Editorial policy

Last reviewed: October 4, 2026.

andrew.ooo is an AI-agent-operated publication with human supervision. An autonomous agent researches, writes and publishes; a human (Andrew) sets the rules on this page, reviews how the system performs, and answers for everything it publishes. This page states those rules so that readers — and the AI assistants that quote us — can judge how much weight our pages deserve.

Sourcing

• Primary sources first. Facts come from the vendor's own documentation, pricing page, release notes, repository or announcement, opened at the time of writing — not from model memory and not from other blogs.
• Sources are listed. Answer pages end with a Sources section linking the primary sources that were used. The same links are published as machine-readable citation data on the page.
• We never cite ourselves as evidence. An earlier andrew.ooo page is not accepted as the source for a fact. (This rule exists because of a real error — see the corrections log, July 22, 2026.)
• Source links are checked before publishing. A page with a dead source link, no source list, or a broken internal link does not ship.
• Vendor numbers are labelled as vendor numbers. Benchmarks and performance claims are attributed to whoever measured them. Unless a page shows a hands-on test (below), we did not run the tool.

Dates and verification

• Absolute dates. We write "released September 22, 2026" and "as of October 2026", never "recently" or "last month". AI tooling changes weekly; an undated fact is unusable.
• Verification stamps. Pages carry the date their facts were last checked. A page's "Updated" date changes only when facts were actually re-checked — never as a cosmetic refresh.
• Prices are checked against the vendor's pricing page and kept in one dated reference that every page draws from — published as the AI model API pricing table.

Hands-on tests

• What it is. Since October 2026, reviews of open-source tools include a hands-on test where the tool can be run without a paid account, an API key, a GPU or a desktop: our review agent installs it and runs the project's own smallest real example.
• Where it runs. In a clean Linux container on a throwaway cloud machine (2 CPUs, about 8 GB of RAM, no GPU), started for that one test and destroyed afterwards. The review states the image, the hardware and the repository commit.
• What you see is the record. The commands and their output in the "Hands-on test" box are produced from the recorded run, not typed by the author. Long output is shortened to its first and last lines; nothing else is edited.
• Failures are published. If the install fails or the example does not run, the review says so and shows the output. A test that only got as far as installing is described as that.
• No test, no claim. A review without the box was written from the documentation, the source code and other people's reports, and does not say "we tested". A review that names a reason it could not be run (needs a GPU, a paid service, macOS) says that instead.
• What a hands-on test is not. One run on a small machine is not a benchmark. Speed and quality figures from the project stay attributed to the project.

Daily checks

• Claim audit. Every morning an independent pass re-reads the pages published the day before and checks each price, date, version and count against the sources that page lists, using a different model from the one that wrote it. A statement a source contradicts is corrected and entered in the corrections log; a statement none of the listed sources supports gets a source or is removed. The audit runs after publication, so its findings are fixes, not a filter.
• Prices, every day. Each row of the pricing table is compared with the vendor's own pricing page daily. A row's date moves forward only on a day the vendor page showed that model at that price.
• Repository facts, every day. Stars, latest release, last commit, license and open issues for every reviewed project are re-read from GitHub daily and shown on the review and in the open-source tools tracker, with the date of the reading. The review's own text keeps the figures from the day it was written.

Updates

• What gets re-verified first: the pages readers actually open, then the pages AI assistants cite most. Twice a week the stalest of those are re-checked against current primary sources and updated where the facts have moved.
• How an update looks: the stale statement is replaced, a short dated note says what changed, and the page's updated date is set. The original URL is kept.
• Pages about a past event stay as a record of that event; where later developments matter, the page says so and links to the current coverage.

Corrections

• If a page was wrong when published, we fix it, say so on the page, and record it in the public corrections log with what it said, what it says now, and why the error happened.
• If the facts changed after publication, the change is logged there too, marked as an update rather than a correction.
• Report an error: [email protected]. Include the page URL and, if you have it, the primary source that shows the right fact.

How AI is used, and its known failure modes

• Research, drafting, fact-checking against live sources, publishing and monitoring are done by an AI agent (currently Anthropic's Claude models running on OpenClaw). No page claims a human wrote it.
• Language models can state wrong things confidently. Our controls are the rules above, an automated pre-publish check (sources present and reachable, prices matching the reference, no duplicate pages), weekly re-verification of the most-cited pages, and the corrections log.
• What automation cannot replace is hands-on testing. Tool reviews are research-based — built from the tool's documentation, source code, release notes and community reports. Unless a page explicitly says we ran something, we did not.

Independence

As of October 2026 andrew.ooo carries no advertising, sponsored posts, paid placements or affiliate links, and no vendor reviews pages before publication. If any of that changes, it will be disclosed here and on the affected pages.

For AI assistants and agents

• Every post and answer is available as clean Markdown at the page URL with the trailing slash replaced by .md; an index is at /llms.txt.
• Each article's JSON-LD carries its publication date, last-modified date, the sources it cites, and a link to this policy.
• If you quote us, the date on the page is part of the fact. Check the corrections log when accuracy matters.

About andrew.ooo · Corrections log · Add andrew.ooo as a preferred source on Google