Position Paper
The DAMREV Position Paper on AI Models in Valuation
Model risk and the defended NAV: where a model may help value a tokenized asset, and why a person still determines the figure
The rule does not change; the controls around it must. A model proposes and never decides: a figure from Equilibrium or from any AI system is a proposal until the person or firm that the applicable regime makes responsible has formed its own opinion and signed it.
AI may build, review, challenge, extract, and draft, but counts as none of the hands, and no AI agent writes or publishes a valuation or a NAV record.
Version 1.0Date 16 November 2026Pages 78Sections 16Sources 108

At a glance
Who stands behind the number?
A model proposes. It never decides. The figure that counts is formed and signed by a person or a licensed firm that answers for it, and nothing a model or an AI system produces is ever presented in its place.
- The engine proposesA model's output is a proposal.
- The validator challengesInputs, calibration and the figure.
- A person signsA licensed valuer forms the value.
What is inside
What the paper covers
- What a model is, and what AI addsFive kinds of model need to be told apart, from classical and simulation models to machine learning and generative and agentic AI. An agent-based simulation contains no AI agent; it is a classical model whose participants are rules in code.
- Where AI touches a valuationAI touches a valuation in three ways: to assist, to estimate, and to generate. Assisting and estimating uses are allowed under controls, and generating uses are not allowed anywhere near a value.
- The fund rules that already reach modelsBinding EU law and South Africa's hedge fund determination already require that a valuation model be explained in the valuation policy and validated before use by someone who did not build it. Neither rule mentions AI, and neither excludes it.
- How AI fails with numbersThe published evidence shows language models reading the wrong source, miscalculating confidently, giving different answers to the same question, and following instructions hidden in the documents they read. Each finding supports a control in this paper.
- Three hands, two layersIn the model layer, a second pair tests how the engine is built and behaves, and never gives an opinion on any instrument's value. In the value layer, an independent validator challenges each strike and a licensed valuer or responsible person forms and signs the value.
- A figure belongs to an engine versionEvery figure carries the engine version that produced it, and the record lets a competent outsider verify a value without recreating it. A defect means affected figures are re-run, a refinement applies to future strikes, and the original record is never overwritten.
- Version
- 1.0
- Date
- 16 November 2026
- Pages
- 78
- Sections
- 16
- Sources
- 108
The positions
The seven positions
- A model's output is a proposal, never a value
- Three hands, two layers
- AI tests the work; it never signs it
- A figure belongs to an engine version
- A value must be verifiable
- For the group's own instruments, only the licensed valuer's opinion is the independent valuation
- DAMREV claims model-assisted, human-determined, and nothing more
The paper's key figure
From model output to value
Figure 1. From model output to value: the steps between. AI may sit beside any step as a tool. It never occupies a step.
Step 1
The engine proposes
A figure with its method, inputs, uncertainty and engine version.
Step 2
The second pair tests the model
Design and behaviour across the stress battery and the benchmark set.
Step 3
The validator challenges the strike
Inputs, calibration and the proposed figure.
Step 4
A licensed valuer or responsible person signs
Forms the value and answers for it.
Step 5
Record and disclosure
Enough for a competent outsider to verify without recreating.
Contents
How is the paper laid out?
16 sections. The opening of Section 01 is open to read here. The full paper comes by email.
- Executive summaryp. 9
- Seven positionsp. 13
- 01The question: when a model's output may become a valueOpenp. 17
- 02What a model is, and where AI touches a valuationp. 19
- 03The fund rules that already reach modelsp. 22
- 04The model risk texts, and their split on AIp. 25
- 05The valuation standard setters on AIp. 28
- 06AI regulation, and where it stops short of valuationp. 31
- 07How AI fails with numbersp. 34
- 08Three hands, two layersp. 37
- 09AI tests the work; it never signs itp. 40
- 10A figure belongs to an engine versionp. 42
- 11A value must be verifiable: the recordp. 45
- 12Vendors, data, and publicationp. 48
- 13DAMREV's own instrumentsp. 50
- 14Disclosure and claim disciplinep. 52
- 15Questions for counselp. 55
- 16Sourcesp. 58
- The position on one page & Appendicesp. 57
Read the opening of Section 01
This paper surveys models and AI in valuation from primary and authoritative sources: supervisory model risk guidance, fund and asset valuation rules, AI rules and standards, recordkeeping, the valuation standard setters, and the evidence on how AI fails with numbers. It reaches seven positions.
1. A model's output is a proposal, never a value. Whatever produces it, a figure from Equilibrium or from any AI system is a proposal until the person or firm that the applicable regime makes responsible has formed its own opinion and signed it. Where a market is active, the market price governs and no model replaces it.
The rest of the paper is sent by email. Request the paper
How to read it
How do you get the full paper?
The same four steps as every DAMREV paper. The paper is sent by email, so each copy goes to a named reader.
Step 1 · You are here
Read the preview
The paper's page carries the summary, the contents and who the paper is for.
Step 2
Request the paper
Give your name, work email and organisation, and consent to receive it.
Step 3
Open the link
A link arrives by email. It expires 24 hours after it is sent.
Step 4
Read the paper
The link opens the paper in a protected reader on damrev.com, for reading only.
Request the paper
Where do we send the paper?
The form is the same one used for every paper. Further contact is optional, and the paper is sent either way.
Request the AI Models in Valuation Position Paper

What you receive
- One email from DAMREV with a personal link to this paper.
- The link opens the paper in a protected reader on damrev.com. It is for reading only and cannot be downloaded or printed.
- The link expires 24 hours after it is sent. You can request a new one at any time.
- The email carries the link only. Anything further needs your separate consent.
Common questions
What else do people ask?
Short answers to the questions people ask most about this topic.
Why is the paper sent by email?
So that each copy goes to a named reader. The request works the same way for every DAMREV paper.
What if my link has expired?
Request the paper again from this page, and a new link is sent to the address you give.
Do I have to agree to further contact?
No. Further contact is optional, and the paper is sent either way.
Can an AI system produce a valuation?
No. A figure from Equilibrium or from any AI system is a proposal until the person or firm that the applicable regime makes responsible has formed its own opinion and signed it. No AI agent writes or publishes a valuation or a NAV record.
Does an AI review count as validation?
No. AI may build code, review it, challenge a model, extract figures, and draft explanations, but it counts as none of the hands and toward no approval tier. DAMREV never calls an AI review an independent validation.
Can a figure extracted by AI enter a value?
Only as a sourced figure: one a person has tied to the document it came from, with the link recorded. Its authority is the document, not the model.
What happens to earlier figures when the engine changes?
It depends on why it changed. A defect means every figure that rests on the engine is re-run and the NAV error process applies. A refinement applies to future strikes and to figures not yet struck, and the original record is never overwritten.
Do the existing fund rules reach AI models?
In the EU, and for South African hedge funds, an AI model that contributes to a fund's value is already a valuation model under binding law. Neither rule mentions AI, and neither excludes it.
Is the paper advice or an offer?
No. The paper is provided for information only. It is not an offer of any token or financial instrument, and it is not legal, tax or investment advice.
Where next
Where should you go next?
The papers and pages below carry the detail this paper leaves to them. Each paper is requested from its own page.
- Valuation of Tokenized Assets Position PaperHow the value of a tokenized fund or asset is proposed, challenged, determined, and published, and why it is a judgement, never a fact.
- EquilibriumEquilibrium combines oracle data with human verification to model token economies and express a fair-market-value view, never an investment signal.
- Independent ValidationHow a smart contract assessor signs an attestation for each contract, and how Sentinel enforces that verdict on-chain.
- The Structuring ReviewWhether your asset, claim, and structure are ready to encode. A named person reviews your answers to nine questions.
References
- 001.Board of Governors of the Federal Reserve System, SR 26-2: Revised Guidance on Model Risk Management, 17 April 2026, with attachment. https://www.federalreserve.gov/supervisionreg/srletters/SR2602.pdf
- 019.Commission Delegated Regulation (EU) No 231/2013, Chapter III Section 7 (UK onshored text). https://www.legislation.gov.uk/eur/2013/231/chapter/III/section/7/data.xht
- 070.IPEV Board, International Private Equity and Venture Capital Valuation Guidelines, December 2025 edition. https://www.privateequityvaluation.com/Portals/0/Documents/Guidelines/2025%20IPEV%20Valuation%20Guidelines.pdf
Last updated:






