# Exhibit thirteen — fetch receipts The measured fetch evidence behind movements one and three: what the major publishers serve, how big the largest `llms-full.txt` is, and what thirty days of our own server logs say about who asks for these paths. **What this file is.** An extract from the exhibit's recon pass (`recon/inventory-history.md`), assembled for publication because the page points at this evidence and a pointer should reach something. Every figure, table row and quoted limit below is carried over unchanged from that document. What was left out: the recon pass's editorial section on what the article may and may not claim, its findings about unrelated estate systems, and the specific container and log-volume names of the serving host — infrastructure detail that is not evidence for anything on the page. The method is described without them. Nothing here was re-measured for publication. Where the recon pass marked a limit or an inconclusive result, the mark is carried over too. --- ## 1. What the major publishers serve Live verification, fetched **2026-08-16**, all primary — our own `curl`. | Property | `/llms.txt` | bytes | notes | |---|---|---|---| | docs.anthropic.com | **200** | 58,752 | also `llms-full.txt` **200 — 32,161,074 bytes** | | docs.stripe.com | 200 | 89,885 | | | vercel.com | 200 | 211,285 | | | mintlify.com/docs | 200 | 54,662 | | | developers.cloudflare.com | 200 | 15,714 | emits `rel="alternate" type="text/markdown"` | | docs.perplexity.ai | 200 | 42,838 | Perplexity *publishes* one | | docs.mistral.ai | 200 | 14,658 | | | docs.github.com | 200 | 28,674 | served as `text/markdown` | | developers.openai.com | 200 | 5,853 | | | cookbook.openai.com | 200 | 19,591 | | | llmstxt.org | 200 | 637 | the spec's own file | | www.answer.ai | 200 | 760 | | | openai.com | **403** | — | Cloudflare bot block — **inconclusive**, not evidence of absence | | platform.openai.com | 404 | — | | | ai.google.dev | 404 | — | | | developers.google.com | 404 | — | | | gemini.google.com | 404 | — | | | google.com | 404 | — | | | huggingface.co | 404 | — | | ### 1.1 The `llms-full.txt` size receipt `docs.anthropic.com/llms-full.txt` is **32,161,074 bytes ≈ 30.7 MiB ≈ ~8.0 million tokens** at the four-bytes-per-token rule of thumb. Verified by `HEAD` on **2026-08-16** (`content-type: text/plain`, after a 301→308 redirect chain). The byte count is measured. The token figure is an **estimate** from the four-bytes-per-token rule of thumb, and is the only token figure on the exhibit page not produced by a tokenizer — the page says so where it prints it, and every dollar figure derived from it inherits the same error bars. ### 1.2 One correction to the spec's own adoption claim llmstxt.org's v2 page claims OpenAI, Anthropic *and* Gemini publish their own files. Anthropic and OpenAI check out (OpenAI on the `developers.` and `cookbook.` subdomains, not the apex). **Google does not, at any of the five obvious paths tested above.** Stated as "not found at the obvious paths" rather than "false" — a file could live somewhere untested — but the spec's own adoption claim does not survive a spot check. --- ## 2. Method and window for the log counts The estate's reverse proxy writes one JSON-lines access log per vhost. They were read **read-only**; nothing was written or rotated. **Retention is set to `roll_keep_for 720h` = exactly 30 days.** No rotated files are present (none has reached the roll size), so each log begins at its own creation or at the retention edge: | Log | First entry | |---|---| | apex (strata2signal.com) | 2026-07-17 04:21:45Z (at the 30-day edge) | | rulesage | 2026-07-17 05:37:48Z | | amble | 2026-07-24 04:57:31Z | | **research** (this site) | **2026-08-11 21:30:39Z** — recreated, so only a 5-day window | Total logged volume ~91 MB across **20 vhost logs**. **The exposure caveat, stated up front.** The estate's `llms.txt` files were deployed **2026-08-15 06:03:54 GMT** (their `Last-Modified`; the research file re-written 19:12:47Z the same day). At the time of this pass they had been live **~31.6 hours**. Any statement of the form "crawlers did not fetch *our* file" is therefore weak. The strong finding in §4 is a different statement, and it holds across the full thirty days: **crawlers never asked for the path at all — not even speculatively, not even to receive a 404.** --- ## 3. Every `llms*` request in the window 183 request rows across all vhosts, spanning **2026-07-21 07:46:24Z → 2026-08-16 13:37:12Z** (165 × `/llms.txt`, 18 × `/llms-full.txt`). By user agent: | Count | User agent | What it is | |---|---|---| | 118 | `curl/8.18.0` | **ours** — this recon pass and earlier rollout verification | | 17 | `coherence-pass/1.0 (estate audit; …)` | **ours** | | 15 | `…claude-code-estate-truth-pass/1.0 (+read-only audit)` | **ours** | | 15 | `claude-code-estate-truth-pass/1.0` | **ours** | | 1 | `s2s-site-pass/1.0` | **ours** | | 4 | `…compatible; BuiltWith/1.4; …` | commercial tech-stack profiler | | 1 | `Mozilla/5.0 (compatible; Dataprovider.com)` | commercial data profiler | | 12 | assorted generic Chrome / Safari / Firefox / Android UAs | mostly datacenter IPs (AWS, Azure, Alibaba, OVH) | **166 of the 183 rows — 91% — are the estate's own agents and probes.** Strip those and **17 genuinely external requests remain in 26 days**, none of them from an AI system. The only *identifiable* external requesters are **BuiltWith** and **Dataprovider.com** — tech-profiling and lead-generation crawlers, not model crawlers. This matches Ahrefs exactly: SEO/audit tooling is the dominant consumer, AI assistants a rounding error. Four of the external rows are `502` against the bare host name where no vhost is configured — internet background noise probing a well-known path. Evidence that *something* out there speculatively looks for this file, just never a model crawler. --- ## 4. The headline: crawler traffic vs `llms.txt` requests Counted case-insensitively over all 20 vhost logs, full 30-day window. | Crawler | Total requests | `/robots.txt` | **`llms*`** | |---|---:|---:|---:| | GPTBot | 2,334 | 1 | **0** | | Amazonbot | 2,150 | 64 | **0** | | Meta-ExternalAgent | 2,113 | 0 | **0** | | Googlebot | 1,301 | 257 | **0** | | OAI-SearchBot | 690 | 429 | **0** | | ChatGPT-User | 609 | 5 | **0** | | **ClaudeBot** | **594** | **161** | **0** | | PerplexityBot | 255 | 0 | **0** | | MistralAI-User | 233 | 0 | **0** | | Google-Extended | 214 | 0 | **0** | | Applebot | 98 | 4 | **0** | | CCBot (Common Crawl) | 95 | 0 | **0** | | bingbot | 94 | 35 | **0** | | Bytespider | 54 | 0 | **0** | | Claude-User | 32 | 0 | **0** | | Diffbot | 12 | 0 | **0** | | **Total** | **~10,878** | **956** | **0** | Also present: SemrushBot 128, AhrefsBot 55, BuiltWith 38, Dataprovider 13. Returning zero across the whole window: Omgili, ImagesiftBot, AI2Bot, Timpibot, DuckAssistBot. **All 16 named crawlers above: 10,878 requests, 956 of them `robots.txt`, 0 `llms.txt`.** Excluding the two classic search crawlers (Googlebot 1,301 / bingbot 94), the **AI-associated crawlers alone made 9,483 requests, 664 of them `robots.txt`, and 0 `llms.txt`.** The exhibit page leads with the AI-only framing and prints the wider one beside it; a re-measure must state which framing it used and reuse the same rule. The ClaudeBot line, as the recon pass put it: > ClaudeBot asked our servers for `robots.txt` 161 times in 30 days. It asked for > `llms.txt` zero times. **Verification performed** — the UA strings are real, not substring artefacts. Sampled actual log lines for the ambiguous matches: - `Mozilla/5.0 (compatible; Google-Extended/1.0; +http://www.google.com/bot.html)` — genuine, and notable: Google-Extended is usually described as a robots.txt *policy token*, yet it fetches here. - `Mozilla/5.0 (compatible; PerplexityBot/1.0; +https://perplexity.ai/perplexitybot)` - `…Chrome/145.0.0.0 Safari/537.36 (compatible; meta-externalagent/1.1 (+http…` - `CCBot/2.0 (https://commoncrawl.org/faq/)` - `Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; MistralAI-User/1.0` **Honest limits on this count:** - Counts are line matches on the UA substring; one request per line. A UA string appearing in a referrer or URL would inflate a count. Spot checks found none, but the totals are "requests whose log line contains this token", not a verified UA-field parse. - **User agents are trivially forgeable.** These counts measure claimed identity. A cross-check against published IP ranges would harden it; not done here. - `robots.txt` counts are per-vhost, and the estate has **13 vhosts with no robots.txt**, so many of those 956 requests received 404s. That does not weaken the finding — it strengthens it: these crawlers keep asking for a file that is often not there, and still never ask for `llms.txt`. - The raw access logs are **not published**: they hold visitors' IP addresses. --- ## 5. Measured again after publication — an interim recount **Added 2026-08-17.** Window: **2026-08-16 13:37:12Z** (the last row of §3's count) → **2026-08-17 23:28:19Z**, ~34 hours. Same twenty vhost logs, read-only; same rule as §3: every `llms*` request row, the estate's own agents and probes stripped first (`curl/8.18.0`, `coherence-pass`, `claude-code-estate-truth-pass`, `s2s-site-pass`). 199 rows landed in the window; 136 were the estate's own (this workshop re-verifies its pages constantly, and counts itself out the same way §3 did). **63 genuinely external — against 17 in the entire 26-day window above:** | class | count | note | |---|---:|---| | browser user-agents | 37 | ordinary Chrome / Safari / Firefox readers | | Meta-ExternalAgent | 24 | this path's first AI-crawler fetches ever logged here | | Googlebot | 2 | | | ClaudeBot / GPTBot / OAI-SearchBot / ChatGPT-User / PerplexityBot / Amazonbot | 0 | unchanged from §4 | 62 of the 63 landed on research.strata2signal.com, 1 on the apex. Status: 59 × 200, 3 × 304 (conditional revisits), 1 × 404. No theory of cause is offered: the dates are printed and the reader can line them up against the page's own publication date. This is an interim reading; the registered 30-day re-measure keeps its original schedule.