Every page on this website — including this one — is AI-assisted content. We don't hide it. We publish under the same discipline we argue for in our Article 50 post: machine-readable markers, visible disclosure, human review, and traceable accountability.
This page explains what's AI-generated, what's human-reviewed, how to check, and how to tell us if something is wrong.
What's AI-generated on this site
Most of the prose you read on this site was drafted or substantially revised by Claude (Anthropic). Blog posts, marketing pages, and product descriptions all pass through AI before they reach human editorial review. Technical content (the glossary, architecture descriptions, pricing mechanics) is transcribed more directly from internal documentation, but the final copy is still AI-shaped.
The two exceptions:
- Quoted standards, regulations, or third-party claims — verbatim text from external sources is attributed and not paraphrased by AI.
- Numbers and facts tied to real-world records — pricing, dates, legal entity details, incident counts — these are read from source records, not generated. Where a statistic is inferred rather than directly measured, the confidence level is declared.
What "human-reviewed" means, concretely
Before any page here goes live, it passes through a four-gate editorial process. Every gate has an owner. Agents do the QA; humans write the body and the Founder approves.
G1 · Factual. Every claim is checked against documented evidence — the practice brain, architecture documents, contracts, records. Any claim that can't be traced to a source is downgraded or removed.
G2 · Voice. Copy is reviewed against the ORCA HQ tone-of-voice standard. Banned words removed. Punchline-first structure enforced. Honesty rule applied (limitations and failure modes acknowledged, not hidden).
G3 · Brand. Visual rendering checked against brand guidelines. Highlight yellow reserved for one phrase per section. No stock imagery, no AI-generated illustrations.
G4 · Founder approval. The Founder personally reviews on a gated preview URL and flips the Directus content item from status: review to status: published. No one else can publish.
None of these gates are theatre. Every one is a real decision point that has sent content back to the author for revision. The blog post you may have arrived from — "Your AI governance architecture IS your Art.50 strategy" — went through all four gates twice before it reached this page.
Machine-readable provenance
Art.50(2) of the EU AI Act calls for AI-generated output to be marked in machine-readable format. The Code of Practice expects both visible and invisible disclosure layers. Ours are:
Meta tags — every page carries:
<meta name="ai-content-declaration" content="ai-assisted-human-reviewed"> <meta name="ai-content-source" content="humanModifiedAlgorithmicMedia"> <meta name="ai-content-producer" content="Claude (Anthropic)"> <meta name="ai-content-review" content="human-editorial + Founder approval">
JSON-LD — a Schema.org CreativeWork block on every page declares the creator, producer, and a link to the IPTC digital-source-type vocabulary term humanModifiedAlgorithmicMedia. This is the form automated crawlers, fact-checkers, and AI provenance tools can read without running JavaScript or parsing prose.
Visible disclosure — the line at the bottom of every page ("AI-assisted content · Human-reviewed · Founder-approved") is the human-visible counterpart. It links back here.
Confidence levels on claims
Our governance platform distinguishes five confidence classes for any piece of knowledge: Definitive (primary source), Grounded (documented evidence), Inferred (AI-generated reasoning), Advisory (weaker evidence), and Insufficient (not yet verified).
On this website we apply the same taxonomy informally: marketing prose is Grounded (based on documented product state); future-tense roadmap language is Advisory; statistics from third parties are marked with their source; anything we can't verify doesn't get published.
If you spot a claim on this site you think is wrongly graded, we want to hear. (See below.)
How to report an inaccuracy
Use the contact form with the URL and the specific claim you're questioning. We treat corrections seriously, respond within two working days, and — if the correction stands — update both the content and the corresponding entry in our practice brain so the same error doesn't recur.
We don't consider this optional. A platform whose pitch is "governed intelligence" cannot be casual about the accuracy of its own public surface.
Why we do this
Article 50 formalises something we think is already the right thing to do. The regulatory pressure is helpful — it makes the discussion specific — but the principle precedes it: if AI generated the content, the reader has a right to know, and the author has a duty to take responsibility anyway.
Visible markers close the first half. Human review and named accountability close the second. We apply both to ourselves because we ask our customers to apply them to their own AI-assisted work. "Do as we do" is a more honest pitch than "do as we say."