The Standard.

Every study, white paper, and cookbook on this site is produced under one written standard: claims trace to sources, numbers trace to models, confidence is graded rather than implied, and corrections are made in the open. This page is that standard, published so readers can hold the work to it.

The Method

Map · Model · Build · Publish

Map: Every piece starts as a question with a scope, not a topic. The evidence sweep covers regulation, standards, peer-reviewed work, technical documentation, and filings before it touches commentary. What the sweep cannot support gets named as a gap, not papered over.
Model: Load-bearing numbers earn their place: a cost model, a calculation that can be redone, or a documented assumption with its source and access date. A claim that cannot survive its own arithmetic does not ship.
Build: Where the paper would normally stop, the studio builds the first useful version: runnable code, a harness, an evidence run. Cookbooks ship their complete source beside the article so the reader can run what the text describes.
Publish: Full text, openly, with publication and revision dates on the page. Citations trace load-bearing claims to their sources. Nothing sits behind a paywall or a form.

Confidence Grades

Claims are graded, not implied

A research site that never distinguishes fact from inference is asking to be trusted on tone. Load-bearing claims in this archive are held to a five-grade scale, and revision moves a claim up or down the scale in public.

Primary evidence: Supported directly by regulation, standards, technical documentation, filings, or measured results.
Corroborated: Supported by multiple independent, credible secondary sources.
Reasoned inference: A conclusion argued from available evidence. The reasoning is on the page and can be attacked.
Early signal: Weak or partial evidence. Flagged as such, held loosely, revisited on revision.
Thesis: A speculative position the studio is willing to be wrong about in public.

The Source Register

In order of authority

Sources are weighed by type before they are weighed by claim. A vendor page and a regulation are not the same kind of evidence, and the writing does not treat them as interchangeable.

Regulation & law: Directives, acts, and their implementing texts.
Standards: ISO, IEC, IEEE, and industry standards bodies.
Peer-reviewed work: Journals and refereed conference proceedings.
Technical documentation: Datasheets, reference manuals, and engineering specifications.
Vendor material: Product pages and announcements - used with their bias priced in.
Market & press reporting: Named outlets only, never the sole support for a load-bearing claim.
Original analysis: The studio's own models, calculations, and evidence runs - published with the piece.

The Pipeline

One person, operating a research desk

This archive publishes at a rate one person could not sustain by hand — and does not pretend otherwise.

Lyon Industries runs an agent-assisted research pipeline. Frontier AI systems, operating under written research charters, do the sweep: literature, datasheets, standards, filings, and pricing, recorded against sources at access time. They draft evidence maps, build first-pass models, and run the harnesses that generate reproducible artefacts.

The judgment stays human. Every load-bearing claim is verified against its source before publication, every model's assumptions are reviewed, and every piece is edited and signed by Christopher Lyon. Where verification fails, the claim is cut or downgraded — the archive has build tooling that refuses to publish work that breaks these rules.

The tooling sets the pace. The standard sets the bar. Treating the pipeline as something to hide would itself fail the standard, so it is documented here instead: this site is also a working demonstration of what one person can operate when the machinery is built properly.

Corrections & Revisions

Made in the open

Every article carries its publication date, and a revision date once it has one. Material corrections change the revision date; nothing is silently patched. Claims that age badly are revisited and regraded rather than quietly removed.

If you find an error — in a number, a citation, or a claim — say so. Corrections are acknowledged, credited when wanted, and folded into the page with a dated revision.

Editorial standard

“If a peer can verify it in an afternoon, it is worth publishing. If they cannot, it is not finished yet.”

Report an error About the author Read the studies →