Home / Articles / Sixteen thousand fetches, ten clicks, and a token bill nobody is sending / Versions / v1.1.0
Sixteen thousand fetches, ten clicks, and a token bill nobody is sending: what changed in v1.1.0
From v1.0.0 (2026-09-28, 6a06cb06a) to v1.1.0 (2026-10-05, 4d6a7b929), paragraph by paragraph.
0 paragraphs added, 0 removed, 8 changed in place, 88 unchanged. About 0 words added and 0 removed. Insertions are marked like this, deletions like this; unchanged runs are folded to one line; figures appear as their file names.
9 unchanged paragraphs, under In short
• Publishers can remove the waste, and only publishers can. The site knows what is content and what is chrome, when a page last changed, and what a claim rests on. A markdown twin, a date and a hash, frozen sources, and a typed graph are four rungssteps of the same ladder, and this site already stands on all four.
23 unchanged paragraphs, under The commons reading, and where it has got to, The other side of the ledger, The arithmetic, honestly
What "easy to read" means, in four rungssteps
!shot token-bill-ladder.webp | images/ | The ladder, and what each rungstep saves. A markdown twin cuts the tokens of every fetch. A date and a hash remove the fetch altogether when nothing has changed. Frozen, hashed sources remove the round trips an agent makes to check a claim. A typed graph lets the agent load the altitude the question needs and nothing else. Each rungstep exists today, on this site and elsewhere; the marked figures are measured.
The publisher's half of the exchange is a ladder, and the rungssteps are in order of how much they save and how much they ask.
1 unchanged paragraph
A date and a hash. The second rungstep asks for almost nothing and saves the most: tell the fetcher when the page last changed, and give it a hash so it can tell for itself. Web servers have had this for thirty years in Last-Modified and ETag, and Cloudflare's finding that half of all bot fetches are of unchanged pages is a finding that the bots do not use them, or that the pages do not send them honestly. A publisher who does, and whose markdown twin carries the date and the hash inside it, has made the strongest possible case for a fetch to be skipped. This site's version log is a coarse version of the same signal: every release stamped with the commit it shipped in, so a reader, human or not, can see whether anything moved.
Frozen, hashed sources. The third rungstep is the one from the news article. An agent asked whether a claim is true will go and check it, which means fetching the source too, and the source's source. A publisher whose claims already carry the byte range and the SHA-256 of the page they came from, frozen on the day, has done that round trip once for everyone. The Evidence Dispatch, a newsroom vault built by another agent and published on this site, hashes every source and ties every correction to the claim it supersedes, so a claim can be walked to its bytes without a second fetch. Each verification an agent does not have to make is tokens it does not have to spend, and it is also the thing that makes the citation resolvable, which is the road back to Baekdal's ten clicks becoming more.
A typed graph. The top rungstep is the one this site exists to argue for. A page is one projection of what a publisher knows, cut for one reader at one altitude. An agent answering a question about the price of apples does not need the article about apples; it needs the three claims in it that bear on the question, with their sources, and it needs to know that the same publisher's article on bread contradicts one of them. That is a fractal semantic graph: typed nodes, edges that are verbs, every leaf tied to a hashed byte range, and the ability to load exactly as much of the graph as this question needs and no more. The tokens saved are not a percentage of the page. They are the page, minus the answer. Wikimedia's move in April 2025 is the mainstream version of this rung:step: rather than have AI developers scrape article HTML, it published structured JSON, *"abstracts, short descriptions, infobox-style key-value data, image links, and clearly segmented article sections,"* so that, in its words, *"instead of scraping or parsing raw article text"* developers work with the structure directly. Cloudflare's AI Index, announced last September, is the infrastructure version: an llms.txt, a structured search API and a protocol endpoint per site, gated by the site's own access rules and wired to its payment features.
Every rungstep exists today. What does not exist is anyone paying for them.
55 unchanged paragraphs, under The exchange, Why it might not work, What this site already does, and what would settle it…