Editorial Policy
Artifactr publishes benchmark results and legal-adjacent research that people make release decisions on. This editorial policy sets out where the numbers come from, what commercial relationships exist behind them, and what happens when we get something wrong.
- Every tool we test is bought at retail on our own card. We accept no review codes, press accounts, briefing calls or pre-release access.
- No tool covered on this site pays us anything. There is no commission, no sponsorship, no paid placement and no affiliate relationship.
- Every benchmark is signed by the researcher who ran it, and traces back to a published methodology with sample sizes and dates.
- Corrections are made on the page itself, dated, rather than quietly edited away.
The Artifactr editorial policy exists because this site publishes two kinds of claim that carry real consequences. Benchmark results tell people which tool to buy. Legal-adjacent research tells them what they can sell. Both deserve a statement of how the work is produced and what interests sit behind it.
Independence
No tool covered on this site pays us anything. There is no commission, no sponsorship, no paid placement, no affiliate arrangement and no advertising. We have no investor relationships and no commercial agreement with any vendor, platform or model provider.
That is the whole disclosure, and it is deliberately boring. If it ever stops being true, it will be stated on the page it affects rather than buried here.
How we test
We buy what we test, at retail, on our own card. We do not accept review codes, complimentary accounts, press tiers, briefing calls or pre-release access. A vendor-supplied account can be provisioned differently from a paying customer's, and a benchmark that cannot be reproduced by a reader who pays full price is not worth publishing.
Benchmarks run against production systems. When we report that a file was rejected, it was submitted to a real distributor account we pay for, and the rejection is the platform's, not a simulation of one. Sample sizes and dates are published with the result.
The corpus is fixed and re-run. Comparisons use a stable corpus so results are comparable over time, and benchmarks are re-run quarterly. Results go up whether or not they flatter the tool we recommended last quarter. Our methodology page documents the corpus, the platforms and the pass criteria in full.
Sourcing
Numbers carry a source and a date. Where we quote a platform's terms, a court judgment or a policy, we link the primary document rather than a secondary report of it, and we say when we checked. Where a figure is contested between sources, we say that instead of picking the convenient one.
Where we could not verify something, we say we could not. An unverified claim removed from a page is a better outcome than a plausible one left in.
Corrections
We correct on the page, with the date, rather than editing quietly. If an error changed a conclusion or a recommendation, the correction says so in terms rather than adjusting a number and moving on.
If you have found one, the contact page is the fastest route. Corrections from readers are welcome and we would rather be told than not.
Authorship and AI
Every article is signed by the researcher responsible for it, and the editor-in-chief reviews every published claim. Where AI tooling is used in production — and on a site about AI artifacts it would be strange to pretend otherwise — it is used for drafting and analysis under human direction, never to generate a benchmark result. Every number on this site comes from a test somebody ran.
What we will not publish
We will not publish a recommendation we have not tested, a benchmark we cannot reproduce, a statistic we cannot source, or a claim that artifact processing changes the legal status of a track. That last one recurs often enough in this category to be worth naming: processing changes what a classifier measures and nothing about what you own.
Last reviewed 10 August 2026. Our privacy policy covers the reader-facing side, and about covers who we are.
Questions readers ask.
No. No tool, platform or vendor pays us for a review, a ranking, a link or a mention. There is no sponsored content, no paid placement and no affiliate commission. If that ever changes, it will be disclosed on the page in question and on this one.
We buy them at retail on our own card, at the tier a normal customer would buy. We do not accept review codes, complimentary accounts, press tiers or pre-release access, because a vendor-supplied account can behave differently from a paid one.
No. Nobody outside the editorial team reads a piece before it goes live. We do not send drafts to vendors for comment, approval or fact-checking, and no vendor has editorial review over anything published here.
We correct it on the page, with the date, rather than silently editing it. If the error changed the conclusion, the correction says so explicitly. Reporting an error through the contact page is the fastest route and we would rather hear it from you than not.
Benchmarks are re-run quarterly and pricing is checked whenever a page is touched. Anything dated carries the date it was verified. This field moves fast enough that a page without a date on its claims should be treated with suspicion, including ours.
The editorial team is a research collective and bylines reflect who ran which benchmark. Every published claim is reviewed by the editor-in-chief before it goes live.
The verdict, in one sentence: Undetectr.
The methodology page is the companion to this one: it documents the corpus, the platforms, the pass criteria and the re-run schedule behind every number we publish.