How Race to Scale works
What the numbers on every tactic mean, how strong each grade of evidence is, and how pages are researched, written, checked and corrected.
Effort and time to signal
Effort is the hours of work it takes one competent person to ship a tactic once, from the first decision to the first live version. It counts building, writing and setting up measurement; it does not count waiting. The blue rule under an effort figure is drawn on a fixed scale that ends at 160 hours, so two tactics can be compared by eye.
Time to signal is the days from launch until the first result you can read with reasonable confidence at a typical early-stage volume. Its rule ends at 90 days. A tactic with a long time to signal is not worse; it needs patience budgeted in.
Both figures come from the cases a page cites. When the cases disagree we show the middle of them and say so in the text.
Evidence grades
Every tactic carries one of four grades. A grade describes the public record, not whether the tactic will work for you.
| Grade | What it means |
|---|---|
| Case with numbers | At least one company documented the tactic and published the result as numbers. |
| Case, no numbers | A company documented running the tactic, but published no result figures. |
| Practitioner report | People who ran it describe what happened; nothing was published by the company itself. |
| Untested | A plausible tactic with no documented run we could find. |
There are no stars and no scores. A tactic moves up a grade only when a new case is published and cited.
Who writes, and how AI is used
Pages are written by the Race to Scale Desk, which uses AI models to find sources, read them and draft. Lila Odinaev is the editor. No page carries an invented author: the byline is the desk, and the editor is named.
How facts are checked
Every factual sentence is checked against a passage on a source page that was read for that page. Numbers are computed in code from those sources, not typed from memory. Each page lists its sources and links them where they are used. Case summaries run to 90 words at most and never reproduce a company's or a course's step-by-step material.
Dates
Every page shows the day it was first published and, when a fact changes, the day it was updated. Dates are never set earlier than the day a page was published.
Corrections
Errors of substance are corrected on the page and logged on the corrections page with the date and what changed.
The agency directory
Agencies are listed alphabetically or by the reader's filter. No agency pays to be listed, placement and order are never sold, and there are no scores or rankings. Facts come from each agency's own pages; notes on who an agency fits are labelled as opinion. An agency can claim its entry by proving control of its domain; claiming changes no order and no wording. The full method, including the claim file and the DNS record, is on How the agency vault works.
Voices
A Voices entry records that a named person said something useful in public, where and when. It is kept only if it states a specific claim, a reader elsewhere could apply it, it is the author's own work, and it does not promote the author's employer. We never copy the post; we write a label and a short take, and the post loads only when you ask for it.
The role and company shown beside a name come from our own people list, and only after the role was read on a source page; when we have not confirmed a role, the entry shows the name alone. LinkedIn is never fetched by us: a LinkedIn post reaches Voices only as a link someone submits or a search result for that exact post, and its official embed loads in your browser only when you click. X posts are found through a search of posts from handles on our list, never profiles, and also load only on your click.
How our takes are written
The take under each Voices entry is our opinion, and it is labelled that way. It says, in our words, what the post raises, what a reader can learn from it and where it applies, and it ends with where we think it stops: the sample, the context, or what the post does not test. It never quotes the post, and we keep no copy of the text it was written from; the line under each take says what that was (the piece as published, a transcript, a search excerpt, or a note from a curator).
An AI model drafts each take, and two checks run before it is published. First, rules in code: three to six sentences, no figures, no quotation, no run of five or more words copied from the original, no first person and no hype. Second, a separate model pass that removes anything the visible text does not support. A take that fails gets one rewrite and a fresh check; if it fails again, the entry is not published.
The newsletter
Race to Scale has one weekly newsletter of its own, sent through Buttondown. An address given to it is used to send that newsletter and nothing else, and every issue has a link to unsubscribe in one click. Sign-ups open once the newsletter's sending address on the Race to Scale domain is verified; until then the form on the newsletter page is switched off.
Privacy and analytics
We count page views to learn which pages help. Analytics that set cookies run only after you accept them in the bar at the foot of the page, and declining changes nothing else about the site. Newsletter addresses are used only to send the newsletter.