Standards

Evidence readiness, mapped to the frameworks your auditor already uses

EvalQA provides technical verification evidence, not statutory certification. This page says which framework each part of a bundle speaks to, and corrects two things earlier drafts of our own material got wrong: the NIST publication identifiers and the EU AI Act deadlines.

Last updated 13 September 2026 · Protocol v9.3

What we will not say

We do not sell regulatory panic. In 2026 the EU AI Act implementation timetable was amended by Regulation (EU) 2026/1744: obligations for Annex III high-risk systems now apply from 2 December 2027, and for Annex I systems from 2 August 2028 CLAIM: EXTERNAL VERIFICATION REQUIRED. Anyone telling you an analytics agent must be certified this quarter is selling something. What we sell instead: the earlier you have reproducible suites, invariant checks and versioned evidence trails, the cheaper every future internal audit, board review and statutory review becomes.

The mapping

FrameworkWhat it isWhat in a bundle speaks to itOur claim
NIST AI 100-1
AI Risk Management Framework 1.0
Voluntary US framework: Govern, Map, Measure, Manage.Consequence tiers and threat model (Map); invariant results with Wilson intervals, ICFY, RWAC (Measure); residual-risk acceptance formulation, recall protocol (Manage).Evidence aligned to Measure and Manage. Earlier material cited NIST SP 1270 here by mistake; SP 1270 is Towards a Standard for Identifying and Managing Bias in AI (2022) and is not the RMF.
NIST AI 600-1
AI RMF: Generative AI Profile
Companion profile for generative systems.Failure domains 13 (prompt injection), 14 (sensitive information exfiltration), 10 (catalog hallucination), 15 (cross-turn memory) map to the profile’s suggested actions on confabulation, data privacy and prompt injection.Evidence aligned. Two of those four domains are still UNVALIDATED in our benchmark and the bundle says so.
NIST AI TEVV
Testing, Evaluation, Verification & Validation
Empirical evaluation protocols from the NIST AI Resource Center.Sequential testing policy, the sequestered calibration benchmark, the stratified Sev-1 gate.Protocol adopted.
ISO/IEC 42001:2023
AI management system
Certifiable organisational management-system standard.Bundles are the technical verification evidence an AIMS points to for controls on monitoring and validation of AI systems.We are not ISO 42001 certified and a bundle does not make you so. ISO 42001 certifies organisational process, not individual query correctness.
IEEE 1012-2024
System, software and hardware V&V
The standard for independent verification and validation.The Independence Charter (technical, managerial and financial independence), evidence traceability in the twelve-element assurance case.Architected with reference to IEEE 1012 IV&V principles PENDING FORMAL THIRD-PARTY IV&V ACCREDITATION.
OWASP Top 10 for LLM Applications & AI Agent SecurityCommunity threat catalogue.Prompt injection, tool abuse, privilege escalation and memory poisoning are tested as failure domains 12, 13, 14, 15.Coverage of the listed classes; not a penetration test.
EU AI Act
Regulation (EU) 2024/1689 as amended by 2026/1744
Risk-based statutory regime.The assurance case is structured so its elements can populate Annex IV technical documentation where an analytics agent is in scope; role classification (provider vs deployer, Art. 3 / Art. 25) is assessed case by case and never assumed.Readiness evidence only. Dates above for counsel to verify.

Role classification, case by case

Whether an enterprise deploying a Cortex Agent is a “provider” or a “deployer” under the AI Act depends on actual modification and branding, not on a category. We assess it per engagement under a written protocol (Art. 3 and Art. 25) and record the conclusion in the assurance case. We do not classify by assumption, and we do not classify for you: that is a legal determination your counsel makes on our technical record.

What a bundle can and cannot do for an audit

Can

  • Show that the agent was tested against a frozen, signed denominator on a stated date
  • Show who reviewed what, with competency signatures and agreement statistics
  • Show that a defect found was remediated and re-tested (the retest bundle references the original)
  • Show the chain of custody: commitment, nonce, dual signature, notary record

Cannot

  • Certify compliance with any statute or standard
  • Replace the deployer’s own risk-management system
  • Speak to domains marked UNVALIDATED beyond “defined and tested where present”
  • Substitute for legal advice on role classification or deadlines

See the appendix in a sample bundle → Methodology