Editorial policy

Last updated: 5 October 2026

PriceProof is produced by a data pipeline with a language model writing part of the prose. That is unusual enough that it deserves a plain explanation rather than a footnote, and it is set out here in full. Google's guidance penalises hiding automated production, not using it — and a reader has a right to know how a page was made.

Who is responsible

The publisher is Lars Cornelissen, named with a biography on the about page. He is the person accountable for everything published here, including the text written by the model, and the person who decides what is published.

Responsibility does not transfer to the software because software wrote some words. Every page published on this site is published deliberately, under a named publisher, who answers for it.

How AI is used — and the four things it is allowed to do

The pipeline computes the evidence first. Figures are extracted from public award records by code, checked, and written to an evidence file for each product. Only then does a language model see anything, and what it sees is that product's evidence file and nothing else.

Within that boundary the model is permitted to do exactly four things:

Write the opening sentences of a product page, from the computed figures.
Write the "what to watch at renewal" guidance for that product.
Write the FAQ answers, each citing the awards it draws on.
Rewrite a sentence that failed a verification check — at most twice.

That boundary applies to product pages. It is widened in exactly one place, the newsroom, and only in the way described in the next section.

What AI is never allowed to do

It does not choose which products are published or rejected. Code does that, against a fixed threshold.
It does not choose which awards appear. Code does that.
It does not introduce a number, vendor, agency, product or award that is not already present in the evidence file it was given — or, in the newsroom, in a fact it was given that has been verified against a fetched public source.
It does not adjust, round, estimate or interpolate a figure. Numbers are copied or they are wrong.
It does not decide whether a page passes verification.

The newsroom

The newsroom carries two kinds of piece, and the page tells you which is which.

Most are written by a person and signed by them. The rest are written and published by the PriceProof Records Desk — an automated desk belonging to the pipeline, not a person. Those pieces carry a desk byline because there is no person to sign them, and a named editor who did not write the piece would be a worse answer than no name at all.

Every desk piece is published without human review. Nobody reads it before it appears. This is deliberate. It is stated here, and in the methodology, because you are entitled to know it before you read a desk piece rather than after. Every desk piece is marked as one at the top of its own page, under its desk byline, and lists what it was built from at the end.

A desk piece is triggered by a new award in the public record that cleared a notability bar fixed in advance — a large obligation, a sole offer, a first purchase by an agency, a new contracting vehicle. A quiet week produces nothing.

The desk researches the agency, the recipient and the product from public sources: the award record itself, the agency's and the recipient's own published records, the vendor's public pages, government and press coverage, and reference works. Sources in every one of those classes must be public and citable.

How a fact survives. For every claim it draws from the web, the pipeline requires the drafting model to return the verbatim quotation it is relying on and the page that quotation came from. Code then fetches that page and checks that the quotation really appears in it. A fact whose quotation cannot be found in the page is discarded, and a number that does not appear inside the quotation is discarded with it. The model never writes prose in that step, so it cannot phrase its way past the check — a fabricated fact is simply dropped, and the article is written without it.

What a desk piece may never do.

It may not present a contract total as the price of a product, a rate, or a cost per seat. A contract total is what one agency was awarded for one scope of work. This rule holds for desk pieces exactly as it holds for product pages.
It may not carry a figure that traces neither to the award's evidence nor to a cited source.
It may not be published on fewer than two researched documents. The award record itself does not count toward that two. The floor is on documents rather than on publishers, so it can be met by two pages of one vendor's own site — which is sometimes all the public record holds for a single award. Every document the piece used is listed at the end of it, so you can see for yourself how independent the sourcing is.
It may not repeat a product page or a previous desk piece. A piece that largely restates an existing page is refused and not published.
It may not be signed by a person, or carry an editor or reviewer field.

There is no way for a vendor to influence a desk piece. No vendor sees one before publication, none is asked for comment as a condition of publication, and none may commission, suppress or alter one. A vendor objection is handled as a correction report and nothing more.

Because a desk piece is unreviewed, mistakes are more likely in the prose than in the figures. The figures are machine-checked against the record before the build will publish; the prose is checked for support, length and originality, but not for judgement. If a desk piece is wrong, the correction process below applies, and we would rather withdraw a piece than defend it.

How the writing is verified, mechanically

Every number appearing in model-written prose is extracted and matched against the evidence file for that product. An unmatched number fails the check, and the sentence is rewritten. If it fails again, the sentence is dropped. A page does not build with an unverified number in it — the build stops.

Every qualitative sentence is checked against the awards it cites, on a deterministic measure; sentences below the threshold are rewritten and then dropped if they still fail. Claims that cannot be supported by the named awards do not survive to publication.

The consequence of this design is that the model's contribution is style, not substance. The substance — every figure, every award, every agency, every date — is computed by code from public records, and the model is fenced inside it.

Where the data comes from

All award data comes from USAspending.gov, the official public source for US federal spending, maintained by the Treasury Department. Every award figure on a product page links to its record on that site so a reader can verify it directly.

We use only public records. We do not use paywalled analyst research, leaked or confidential vendor price lists, information covered by a non-disclosure agreement, scraped private data, or anything a vendor tells us privately. Vendor list prices, where shown, are read from the vendor's own public pricing page or published pricing document, and the quote and capture date are recorded.

Verification and freshness

Each product page carries the date it was last verified. Award data is refreshed periodically by re-querying the source; figures are recomputed rather than patched by hand. When a figure changes, the change is recorded and appears on what changed, with the date and the record behind it.

List prices are re-checked against the vendor page when the pipeline runs, and a list price whose source page no longer resolves is withdrawn rather than left standing on an old number.

Corrections

Anyone can report an error to [email protected]. The process is:

We identify the specific published figure and the specific record it claims to come from.
We compare the two, ourselves, against the source — not against a preference about what the figure should be.
If the figure is wrong, it is corrected at the source of the computation, so that every page drawing on that record is corrected too, not merely the page where the error was noticed.
Material corrections are noted in the affected page's change history.

Corrections are treated as the highest-priority work on the site. A publication whose entire claim is that its figures are traceable has to be fast and ungrudging when a trace fails.

Independence and funding

PriceProof is independent of the vendors it writes about. It accepts no payment, sponsorship, affiliate commission or free product from any vendor mentioned, and no vendor has any influence over what is published, how a product is described, or which figures appear. Vendor objections to a figure are treated as a correction report and nothing more: if the record supports the figure, the figure stands, and we will explain the method but not negotiate the number.

The site is funded by advertising, which is disclosed on the privacy policy page along with how advertising cookies work and how to opt out. Advertising is served by a third-party network and has no bearing on editorial content. Advertisers do not see pages before publication and cannot commission, suppress or alter an article.

Standards we hold ourselves to

Every figure traces to a record. No figure appears on a page without a matching entry in the evidence file, and no page is published below the evidence threshold.
A contract total is never presented as a price. Every figure is labelled by type, and different types are not compared with one another.
Nothing is published that fails a gate. The pipeline fails closed: a product without enough evidence gets no page and a recorded reason, not a thin page.
No page is filler. Every product page contains its own award rows — facts that appear on no other page — and must carry substantial original reporting above a minimum length.
The method is public. The methodology page documents the codes, thresholds and matching rules well enough that a reader could reproduce the work. Its limits are documented there too.

Contact

Editorial questions, corrections and complaints: [email protected].