already-cited README.

Rounds 19, 20, 23, 25 and 27 already covered seventy claims from rounds 1–26. This pass sampled version, language, audience and template claims from the Round 26 Croissant / Drata / Astah listings and the Round 27 ERAN / CycloneDX / NNV / ATLAS / Inspect / Hyperproof / Azure Content Safety listings — none of which sat in the Round 27 table. Zero corrections. Zero live re-fetches blocked this pass. The Drata Help Center article is JavaScript-heavy; the 1 April 2025 date and AIMS / ISO/IEC 42001 sentences were present on the live HTTP 200 body. The Drata marketing product page remains 403-blocked on some automated checker runs.

Round 18 (14 October 2026): added cited articles on proposed FAA Part 108 BVLOS versus DoD Directive 3000.09, ISO/IEC 23894 and 5338 for engineering teams, and DFARS/FAR supplier flow-down (no AI- or Super Intelligence-titled clause found in live Part 252 text). Folded the aviation publisher-map into the FAA/EASA timelines page and expanded thin Australia Defence, EASA levels, ISO/IEC 23894 and NIST AI RMF pages. The Model Card Toolkit listing note above replaces an earlier “not listed” sentence from before that GitHub README listing.

Questions

Keeping it current

Each listing and explainer shows the date its sources were last checked. Tool and framework pages also show a Last verified date from src/data/freshness.json, written by npm run check:freshness when that check is run in an editorial pass. Directory changes are recorded on the updates page, the What changed this month digest generated from that log, and the RSS feed. A corrections and submission process will be published at launch; none is open yet.

Between editorial passes, npm run check:freshness re-fetches every cited URL and every tool evidence quote. It reports DRIFT when a quoted phrase is no longer on the live page, and BLOCKED or UNCHECKABLE when the page cannot be read (403, PDF, empty body, CAPTCHA). HTTP 500 and network timeouts are blocked, not failed, so a flaky issuer site does not fail the weekly job. The script does not rewrite listings. A weekly GitHub Action runs the same check and opens or updates a single Freshness review issue; it reopens that issue when there is drift or a hard failure and closes it when the run is clean. If the check does not write a report, the issue still posts so the standing ticket does not go silent. The workflow does not edit content and does not commit freshness.json (CI skips writing that file). How to run the check locally is documented in this repository's README. Go-live steps that a human must not run from a pull request are in the repository file docs/LAUNCH.md.

Site search is a static Pagefind index built at build time and queried in the browser. It does not call an external search API and it does not refresh issuer documents by itself.

Sponsorship

Sponsored placements will be labeled "Sponsored", kept visually separate from listings, and will not change which tools are listed or how they are described. That rule is in the listing policy. The sponsor media kit restates the audience and those rules and does not publish traffic or prices. No sponsor is placed today. The sponsor slots on this site are placeholders. To point us at a public URL, use the suggest a tool placeholder (not connected yet).