2028
From 1 January 2028, Illinois requires large frontier AI developers to obtain an annual independent audit.
Latham & Watkins, July 2026 In development · The auditor’s discipline, applied to AI
Independentassurancefor AI.
Beyond the claim.
What the evidence shows, not what was said.
Givmark is bringing chartered-accountant audit standards to AI systems.
See the methodWhy it matters
Claims about AI are multiplying. Independent checking of them has barely begun.
The claims
AI developers publish safety frameworks, model cards and risk reports. AI agents report on the work they have done. Each is a claim.
The requirement
Regulation is beginning to ask for some of them to be checked independently. Illinois’ AI Safety Measures Act, signed in July 2026, requires the largest frontier AI developers to obtain an annual independent audit from 2028, including an assessment of their internal controls.
The method
Audit has a tested answer: agree the scope, gather the evidence, test the controls, and sign a report that says what was and was not examined. We are building that discipline for AI, with technical testing by or with specialist evaluators.
- Safety framework
“Every model is evaluated before release.”
Evaluation recordsTraced to the record
- Controls statement
“Safeguards operated all year.”
Approval logs, sampledTested across the period
- Agent task summary
“Task completed. No issues.”
Operator logGap: logged incident not reported
$500m
Annual revenue, with affiliates, above which a frontier developer is covered by the Illinois audit requirement.
Morrison Foerster, July 202610 to the power of 26
Operations of training compute above which a model counts as a frontier model under the Act.
Morrison Foerster, July 2026Figures from law firm summaries of Illinois Senate Bill 315. Not legal advice.
The questions
Five questions assurance asks.
Claims and evidence
Is it supported?
Safety frameworks, model cards and risk reports make claims. An independent check traces each one to the evidence behind it, or records that the evidence is missing.
Evidence examined
Published claims · internal records · test results
Givmark / Working paper
01 / The claim
Published safety framework · section 4.2 · Developer A
“Every new model is evaluated for dangerous capabilities before release.”02 / The evidence
Release log for the period: three model updates. Evaluation records found for two.
03 / Claim not yet supported
One release has no evaluation record. Ask why before relying on the claim.
Internal controls
Do the safeguards work?
A control on paper is not a control in practice. An independent check tests whether safeguards were designed well and whether they actually operated, across the whole period.
Evidence examined
Policies · approvals · logs
Givmark / Working paper
01 / The claim
Release control · approval matrix · Developer A
Every release requires prior sign-off from the safety lead.02 / The evidence
Sample of 12 releases: 11 approved in advance, 1 approved two days after release.
03 / Design versus operation
The control exists but did not operate once. Test whether the exception was found and fixed.
Known answers
What happened, not what was reported.
Some questions are settled only by testing. Where the right answer is known in advance, what a system reports can be compared with what it actually did. Technical testing is done by or with specialist evaluators.
Evidence examined
Reported result · known answer · difference
Givmark / Working paper
01 / The claim
Synthetic bookkeeping task · agent close report, line 3
“Bank reconciled. Unreconciled difference: nil.”02 / The evidence
Known answer: £8,412.00 posted to suspense, no source document.
03 / Reported versus real
The report and the books disagree. The gap is recorded with the number of runs behind it.
Incidents on record
What went wrong, and was it said?
Real incidents are evidence too. Each case is recorded with its sources, graded on the strength of that evidence, and open to a right of reply.
Evidence examined
Sources · evidence grade · response
Givmark / Working paper
01 / The claim
Incident register · entry 07 · operator log
Logged: an AI agent deleted a test database during a routine task.02 / The evidence
The agent’s own task summary: “Completed with no issues.”
03 / Graded on its evidence
The summary left out a logged incident. Record both sources, grade the evidence, and note what is still unknown.
Calibrated reporting
Say what was examined.
A report says what was examined, what was not, and how confident its conclusion is. It is signed by the person accountable for it.
Evidence examined
Scope · limits · signature
Givmark / Working paper
01 / The claim
Draft report · conclusion, paragraph 3
Draft wording: “Controls are effective.”02 / The evidence
Work done: two of five control areas tested; sample sizes recorded in the working papers.
03 / Scope before conclusion
Reword to say what was tested and what was not. A conclusion cannot be wider than its evidence.
Where Givmark fits
Testing is one discipline.
Assurance is another.
AI assurance will need many kinds of expertise. We are building one part of it, alongside the specialists who already do the technical work.
DisciplineThe questionHow Givmark relates
Technical evaluation
What can the system do?Specialists test capabilities and safeguards. We intend to work with them, not replace them.
Standards and policy
What should an audit cover?Researchers and policymakers are setting the direction of the field. We aim to contribute a practitioner’s view.
Audit discipline
Is the claim supported?Scope, independence, controls testing and calibrated, signed reporting. In development.
The method
Scope. Evidence.
A signed conclusion.
Chartered accountants have long practice in giving an independent view on what organisations report. We are adapting that method to AI.
01Scope
Agree what is examined.
Write down exactly what is in scope, against which criteria, and what is not. Confirm independence before any work begins.
02Evidence
Gather and test the evidence.
Trace claims to records. Test controls across the period, not on a single day. Technical testing is done by or with specialist evaluators.
03Calibrated report
Report what was found, and its limits.
A conclusion no wider than its evidence, the method behind it, and a named person accountable for it.
A mark with something behind it
Draft principles
Before any mark
- IndependentNo financial tie to what is examined
- EvidencedEvery conclusion traced to its source
- CalibratedConfidence and limits stated
- SignedA named person accountable
Status
Principles in development. No assurance has been issued.DRAFT · 2026-10Givmark’s working principles. Draft, October 2026.
No mark
without evidence.
Four principles we are building to. They are also fair questions to ask of anyone who checks AI, Givmark included.
A professional, independent way to examine claims about AI systems: a written method, a way of testing internal controls, and reports that say what was and was not examined. It is in development. Givmark does not currently carry out audits or issue assurance opinions.
Who is behind Givmark
Founded by a chartered
accountant who builds with AI.
Givmark was founded by Ben Taylor FCA, a PwC-trained chartered accountant with twenty years in institutional finance. He designs, builds and runs production multi-agent AI systems, directing AI coding agents and auditing their output himself.
Professional foundation
PwC · ACA · FCA
Trained as an external auditor at PwC from 2001 to 2004. ACA-qualified in 2004 and a Fellow of the Institute of Chartered Accountants in England and Wales.
Institutional finance
RBS · Citco
Twenty years in finance functions where controls are the job: global transfer pricing and finance change at RBS, then Senior Vice President, Finance at Citco, a global fund administrator.
AI engineering
Production multi-agent systems
Designs, builds and runs production multi-agent AI systems. Our own tools are built with commercial AI models, including Anthropic’s Claude.
In developmentFor evaluators, researchers, funders and policy teams.
Start a
conversation.
Givmark is at an early stage. If you work on AI evaluation, audit or policy, or want to follow the work, write to us.
ben@givmark.org