DesignationISO/IEC FDIS 24029-3
Stage50.00, FDIS registered
Planning date, an estimate16 Feb 2027
Text todayPublic record and DIS view
For AI governance, model risk and 42001 implementers
The statistical robustness standard is one step from publication.
ISO/IEC FDIS 24029-3 gives the methodology for using statistical methods to assess the robustness of neural networks. It is at stage 50.00. This site states what the public record decides, marks every clause the public view withholds, and gives you a dated preparation register.
Read from the real rules, not invented. Every answer on this site points to the official record or the held rule behind it, so you can check any line yourself. Read from the ISO record, the public DIS view, and the held text of ISO/IEC 42001:2023, ISO/IEC 23894:2023, ISO/IEC 25059:2023 and the NIST AI Risk Management Framework.
Read from the real rules222 held rules behind itCheck any line yourselfFree check, no account
No account for the free check. Not legal advice; does not certify.
The record, read 21 October 2026
- Designation
- ISO/IEC FDIS 24029-3, Edition 1
- Title
- Artificial intelligence (AI), Assessment of the robustness of neural networks, Part 3: Methodology for the use of statistical methods
- Committee
- ISO/IEC JTC 1/SC 42, Artificial intelligence
- Stage
- 50.00, 21 August 2026
- Published today
- Nothing. No 50.20, 60.00 or 60.60 date is recorded.
- Text in public
- Foreword, introduction, clauses 1, 2, 3 and 3.1, and the bibliography, in the DIS view
- Planning date
- 16 February 2027, our estimate
The ISO record and the public DIS view answer five questions and no more: what the document is, which references it makes normative, what stability means, why statistical methods are used, and where the public view stops. Everything else is withheld, and this site says so rather than filling the gap.
| Code | Unit | Group | Citation |
|---|---|---|---|
| 1 | Clause 1, Scope | Clauses | ISO/IEC FDIS 24029-3 |
| 3.1 | Clause 3.1, Stability | Clauses | ISO/IEC FDIS 24029-3 |
| I3 | High-dimensional input and why exhaustive testing fails | Introduction | ISO/IEC FDIS 24029-3 |
| P1 | Write the sampling and perturbation plan | Prepare | Our preparation method |
| P4 | Map the work to the AI management system and the risk process | Prepare | ISO/IEC 42001:2023 Annex A (held export |
| S2 | Part 2, the formal-methods companion | Series | ISO/IEC 24029-2:2023 |
29 units, 108 evidence lines. Every unit page names its source and states what it does not cover.
Know what the text says, what it withholds, and what to have ready.
The register turns the public record into the evidence a reviewer will ask for, and separates the recorded dates from our estimate.
Photograph: Startup Stock Photos on StockSnap, CC0.
02
Three things you can do before it publishes
Run the scope and date check
Five questions. It returns an applicability line, the clause behind it, three dated actions and the day the evidence pack is due. It starts blank and returns scope incomplete, never a pass.
Run the free checkRead the stability change
Stability moved from a neural-network property to an AI system property, with invariant output under specified perturbations. A test built on the Part 2 wording is not the same claim.
Read clause 3.1Take the obligation kit
The full register, the 106 evidence lines, the dated plan to the planning date and the sampling and threshold worksheets, in PDF, CSV and XLSX.
See what is in the kit- 3.1
Define the stability claim
Name the AI system, the output that must stay invariant, and the perturbations you specify.
- 1
Name the robustness property
The scope speaks of properties in the plural. Write one sentence per system, and do not borrow a definition you cannot cite.
- P1
Write the sampling and perturbation plan
The public text gives no sample size. Fix the frame, the draws, the seeds and the decision rule before the run, not after it.
- P2
Set the threshold with the risk owner
A result that cannot fail is not an assessment. The threshold is a risk decision and needs a date and a signature.
- P4
Put the result under a control row
ISO/IEC 42001 and ISO/IEC 23894 supply the management system and risk rows that make the result visible at audit.
Three questions here, five in the full check. The answer is computed by rule from the record, not by a model, and nothing is sent anywhere.
The rules this is built from
Nothing here is invented. The product reads the official public record of ISO/IEC FDIS 24029-3 and the held requirements of the frameworks below, by market, and every answer points to the one it came from.
ISO/IEC 42001:202338 Annex A controls, the AI management system the evidence feeds.
ISO/IEC 23894:202354 controls, the AI risk process that treats and monitors the result.
ISO/IEC 25059:202315 controls, the quality model that places robustness under reliability.
NIST AI RMF 1.072 controls, the United States framework the same evidence is shown under.
222 held rules in full, each cited to its control or clause. See the rules that apply to you. See the standards.
Is ISO/IEC 24029-3 published yet?
No. ISO records it at stage 50.00, final text received or FDIS registered for formal approval, on 2026-08-21. No ballot date, publication date or published edition is recorded. The planning date of 16 February 2027 is our estimate, not an ISO commitment.
What does ISO/IEC 24029-3 cover?
It gives a methodology for selecting, applying and managing statistical methods to assess robustness properties of neural networks. The public record states the subject and the three normative references. The detailed clauses are not public.
Does it apply to my AI system?
It applies if the system contains a neural network and you need to assess its robustness properties with statistical methods. The scope does not exclude bought or embedded models. It binds no one by itself; a contract, regulation or management system has to adopt it.
Can I use it for certification?
No. It is a methodology, not a management system standard and not a certification scheme. No certification body can certify against a document that is not published.
What is in the USD 799 kit?
The full register of every public unit with the evidence a reviewer asks for, the 106 evidence lines, a dated plan to the planning date, and the sampling and threshold worksheets, in PDF, CSV and XLSX. The kit ships the working register blank, with a labelled specimen on the site.
How does the change watch work?
USD 49 a month. It emails when the stage, the dates, the public text, the record or the guidance changes, with the diff. The refresh script is the same one that checks the record, so a change is caught by a machine before it is emailed.
Is this legal advice or a certification?
No. It is a reference and a working register. It does not reproduce the ISO text, it does not certify, and it does not decide your risk appetite.