Independent AI transformation assurance

Know what your AI changed.

When a model is asked to change one thing in a message, it often changes several others at the same time. Tone turns more emotional, commercial pressure rises, nuance disappears, a fact quietly shifts. Kronaxis measures what actually moved, what drifted, and whether the result stayed inside the bounds you set, with independent readers rather than the model grading itself.

Provable, not plausible. Grounded in published research, the Distinct Fields series, ten papers on Zenodo with a full replication package.

The problem

A prompt tells you the request. Not the full effect.

Generative AI now rewrites customer communications, personalises content, summarises evidence and acts on behalf of organisations. Governance can see the prompt and the policy. It rarely sees the full behavioural effect of the transformation itself.

Gap 1 · The effect

The prompt is not the outcome

A request to change one property can move several. Teams review the instruction, not what the model actually did to the text.

Gap 2 · Self grading

The model should not mark itself

Model self evaluation is not enough for high consequence use. The system that produced the output should not be the sole judge of it.

Gap 3 · Verification

A policy is only as good as its check

"Change X, do not change Y" is useful only if somebody independently verifies whether that is what actually happened.

What Kronaxis does

Measure, compare, verify, report.

01

Measure

Score the source and the transformed output on frozen, auditable dimensions, fixed before any result is opened.

02

Compare

Quantify the intended movement and every material collateral movement, so what changed beyond the request is visible and measured, not assumed.

03

Verify

Use independent readers and simple baselines, not the generating model, and require agreement before a finding counts.

04

Report

Return a clear decision, with the evidence behind it and a practical remediation plan for governance, risk and the model owners.

The engagement

A ten day audit, one workflow, not a platform.

Kronaxis sits beside your existing AI workflow. You do not replace your model, orchestration or governance stack. We take a bounded sample, measure the transformation independently, and show where the system is controlled and where it is not. A first audit is typically five hundred to five thousand source and output pairs.

You provideOne AI workflow, the intended change, the changes that must not happen, and a representative sample of source and output pairs.
We returnA transformation movement matrix, a collateral drift register, an independent reader agreement summary, a policy exception register and a remediation plan.
PASS

Intended movement observed and collateral drift stays within the agreed bounds. Proceed with routine monitoring.

PASS WITH CONTROLS

The core transformation works, but defined drift or failure clusters need controls. Apply thresholds and guardrails, then retest.

FAIL

Material unintended change or a policy breach is systematic. Do not scale until it is remediated and re audited.

£7,500 to £15,000 + VAT

Fixed scope founding audit. The final fee is confirmed after a short scoping call and a sample check, and depends on dataset size and workflow complexity. A paid proof of value, not a commitment to replace your stack. If it reveals a repeatable control need, the next step is continuous verification.

Why Kronaxis

The measurement is the research, not a wrapper.

The published work separates measurement from controllability: a property can be easy to recognise yet hard to move on its own. The programme has repeatedly measured collateral drift across independent model families and domains, and its measurement instruments are being grounded against independent human judgements, with failures kept on the record rather than repaired after the fact. The compliance architecture separately distinguishes what can be formally enforced from what remains judgement.

We do not trust a single model's opinion. We require independent models to agree before we report a result, and we say plainly which findings are proven and which are judgement.

Boundaries, stated first

What this is, and what it is not.

What it is

An independent assurance engagement

An evidence backed decision on one workflow and a bounded sample: what moved, what drifted, whether it stayed inside your policy.

What it is not

Not a legal opinion or certification

It does not claim behaviour beyond the sampled workflow. Where a criterion is learned or judgement based, we report it as judgement, not proof. Formal or cryptographic components are described as proven only where the mechanism supports it. We do not claim to read minds or steer people.

A good first question

Give us one AI workflow. We will show you what it changed beyond what you asked.

The fastest way to a fixed scope and fee is a twenty minute call and a representative sample. One workflow, one defined risk question, one independent answer.

Start the conversation

Tell us about one AI workflow.

A short message is enough. A person replies, not a bot, and we confirm a fixed scope and fee after a short call.

Prefer email? Write to hello@kronaxis.co.uk.