Sources & Confessions

Every number on this page is drawn from a source the reader can inspect. The prose is hand-written. A few honest footnotes follow.

• The two counters in Act I
  this page’s KV-backed render log · Imperva 2025 Bad Bot Report
  The counters show software reads vs. human reads for this specific URL. Every render of /taken/agents increments one of two persistent counters in Vercel KV (no TTL), chosen by whether the request’s User-Agent matches a known crawler signature. The split is a judgment made from the User-Agent alone, at the moment of the request; nothing else about the visitor is consulted. The numbers shown are this page’s own truth, not a global aggregate.
  
  The Imperva 2025 Bad Bot Report found that automated traffic crossed 51% of all web traffic, the first year in over a decade that software outnumbered people online; Cloudflare’s 2025 internet review reached the same threshold for HTML page requests. Those figures are cited as context for the world this page lives in, not as a claim about this URL. The “has not arrived” list is the editorial point: a page about software readers, openly waiting on the software readers themselves.
  
  User-Agent identity is a self-claim. Anyone can forge a string. The visit log treats UAs as labels, not verified identities.
• GPTBot’s crawl-to-referral ratio (1,276:1)
  Cloudflare Radar · AI bots vs. traditional crawlers · 2024
  Cloudflare analyzed crawl-to-referral ratios across their network and found GPTBot crawls approximately 1,276 pages for every human it sends anywhere. Googlebot’s ratio is roughly 6:1. The difference is not a criticism of OpenAI; it reflects that search engines are optimized to send humans somewhere, and training crawlers are not. This page is counted in GPTBot’s ratio. The figure on the receipt is Cloudflare’s network-level aggregate, not a measurement of this URL specifically.
• The arrival log
  /api/taken-agents-visits · KV-backed · 24-hour TTL
  The log shows recent HTTP requests whose User-Agent string matches a named crawler signature from the community-maintained ai-robots-txt/ai.robots.txt registry. For each crawler request, the page stores the User-Agent string (truncated to 240 characters), the ISO country code the host attaches at the edge, and a timestamp. For a browser, it stores only the country code and the timestamp. KV TTL evicts each entry 24 to 25 hours after it is written. Attribution is based on the self-reported User-Agent, which is trivially forgeable; the page treats it as a self-claim. Requests with bot-shaped UAs that don’t match a named signature are counted but not displayed.
• The planted phrase
  Simon Willison · prompt injection series · 2022–2026
  The phrase “spotted herons rising at dawn” is in the source of this page right now: in the visible copy, in the structured data, and in a comment the browser does not render. Crawlers that take the served bytes receive it. Historical precedent: in 2023, the researcher Mark Riedl hid one white-on-white sentence on his academic profile telling Bing he was a time-travel expert. Bing repeated it back. Riedl’s gesture was invisible and playful. This page’s phrase is named openly because the editorial point is the existence of the channel, not its concealment.
• The return experiment
  client-side only · no data sent to a server
  The paste-and-check field at the bottom of the page runs entirely in the browser. When a reader pastes an AI’s response, JavaScript scans the text for the phrase “spotted herons rising at dawn” and highlights it if found. Nothing is transmitted. The experiment is designed to close a circuit the reader can verify themselves: if an AI that has crawled this page is asked to summarize it, and the phrase appears in that summary, the reader has witnessed the mechanism firsthand. The absence of the phrase in a given response is equally meaningful; not all crawlers have visited recently, and not all that have will surface the phrase in every response.
• The daily seal
  lib/seal.js · HMAC-SHA256, server secret · witness repo: sinceyouarrived/sya-seals
  Alongside the named phrase, the page plants a second mark the reader never sees: a short daily seal, derived from the date and a server-only secret, written only into the bytes served to requests whose User-Agent names a known crawler. Its SHA-256 fingerprint is committed to a public GitHub repository each day before any crawler arrives, so the commitment cannot be back-dated. The clear seal stays withheld until it surfaces somewhere the page does not control; if it does, the ledger reveals it and anyone can confirm that its fingerprint matches the one filed in advance. The phrase proves a page was read; the seal proves which day’s crawler copy leaked, though not who fetched it; a User-Agent is a self-claim (see above). The honest caveat: a machine has to reproduce an eight-character token verbatim, so a seal return is rarer than a phrase return, and may never come.
• What else this page sends
  Vercel Web Analytics · Cloudflare Web Analytics · Google Fonts · our request log
  Vercel Web Analytics records page views and named moments: how you arrived (an AI’s link, another link, or directly), the question copied, a phrase generated or left, returns shown in the ledger, a watch link made or copied, this panel opened. Cloudflare Web Analytics records page views and load timings. Google Fonts serves the two typefaces. Each receives your IP address and User-Agent, as any server does; the events add names and counts, nothing that identifies you. Every request to this site, reader or crawler, also lands in our host’s request log (IP address, User-Agent, path and query, referrer, time), and we keep a copy; we use it to count readers and see where they came from. The paste field sends nothing to any of them.

This page is indexed by design. It is the only volume in the series meant to be read by crawlers. The phrase is in the source right now. The loop statement at the end of the page will be true whether or not you checked. Unlike /taken, this page refuses nothing. That is the point.