Trust, AI & model risk

The scoring is deterministic. The model narrates — it does not decide.

The moment you say “AI” to a regulated institution, model risk management engages. This page is written for that conversation.

For model risk management

Five properties your MRM function will want to test.

The number that reaches your board is not produced by a language model. It is produced by a deterministic engine that an independent reviewer can re-run and reproduce.

01

Deterministic and reproducible

Findings, severity, maturity levels and the sequenced roadmap are computed by rules, not generated. The same evidence produces the same score every time, so the engine can be independently validated.

02

A closed evidence set

The model is handed a bounded set of your own evidence and is not permitted to reach outside it. It composes language over facts it was given; it never supplies the facts.

03

Cited, then verified

Every statement has to cite the evidence it rests on, and a verification pass checks those citations after generation. Anything that fails is withheld rather than shown.

04

Zero retention, enforced by refusal

Answers are composed against an endpoint that retains nothing. If that guarantee is not explicitly confirmed in the configuration, the platform refuses to answer rather than send your estate somewhere that might keep it.

05

No training on your data

There is no fine-tuning on client data and no cross-client learning. The model is stateless between requests; nothing an answer touched survives it.

06

A human approves what executes

A rule extracted from one of your policy documents is a proposal. It never runs until a named reviewer approves it, and the approval records who that was.

The usual concern about AI in a regulated process is that it invents. Here the model cannot invent a finding, a severity or a score — it is not the thing that produces them.
Your policies, as enforceable rules

What the platform can execute is constrained by design.

A policy document becomes a rule in a restricted, declarative grammar — not free-form code and not free-form queries.

A closed grammar

Rules are expressed in an allow-listed vocabulary and compiled into safe, parameterised checks. Nothing outside that vocabulary is representable, which is what makes machine-extracted policy safe to run at all.

Proposal, review, then execution

Extraction produces a proposal. A named human reviewer approves it before it can execute, and the approval is part of the record.

Precedence made explicit

Where an internal standard and a statutory obligation disagree, the conflict is surfaced with the resolution guidance attached — rather than being silently resolved in favour of whichever ran last.

Fit with what you already run

FluenBox evidences the data feeding your reports. It does not replace your GRC platform.

The question every regulated buyer asks: where does this sit next to my regulatory reporting chain and my GRC tool?

Next to your GRC platform

Your GRC tool holds the control framework and the risk register. FluenBox holds the evidence that the data underneath those controls is measured, governed and current — and complements it rather than competing with it.

Underneath your regulatory reporting

Your reporting chain produces the submission. FluenBox evidences the data products that submission is built from — per product, per critical data element, on a date.

Above your catalogue and quality tooling

The catalogue, the quality engine and the orchestrator all keep running and keep their jobs. FluenBox reads from them and measures what they produce.

Deployment posture

Including estates that never touch the internet.

Self-hosted and air-gapped

FluenBox runs fully isolated in your own environment. Nothing is fetched at run time; the regulatory and framework knowledge it reasons over is stored locally and is self-contained.

Enterprise identity

Single sign-on against your own identity provider, with strict separation between workspaces.

Licensed content stays yours

Where a framework is commercially licensed, the platform ships its own synthesis of the structure rather than the licensed text. If you hold a licence, you load that content into your own instance.

Bring your model risk and security teams to the pilot.

Thirty days on one slice of your estate is also thirty days for your second line to test every claim on this page.