Trust

On the record.

Depth is only worth having if you can rely on it. These are the mechanisms that make an answer checkable, stated as behaviour rather than assurance.

01 · The register

Eight mechanisms. No adjectives.

01Every figure is calculated from your records, the way your finance team would calculate it. The intelligence decides what to examine; the arithmetic comes from the data itself.
02Every fact traces back through the record to the file, and to the span of the file it was read from.
03In a generated report, every figure is pulled from the underlying record as the page is built. What you read is what your data says.
04Every fact carries when it was true and when we learned it, so the brain can answer what it believed six months ago.
05Nothing is overwritten. A correction is a new assertion and the history survives it.
06Provenance stays visibly separate: your reviewed truth, an agent's private context, and external research are never blended.
07When there is no record, the brain says so and reports the absence as a gap.
08Every run, every model call, every cost and every action is metered and audited, across Kadura Meridian, the connector and any tooling built on your brain.

Example record

What is the current lead time on this component?

No record.

no source covers thisnearest related fact 14 Nov 2025ranked 3rd by impact in gaps

02 · Deployment

Your data. Your walls.

Meridian runs where your data is allowed to be. What it needs from the environment underneath is written down and checked against each of these shapes, so whether it can run inside your walls is answered before anyone starts an investigation.

The same holds when the brain is reached from Claude or ChatGPT through the connector. Every request runs as the person who asked, under exactly the permissions and document clearance they hold in Meridian, and the log records who called what and when, never the question.

Managed by usThe default. We run it, you use it.
Your own cloudInside your own cloud environment, whichever provider you use, including where it has to reach into your private network. The database never gets a public endpoint, and access is granted by identity rather than by a key held by anyone else.
Your own hardwareEvery piece of infrastructure on machines you control. Only model inference leaves the box.
HybridThe brain in one place, extraction wherever the data lives. A site reads its own documents, archives and feeds, with local models where sensitivity demands it, and only the distilled record travels to the brain.
Fully air-gappedNothing egresses at all, with models running inside your own walls.

03 · Models

Models are plumbing.

All model traffic passes through one internal layer, and no part of the application talks to a provider directly. Cloud models, models taken directly from the labs, open models, your own custom deployments, or a mix. Which one handles a piece of work is decided by how sensitive the data is and where the system runs, never by lock-in.

04 · Isolation

Your brain is yours alone.

Every brain is its own isolated deployment: its own database, its own storage, its own credential, its own graph. There is no shared store and nothing joining one customer to another. We do not train models on your data, and it is never pooled with another business.

05 · The benchmark

What the structure is worth.

For anyone who wants to interrogate the machinery rather than take it on trust. Benchmarked internally against roughly eighty million transactional rows of real client data. Both numbers come from the structure, not from a better model.

40% → 100%

Correct answers. A frontier model pointed straight at the raw database gets six in ten wrong, and you cannot tell which six. On the same questions, the brain got every one right.

200s → 17s

Mean time to answer. Speed comes from knowing where to look rather than reading everything.

80m rows

The corpus it was measured against. Real operational data, not a synthetic set.

06 · Discipline

Designed in, not bolted on.

Every consequential action has a person in itThe brain drafts and routes for approval. Nothing leaves your company until someone has approved it.
Judgement stays with your peopleKey-person risk is modelled at the level of roles and relationships, never as monitoring of individuals.
Your systems stay authoritativeThey own what they own. The brain composes across them with consistent meaning.
Explanations carry their confidenceWhere the data supports more than one explanation, you see both.
Truth is reviewedNothing becomes part of the brain without passing review, and what an agent holds privately stays private to it.

Ask us the hard version of any of this.