NXS · Foundations
NS-12 software, model and AI/ML artefacts: what the standard requires and who it serves
NS-12 proposes what software, models, training datasets and agents must carry to be identified, rebuilt, checked and withdrawn. This explainer covers bills of materials, build provenance, model cards, dataset sheets, agent mandates and AI governance clauses.
- As of
- 3 November 2026
- القراءة
- 10 minutes
- الأدلة
- 15 sources, cited in order
- الحالة
- Proposed standard
- الرمز
- NS-12 · Software, model and AI/ML artefacts
Status and scope
A proposed standard. It is not adopted, approved or certified by any authority.
- الرمز
- NS-12 · Software, model and AI/ML artefacts
- النطاق
- What software, models, training datasets and agents must carry to be identified, built reproducibly, attested, evaluated, governed and withdrawn.
- ينطبق على
- Software, services, migrations, workflows, models, ML datasets, evaluation records and agents.
- 1 البرمجيات
- 2 Models and kernels
- 3 Datasets for ML
- 4 الوكلاء
- 5 Governance crosswalk
- 6 القواعد
01What NS-12 software, model and AI/ML artefacts is
What NS-12 software, model and AI/ML artefacts is
NS-12 software, model and AI/ML artefacts is the proposed Nexus Standard titled Software, model and AI/ML artefacts (supply chain, model cards, dataset sheets). Its identifier is NXS-STD-ST-000012 and its current version is 0.1.0, catalogued on 16 October 2026 1. The text is on the NS-12 page.
NS-12 governs software (artefact codes SW, SV, MG and WF), models (MD) and AI/ML artefacts, namely datasets used for training, evaluation records and agents 1. For each it sets what the artefact must carry so that it can be identified, built reproducibly, attested, evaluated, governed and withdrawn 1.
Its parent is NS-1, which gives each artefact its identifier, and its verification follows NS-11 1,2,3. It is a preparatory working draft circulated for comment on 17 October 2026, and is not for implementation 1.
02Why software and models need the same discipline as documents
Why software and models need the same discipline as documents
Risk analysis increasingly runs on code and models. A flood forecast, a credit score or a supply-chain map depends on libraries, training data and model choices that the reader of the result never sees. When something goes wrong, the questions are concrete: which version ran, what was it built from, who trained the model, on what data, and what was it never meant to do?
Answers usually exist somewhere, but in scattered forms: a changelog, a notebook, a slide. NS-12 asks for them as structured records attached to the artefact's identifier, so that they can be found, checked and compared 1. It draws on practices that open communities have already built, such as software bills of materials, build provenance, model cards and dataset documentation.
03Software: bill of materials, provenance and signatures
Software: bill of materials, provenance and signatures
Clause 1 lists what every software artefact carries 1. Its identifier appears in the source header and in a codemeta.json file. Each release has a software bill of materials in SPDX or CycloneDX format, with a digest. A build provenance attestation, in the style of SLSA and wrapped in an in-toto envelope, names the builder and the source revision. Signatures cover the release artefacts. The licence is given as an SPDX identifier, with the legal state. The clauses the software implements under NS-10 are listed by identifier 1.
Releases are versions of the identifier, and the bill of materials and attestations count as evidence under NS-11 1,3. Rule NS12-R1 restates the requirement 1.
Each named format has its own steward. SPDX is an open standard published as ISO/IEC 5962:2021, at version 3.0 4. CycloneDX is published as ECMA-424, driven by the OWASP Foundation and Ecma International's technical committee TC54 5. SLSA sets levels for supply-chain security 6. in-toto is designed to ensure the integrity of a software product from initiation to installation by the end user 7. CodeMeta provides a minimal metadata vocabulary for software 8.
Fig. 01What a software release carries
Figure 1 in words
Five layers from bottom to top. Identifier: in the source header and in codemeta.json. Bill of materials: SPDX or CycloneDX per release, with a digest. Build provenance, highlighted: an SLSA-style attestation in an in-toto envelope naming the builder and source revision. Signatures: over the release artefacts. Licence and clauses: SPDX licence, legal state and the NS-10 clauses implemented.
04Models and kernels: the model card
Models and kernels: the model card
Clause 2 sets the model record 1. It carries a model card covering intended use and out-of-scope use, the training and evaluation datasets by identifier, and metrics with their evaluation records. The card also records known limitations and failure modes, uncertainty behaviour, the provenance of the training run, the licence and legal state, and the supported claim level 1.
A model without a card is not catalogued as a model 1. Scientific kernels, the numerical routines behind many risk calculations, record their golden-value tests by identifier, so that a known input can be checked against a known answer 1.
Model cards come from research practice. The paper Model Cards for Model Reporting proposed them as a framework to encourage transparent model reporting, increasing transparency into how well AI technology works 9. NS-12 keeps that intent and adds two things the paper did not require: datasets cited by identifier and a claim level bounded by a registry 1.
05Datasets for machine learning: the dataset sheet
Datasets for machine learning: the dataset sheet
Clause 3 covers training, validation and evaluation datasets 1. Each carries a dataset sheet with seven parts: motivation, composition, collection, preprocessing, uses, distribution and maintenance. It also carries the NS-3b data profile, with rights and consent states made explicit 1.
The idea follows Datasheets for Datasets, which proposed that every dataset be accompanied by a datasheet documenting its motivation, composition, collection process and recommended uses. Its latest arXiv version was revised on 1 December 2021, the month it appeared in Communications of the ACM 10.
For machine-readable exports, NS-12 points to the Croissant vocabulary where applicable 1. Croissant is an MLCommons format for machine learning datasets, at version 1.0, published on 1 March 2024, and is designed to be modular and extensible 11.
Fig. 02Model card and dataset sheet
Figure 2 in words
Column 1, model card: intended and out-of-scope use; training and evaluation datasets by identifier; metrics with evaluation records; limitations and failure modes; uncertainty behaviour; training provenance; licence and claim level. Column 2, dataset sheet: motivation; composition; collection; preprocessing; uses; distribution; maintenance; rights and consent states.
06Agents: software with a mandate
Agents: software with a mandate
Clause 4 treats an agent as software with a mandate under NS-11 1. The mandate records what the agent may do, for whom and under which clauses 1. Every artefact the agent produces credits it as an agent creator and names the mandate it acted under 1.
Agents never approve, adopt or give formal assurance (rule NS12-R3) 1. That keeps decisions with people and recorded authorities, while letting agents do recorded, checkable work. It also means a reader can always tell which parts of a record an agent produced and under what authority.
07AI governance as evaluated clauses
AI governance as evaluated clauses
Clause 5 sets the governance crosswalk 1. Obligations such as risk management, transparency, human oversight, logging and incident reporting are written as NS-10 clauses that cite their external source 1,12. They are bound to the model's or agent's contextual identifier and evaluated under NS-11. The crosswalk register names the external standards, and their text is never reproduced 1.
Rule NS12-R4 puts it in one line: AI governance obligations are clauses bound to contextual identifiers and evaluated, never prose assurances 1. Several public frameworks state such obligations. NIST released its AI Risk Management Framework on 26 January 2023, for voluntary use, to help incorporate trustworthiness into the design, development, use and evaluation of AI products, services and systems 13. The OECD AI Principles were first adopted in 2019 and updated in May 2024 14. NS-12 version 0.1.0 does not list which external frameworks its register will hold 1.
08How the pieces fit together: an illustration
How the pieces fit together: an illustration
Consider, as an illustration rather than a real case, a flood forecasting model used by a regional programme. Under NS-12 the model has its own identifier and a model card. The card names its intended use, such as short-range river forecasts, and its out-of-scope uses, such as insurance pricing. It lists the training and evaluation datasets by identifier, the metrics with their evaluation records and the known failure modes 1.
Each of those datasets has a dataset sheet saying why it was collected, what it contains, how it was processed and under what rights 1. The code that runs the model is a software artefact with a bill of materials, a build attestation and signatures 1. If an agent prepared the weekly forecast bulletin, the bulletin credits the agent and names its mandate 1.
A reader who doubts a forecast can follow the identifiers from the bulletin to the model, from the model to its data and from the code to its build. Each link is a record, not a recollection.
09Withdrawal and supersession
Withdrawal and supersession
NS-12 also asks that artefacts can be withdrawn 1. Under NS-1 (rule NS1-R3), identifiers are never deleted or reused: withdrawn artefacts keep their record, marked WITHDRAWN, and replaced ones record a supersededBy link 2. For models this matters a great deal. When a model is found to fail in a setting it was never meant for, the record should show when it was withdrawn and what replaced it, so that earlier results can be re-examined against the right version.
Because releases are versions of one identifier, a bill of materials and attestation exist for each version 1. A vulnerability found in one dependency can then be traced to every release that carried it.
10The four rules in one place
The four rules in one place
Clause 6 condenses NS-12 into four rules 1. NS12-R1: software releases carry a bill of materials, a build attestation and signatures as evidence of the version. NS12-R2: models carry a model card, datasets carry a sheet, and both cite their sources by identifier. NS12-R3: agents act under recorded mandates and are credited as agents, and they never adopt or give formal assurance. NS12-R4: AI governance obligations are clauses bound to contextual identifiers and evaluated, never prose assurances.
Fig. 03From duty to evaluated clause
Figure 3 in words
Four rising steps. External obligation: risk management, transparency, human oversight, logging or incident reporting, from a cited source. Clause: written under NS-10, citing the source without copying it. Binding: attached to the model's or agent's contextual identifier. Evaluated, highlighted: attestations under NS-11, never prose assurances.
11Differences and open questions
Differences and open questions
Three points are worth noting for implementers. First, NS-12 allows two bill of materials formats, SPDX and CycloneDX 1. Both are open, but they differ in structure, so a register that compares releases across projects needs a mapping between them. Second, SLSA lists more than one version of its specification on its site 6, while NS-12 asks for an SLSA-style attestation without naming a version 1. Third, NS-12 names Croissant only where applicable, and does not say when a dataset sheet alone is enough 1.
These are matters a later version could settle through an NS-13 proposal 15.
12Who NS-12 serves and how to take part
Who NS-12 serves and how to take part
NS-12 serves those who build, buy or rely on software and models: developers, data scientists, procurement teams, risk officers and regulators. It gives each a fixed list of what to ask for and where to find it 1.
Changes follow the open process on Develop a standard, and comments go through the contact form. The standards are published by The Global Centre for Risk and Innovation (GCRI) as Secretariat. The Global Risks Forum (GRF) and The Global Risks Alliance (GRA) take part as stakeholders like any other. Learners can study related methods in Risk Academy's Technology and Data course.
13Instruments named in or related to NS-12
Instruments named in or related to NS-12
The table lists external instruments NS-12 names or resembles, as published on 3 November 2026. Each is maintained by its own issuer and applies on its own terms.
| الأداة | Relation to NS-12 | الطبعة |
|---|---|---|
| SPDX | Bill of materials and licence identifiers | 3.0; ISO/IEC 5962:2021 4 |
| CycloneDX | Bill of materials | ECMA-424 5 |
| SLSA and in-toto | Build provenance attestation | Live specifications 6,7 |
| Model Cards | بطاقة النموذج | arXiv v2, January 2019 9 |
| Datasheets for Datasets | Dataset sheet | arXiv v8, December 2021 10 |
| Croissant | Machine learning dataset metadata | 1.0, 1 March 2024 11 |
| NIST AI RMF | Source of governance obligations | 1.0, 26 January 2023 13 |
What this page does not claim: NS-12 is a proposed, preparatory working draft that its own page says is not for implementation, and no Nexus Standard has been adopted by the Global Standards Council or any National Standards Council. A model card, dataset sheet or attestation under NS-12 records information; it does not vouch for a model's quality or safety. The SPDX project, OWASP and Ecma International, SLSA, in-toto, MLCommons, NIST, the OECD and the authors of the cited papers maintain their own work; Nexus Standards has no affiliation with any of them, and naming a format is not an alignment claim.
أسئلة
أسئلة وأجوبة
What must a software release carry under NS-12?
A software bill of materials in SPDX or CycloneDX with a digest, a build provenance attestation, signatures, an SPDX licence and the clauses it implements.
Can a model be catalogued without a model card?
No. A model without a model card is not catalogued as a model. The card covers intended use, data, metrics, limitations and provenance.
What goes into a dataset sheet?
Seven parts: motivation, composition, collection, preprocessing, uses, distribution and maintenance, with rights and consent states made explicit.
What may an agent do under NS-12?
Only what its recorded mandate allows. Its outputs credit it as an agent, and it never approves, adopts or gives formal assurance.
What is NS-12's own status?
Proposed. Version 0.1.0 is a preparatory working draft circulated for review and comment, not for implementation, and changes go through NS-13 proposals.
الشروط
المسرد
- Software bill of materials
- A list of the components and dependencies in a software release, in a standard format.
- Build provenance
- A signed record of who built a release, from which source revision and how.
- بطاقة النموذج
- A structured description of a model's intended use, data, metrics, limitations and provenance.
- Dataset sheet
- A structured description of a dataset's motivation, composition, collection, uses and maintenance.
- Golden-value test
- A test that checks a computation against a known correct answer for a known input.
ذو صلة
Further reading across the network
- All Nexus Standards
- NS-12: Software, model and AI/ML artefacts
- NS-1: The UNI and the artefact catalogue
- NS-10: Policies and clauses
- NS-11: Verification and monitoring
- NS-13: Standards development process
- NS-14: Conformance and crosswalk
- وضع معيار
- حوكمة Nexus Standards
- المسرد
- Global Risks Forum
- Global Risks Alliance
المصادر
المراجع
15 sources, numbered in the order they are first cited
- Nexus Standards (2026-10-16). NS-12: Software, model and AI/ML artefacts (proposed). nexustandards.com. Accessed 2026-11-03.
- Nexus Standards (2026-10-15). NS-1: The UNI and the artefact catalogue (proposed). nexustandards.com. Accessed 2026-11-03.
- Nexus Standards (2026-10-16). NS-11: Verification and monitoring (proposed). nexustandards.com. Accessed 2026-11-03.
- SPDX project, The Linux Foundation (2026-11-03). SPDX Specifications. spdx.dev. Accessed 2026-11-03.
- OWASP Foundation and Ecma International TC54 (2026-11-03). CycloneDX Bill of Materials Standard (ECMA-424). cyclonedx.org. Accessed 2026-11-03.
- SLSA, The Linux Foundation (2026-11-03). SLSA: Supply-chain Levels for Software Artifacts. slsa.dev. Accessed 2026-11-03.
- in-toto project (CNCF graduated project) (2026-11-03). in-toto: a framework to secure the integrity of software supply chains. in-toto.io. Accessed 2026-11-03.
- CodeMeta Project (2026-11-03). The CodeMeta Project. codemeta.github.io. Accessed 2026-11-03.
- arXiv (cs.LG) (2019-01-14). Model Cards for Model Reporting. arxiv.org. Accessed 2026-11-03.
- arXiv (cs.DB); published in Communications of the ACM, December 2021 (2021-12-01). Datasheets for Datasets. arxiv.org. Accessed 2026-11-03.
- MLCommons (2024-03-01). Croissant Format Specification. docs.mlcommons.org. Accessed 2026-11-03.
- Nexus Standards (2026-10-16). NS-10: Policies and clauses (proposed). nexustandards.com. Accessed 2026-11-03.
- National Institute of Standards and Technology (NIST) (2023-01-26). AI Risk Management Framework. www.nist.gov. Accessed 2026-11-03.
- Organisation for Economic Co-operation and Development (OECD) (2024-05). OECD AI Principles. oecd.ai. Accessed 2026-11-03.
- Nexus Standards (2026-10-16). NS-13: Standards development process (proposed). nexustandards.com. Accessed 2026-11-03.
السجل
Editorial record
Facts as of 3 November 2026. The NS-12 text, version and status were read on its page that day, and every external instrument was checked at its issuer the same day. The EUR-Lex page refused automated access, so no European Union legal act is cited. This page is reviewed at least every 180 days and whenever NS-12 changes version or status. Text, visuals and data on this page are licensed CC BY 4.0.
Published by the Standards desk, Nexus Standards. Drafted with AI assistance under the Nexus editorial constitution. To report an error, use the corrections page; corrections are dated and listed here.
Cite this article
Standards desk, Nexus Standards (2026, November 3). NS-12 software, model and AI/ML artefacts explained. Nexus Standards. https://nexustandards.com/ns-12-software-models-and-ai-ml-artifacts/
@online{standards2026ns12softwaremodelsan,
author = {{Standards desk, Nexus Standards}},
title = {{NS-12 software, model and AI/ML artefacts explained}},
publisher = {Nexus Standards},
year = {2026},
month = {11},
url = {https://nexustandards.com/ns-12-software-models-and-ai-ml-artifacts/},
language = {en}
} TY - ELEC AU - Standards desk, Nexus Standards TI - NS-12 software, model and AI/ML artefacts explained PB - Nexus Standards PY - 2026 DA - 2026/11/03 UR - https://nexustandards.com/ns-12-software-models-and-ai-ml-artifacts/ LA - en ER -
Republish
Republish or translate this article free of charge, in print or online, under CC BY 4.0: keep the byline and Nexus Standards, link to the original and note any change. Figures, maps and twins made by the institutions carry the same licence; third-party data keeps its own credit line.
Attribution to paste
<p>NS-12 software, model and AI/ML artefacts: what the standard requires and who it serves — Standards desk, Nexus Standards, Nexus Standards. First published 3 November 2026 at <a href="https://nexustandards.com/ns-12-software-models-and-ai-ml-artifacts/">https://nexustandards.com/ns-12-software-models-and-ai-ml-artifacts/</a>. Republished under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).</p>
Comment on this draft
Comments are taken clause by clause, with the exact text you propose and your evidence, through the comment form on this site.