Editorial Standards & Sourcing Policy
How Capital & Compute researches, sources, fact-checks, reviews, and corrects its work, and exactly how and where we use AI.
Capital & Compute covers money and compute, subjects where a wrong number or a stale claim has real consequences. These standards describe how we work so you can judge our work.
What this site publishes
The beat is the economics of AI infrastructure, across six connected areas: what models and coding agents cost per finished task, whether running inference locally beats renting it, what benchmark scores mean once you leave the test set, the data centers and capital behind the models, the memory and GPU supply constraining them, and the markets being repriced by all of it. Coverage extends to adjacent finance topics where the same discipline applies.
What this site does not publish: price predictions presented as forecasts, investment recommendations, sponsored articles written by anyone other than the editorial team, or rankings influenced by payment. Roughly half the site is not prose at all but maintained datasets, and those are held to the same standard as the writing.
How we research
- Primary sources first. Pricing, model specifications, and policy claims are taken from official documentation, provider pages, and filings, not from secondary summaries.
- We test what we can. Where a claim about an agent harness, a local model, or a tool can be verified hands-on, we verify it rather than relying on marketing.
- Numbers are dated. Prices and benchmarks move quarterly. We note when a figure was accurate and revisit posts when the underlying numbers change.
Sourcing and attribution
We link to primary sources so you can check our work, citing them at the point of each claim. We represent every source's status accurately, distinguishing a peer-reviewed study from a preprint from vendor documentation, and we never overstate what a source establishes. Posts built on research or statistics carry a full list of sources at the foot. We quote others fairly and with attribution, and we distinguish clearly between reporting, analysis, and opinion.
In practice that means a set of rules applied to every post. Each external factual claim carries a link at the point the claim is made, with anchor text that names the source rather than reading "click here" or "a study shows". The producer, the nature of the document and the year are stated in the prose, so a claim is legible without clicking and attributable by an AI answer engine that quotes it. A preprint is called a preprint. A vendor's own documentation is called documentation, never an independent finding. Secondary coverage is attributed as secondary. An institutional affiliation is never asserted unless it is verified from the source itself.
Citation links are ordinary followed links. Pointing readers and search engines at the sources a piece rests on is the whole point of citing them, so those links are never marked nofollow to preserve ranking signal. The only links here that carry nofollow are third-party utility links we cannot vouch for, such as the advertising opt-out links in the privacy policy. Any compensated link, were one to appear, would be marked sponsored.
How the trackers and datasets are built
The price trackers, benchmark directory, model registry and provider comparisons are built from dated datasets rather than screenshots. Each figure is verified against a primary source, most often a provider's own rate card, an official leaderboard, a regulatory filing or a published index, and each carries the date it was checked. Every tracker states its as-of date on the page.
Two limits are stated rather than hidden. A tracker is a dated snapshot, not a live feed, so a reading can be stale by the time it is read and anything time-critical should be confirmed at its source. And where a figure is modeled rather than measured, such as a cost per completed task derived from published rates, the assumptions behind the model are stated so the arithmetic can be checked or disputed. A modeled number is never presented as an observed one.
How we treat vendor claims and benchmark scores
A score in a launch announcement is a claim by an interested party, and it is reported as one. Where an independent leaderboard publishes a figure for the same model, the independent figure is the one recorded, and a self-reported number that has not been corroborated is labeled unverified rather than quietly adopted. Evaluation conditions travel with the number: the benchmark version, the split, the reasoning effort setting and whether tools were available all move results by more than the gap between most models, so a figure quoted without them is treated as unusable.
How we use AI
We write about AI tools and we use them. To be transparent: AI assistants may help with research, drafting, and editing, but every published piece is reviewed and fact-checked by a human editor who is accountable for its accuracy. We do not publish unreviewed AI-generated text, and we do not fabricate experience, sources, or data.
Independence and conflicts
Editorial decisions are independent of advertising. Advertisers and affiliate partners have no input into what we cover or what we conclude. Any affiliate relationship that could affect a recommendation is disclosed inline; see our disclaimer and disclosures.
Concretely: sponsors buy visibility and never buy rank. No position in any ranking, comparison, or directory is for sale, and no payment creates or improves a listing. The full statement of what is and is not available commercially is on our sponsorship page, and directory entries are submitted and corrected for free via our submissions page.
Corrections and updates
Errors are fixed promptly and openly. Two things are distinguished. A correctionis a factual error: a wrong number, a misattributed source, a claim that was not true when published. Corrections are noted on the piece rather than silently edited away, and where one materially changes a conclusion, the change is described. An update is new information on a piece that was accurate when written, such as a price that has since moved; updated posts carry a revised date.
Reports of errors are read first and are welcome even when they are inconvenient. Email contact@capitalandcompute.net with the page and the specific claim, and a source if you have one.
Using our data
The datasets behind the trackers are served as JSON at stable URLs so figures can be cited or rebuilt rather than scraped, and an llms.txt describes the site for AI answer engines. Quoting a figure with attribution and a link back is welcome. When you do, please carry the as-of date with the number, for the same reason we do.
Who is accountable
Capital & Compute publishes under its brand rather than individual bylines, but a human editorial team stands behind every post and is reachable at the address above. Read more on our about page.