// Kedion — independent scientific review of AI

What you've been told about your AI and what's actually true are not the same thing.

Independent scientific review of AI in the enterprise — from executive briefings to line-by-line system audits. Whatever depth the situation calls for, the job is the same: close the distance between what your organization believes and what is real.

The same gap, at three depths.

There is a distance between what an organization believes about its AI and what is true. It shows up at every level — and closing it is the work.

Depth 01 — the boardroom

Your executives

…are making decisions from a picture of AI assembled by the people selling it. Vendor decks and press coverage are not a working model of the technology.

Depth 02 — the team

Your teams

…are using tools without a working model of what those tools actually do — so they over-trust them in some places and under-use them in others.

Depth 03 — the system

Your systems

…may be silently wrong in ways standard QA does not catch, because catching them requires reasoning about the behavior of something you cannot inspect line by line.

Down into the mess, then up into a category.

The same sequence closes the gap at any depth. The move that matters is the last one: taking a specific, tangled problem and lifting it into a class of system whose properties are already known.

01

Trace the actual chain of events.

02

Find the step that is the bottleneck.

03

Lift it into a general class of system.

04

Fix it using what's already known about that class.

The method as a diagram A causal chain of four steps runs left to right. The third step is highlighted as the bottleneck. A dashed line lifts from the bottleneck up to a box labelled a known class of system. a known class of system behavior already understood input step ◆ bottleneck the step output
The method, drawn once — trace the chain, isolate the one step that is the bottleneck, then lift it into a class of system whose behavior is already known, so it can be reasoned about and fixed with known tools.

A shipping product
that was quietly wrong.

A commercial generative music system — built on real mathematics, running in production, passing all 373 of its tests. An independent review found six silent defects: the mathematics was implemented correctly, and in six places the shipped operating point sat outside the regime where that mathematics means anything.

The full writeup describes each defect — what the code did, what it should have done, what the user actually experienced, and why nothing caught it — then lifts the six into the general class they belong to.

Read the full diagnosis
The system

A working, shipping generative music system — correct mathematics, under a green suite of 373 tests.

The finding

Six silent defects — the code was right, the operating point was wrong. Output no listener could hear: a constant tone, a dropped note, a click.

The lift

Not six accidents but one class — a model shipped outside the regime its own mathematics assumes. That class predicts where to look in other systems.

Photo of Sean McClure
→ /assets/sean.jpg (4:5)

I'm a scientist. Scientific training is training in a specific thing: taking a system you didn't build, that nobody can fully inspect, and determining what it is actually doing rather than what it was meant to do.

That is exactly the problem an enterprise AI system presents. I come at it from the outside, with no stake in the answer being reassuring. Two things that buys, which a delivery vendor cannot claim: classification — putting a system into a category whose behavior is already understood — and independence, which is the whole premise of an audit.

Sean McClure — PhD, computational chemistry.
Prior enterprise delivery at Ford and CSX via Trace3.
Author; work published on arXiv.

The same method, as tools.

Notes2Tree turns raw material — notes, documents, transcripts — into a hierarchy you can navigate. Dekyon layers that structure onto a body of content so it can be learned from. Not a product line; the practice, compiled.

Find out what's actually true.

Book a diagnostic call. Tell me the situation — a briefing for the executive team, a review of a roadmap, or a production system you suspect is quietly wrong — and I'll tell you what a review would look at.

Work is done under NDA by default, and anything published is de-identified.