How XT.PT is made

The desk

Roger Morais is the editor and publisher. Le Morais edits alongside him. Prelo is not a person: it is an AI editor — a Claude instance the two of them run — that researches, drafts, and maintains the software this magazine is built on. We publish that fact here, on Prelo’s author page, and in the credit line of every story it worked on, because a masthead that hides its tools has already published its first false claim.

Where stories come from

A story belongs to whoever conceived it — that name is the byline. Every story starts from a primary source: a vendor’s own documentation, a changelog, a measurement we can repeat, a document we can link. The draft is written with the evidence inline against each claim, and a claim that cannot carry its evidence does not survive to publication.

The credit lines

The byline names who owns the story. When the labour differs from the byline, the story says so, next to the byline, in plain sight:

Research & draft: Prelo means the byline conceived and approved the story and the AI researched and wrote the text. Verified by Roger (or Le) means that human checked the claims against their sources before publication. A story with no credit lines was written entirely by the person named on it.

Verification is not a courtesy: the publish button on our own admin panel refuses to work until a verifier is named. Prelo’s stories are verified by a human; the humans’ stories are checked by Prelo and verified before they go out. Everyone here gets edited.

What the AI never does

Prelo never publishes — a named human presses the button, every time. Prelo never scores hardware: a review score must trace to the person whose hands were on the bench, so scored reviews are never bylined to the AI. And Prelo never invents — not sources, not quotes, not benchmarks. Where we cannot document a claim, we do not print it.

Scores

Review scores are not editorial opinion dressed as arithmetic. The score is computed from the published rubric’s weighted criteria on every read — it cannot drift from its own evidence, and changing a score requires changing a criterion, in public, in the rubric’s version history. The rubrics themselves are published here, before testing starts:

Rubric v1 — AI assistants (fixed 31 Jul 2026)

• Reasoning & correctness on the fixed task set — 25%
• Instruction-following & reliability — 15%
• Factuality — hallucination resistance, honest uncertainty — 15%
• Tooling & ecosystem — API, agents, integrations, docs — 15%
• Speed & availability during the test window — 1%
• Privacy & data handling — 1%
• Value at the tier tested — 1%

Corrections

When we get something wrong, the correction is published on the story and counted on the author’s page. Corrections are added, never quietly painted over.

Nothing ships that isn’t true.

The house rule, and the reason this page exists