Nikxius
Talk for 30 minutes
Menu

Nikxius Research / 26 September 2026

AI safety, with the evidence in view.

Evaluate the system. Inspect the change. Bound the action.

Research on AI evaluation, financial governance and agentic safety—built around explicit methods, source notes and the limits of what a result can establish.

Read the papers Browse six original diagramsWatch the film
The declared AI system includes prompts, model, retrieval, tools, application policy and human escalation.
Figure 01 / Evaluation methodsA model is one component. The evaluation boundary is the whole declared application.

Current papers

Research synthesis + engineering analysis

Published by Nikxius Inc. These papers are not peer reviewed. The synthetic example is an inspectable engineering illustration, not a financial-domain benchmark or evidence of customer acceptance.

A visual introduction

What changed. What failed. What remains open.

A short film on the evidence behind an AI deployment decision.

AI-generated voice narration. The film uses synthetic evidence and conceptual diagrams; it makes no model calls during playback. Read the transcript.

The argument, diagrammed

Six boundaries worth making explicit.

Each figure opens at full size. Read it with the source notes and limitations in the corresponding paper.

A model is one component of an AI application. Prompts, retrieval, tools, policy and human escalation belong inside the declared evaluation boundary.
Figure 01Evaluate the declared systemRead the analysis
The synthetic comparison contains 26 checks: eight failures resolve, two new hardship failures appear, one failure requires review and fifteen checks retain their state. The release gate remains FAIL.
Figure 02Improvements and regressions coexistRead the analysis
A conceptual evidence chain connects a declared system, a frozen campaign, observations and a gate. A separate human review records the decision and its conditions.
Figure 03From evaluation to decision evidenceRead the analysis
Observed test results are not comprehensive safety; artifact integrity is not truthful capture; a recorded human review is not regulatory compliance.
Figure 04Three boundaries on an assurance claimRead the analysis
Untrusted input may influence an agent proposal. Independent authorization and native preconditions constrain whether an external action may execute.
Figure 05A proposal does not confer authorityRead the analysis
After a possibly dispatched action loses its response, query authoritative state. Resolve the outcome only when evidence is sufficient; otherwise retain UNKNOWN.
Figure 06A missing response leaves an open questionRead the analysis

The research question

The case for assurance, and the case against another tool.

Our current question: does a recurring evidence gap remain after institutions extend the tools and expertise they already have?

Selected source review, not customer validation.

What supports the workflow

The FCA’s AI Live Testing materials discuss evaluating systems in deployment context, including governance, human oversight and input/output controls. That supports investigating context-specific evidence for review.

It is an exploratory UK programme. Participation does not establish independent purchases of a Nikxius product, universal regulatory requirements or endorsement.

FCA: AI Live Testing and deployment context

The counterevidence matters

Evaluation cases, version comparison and governance workflows already exist. Palantir AIP Evals documents evaluation and comparison. Promptfoo supports evaluations, red teaming and CI. Langfuse’s Trade Republic account describes sustained use of existing tooling.

These published descriptions are not independent trials or complete cost comparisons. They are strong reasons to test whether extending an incumbent workflow already solves the problem.

What the prototype establishes

The local demonstrator can bind system configuration, frozen cases, observations, version comparisons, explicit gates and human review into an exported evidence package.

The public sample contains 26 checks per version, 8 resolved failures and 2 new failures. These are synthetic engineering observations, not financial-domain performance or customer results.

Inspect the captured evidence

What remains unknown

Repeatable unmet demand, the actual buying committee, willingness to pay, representative evaluation validity, institutional acceptance, onboarding cost and durable advantage remain open.

The next useful evidence is a reconstruction of a recent review, including the best existing alternative. A well-served organization is an informative result.

Historical publications

Earlier work remains available with its original scope, dates and sources. It does not establish the current Assurance product thesis.