LH003 survey table

The two tables below are the deliverable LH F04's LH003 design section and the dispatch prompt both ask for: one row per candidate observatory, in the column order the dispatch prompt gave (source, layer seen, population, boundary, unit, cadence, licence, what it cannot see, S number or URL), and a second table for candidates considered and excluded, with the reason. Version 0.2 added two columns to the included table, publication date and observation window, beside a last column that now gives each row's date read (see LH003.md's Corrections).

Layer names follow Lighthouse_Research_Design.md (LH F02): activity is computation happening, residue is the altered structure and state it leaves behind, propagation is residue enabling later activity elsewhere and is not itself a surface. F02's "Outcomes" row sits outside the three layers and scores activity and residue against a declared purpose; a few candidates below fit there better than in any of the three, and the layer cell says so rather than forcing a fit. Where a source spans more than one surface, the cell names the primary one and notes the other.

Every cell that could not be confirmed from a source read on the date below is left blank or marked "not stated", per AGENTS.md; nothing here is invented. A fuller note follows the table for any row whose cells need a qualification too long to fit them honestly. Every row was read on 26 September 2026, and the last column says so row by row; where a row also cites an LH F02 source (S number), F02's own check of that link was made on 25 September 2026. Publication date is the date a source's page gives for the version read; observation window is the period of activity its data covers, as the source states it. Both are filled only where the page, as this survey recorded it, states them; elsewhere the cell is blank, with a note in brackets saying why. Only the Anthropic Economic Index page was read again for version 0.2. Version 0.3 takes its changes to the Cloudflare AI Crawl Control, Open Source Insights and Anthropic Economic Index rows from notes/R-0001.md, whose reviewer read those pages as raw HTML on 26 September 2026; no page was fetched again for it.

Included: observatories with a row

Notes on rows above

GH Archive. LH002's own design section already names the qualification that matters most for this survey: event records alone do not give reliable AI attribution, deployment records, or every code diff.

Vulnerability databases. F02 already records that nvd.nist.gov "needs scripts to render" for its S15 entry; this study's own attempt at both nvd.nist.gov and cve.org on 26 September 2026 hit the same wall (a bare page header, no body text a fetch tool could read), so the licence and exact update cadence for both are left blank rather than guessed. The CNA/NVD split (who assigns an ID versus who scores it) came from a WebSearch summary, not a direct page read, and should be checked against a primary CVE Program page before a study relies on it.

Anthropic Economic Index. Version 0.1 gave this row's scope as Free and Pro only without saying which report the scope belonged to. The page read is the introductory report, which its page dates "Feb 10, 2025" and which states: "We also only analyze data from Claude.ai Free and Pro plans, rather than API, Team, or Enterprise users." notes/R-0001.md's raw read of the page on 26 September 2026 confirmed both. That read also found three errors carried from version 0.1: the page says "we don't argue that the uses in our dataset are a representative sample of AI use in general", not the words earlier versions put in quotation marks; the privacy floor of 15 conversations or 5 accounts is not on the page, so it is unconfirmed and may come from the accompanying paper, which this survey has not read; and image generation is absent because Claude cannot generate images, not excluded by the index. The earlier reads, like version 0.2's, used a fetch tool that returns a model-written extract of a page rather than the page itself. The page gives no date range for the conversations it analysed, so the observation-window cell is blank. That later reports widened the scope is the claim of the first review note, notes/review-2026-09-26-first-articles.md, which says later reports include first-party API traffic and that the June 2026 report separates chat, Cowork and API usage; it gives no URL, and this survey has read no later report, so the row describes the introductory report only.

Cloudflare AI Crawl Control. Versions 0.1 and 0.2 listed "the purpose behind a crawl (training vs a live agent fetch)" among what this feature cannot see. notes/R-0001.md read the page as raw HTML on 26 September 2026 (last updated 23 April 2026) and found that it lets a site owner group AI-crawler traffic by crawler, by operator and by "Category: Analyze crawlers by their purpose or type" (observation). The feature therefore attributes traffic to AI operators, for crawlers Cloudflare has identified, and sorts it by each crawler's stated purpose. What it still cannot say is why any single page was fetched, because the category belongs to the crawler, not to the request (interpretation, following the review). No fetch was made for this correction; the page wording above is the review's.

OpenAI, "How People Use ChatGPT." Every figure in this row traces to a WebSearch summary of press coverage and the paper's own abstract, not a direct read of the NBER PDF (which this study's fetch tool could not parse as text) or of the companion openai.com page (which returned HTTP 403). The "10% of the global adult population" figure in particular is press commentary on the paper, not a number this study confirmed in the primary text; treat it as a secondary claim, not a derived measurement, until someone reads the primary source.

Hugging Face Hub. Only the models-the-hub page was fetched directly, and its returned content was too thin to source the download-concentration figures against; those came from a WebSearch summary citing a Hugging Face blog post and a download-analytics doc page, neither fetched directly here. A future check should read those pages directly before citing the 85.6%/80% figures as more than a secondhand report.

GitHub Security Advisories and npm download-count API. Both rows above were confirmed by a direct page fetch after an earlier WebSearch-only pass; they are accordingly more reliable than the OpenAI and Hugging Face rows, which rest on search summaries alone.

Excluded: candidates considered and not given a row

Every fetch attempted for this study, 26 September 2026

Recorded here so a later check can see what was tried and what failed, independent of which row it fed.

Several rows above also draw on WebSearch result summaries rather than a direct page fetch (the OpenAI paper's population figures, the GHSA licence claim before the direct-fetch confirmation, the Hugging Face download-concentration figures, the Shodan/Censys and Stack Exchange exclusion reasons, the CNA/NVD split). Each such case is flagged in its row or note above; a WebSearch summary is a secondary read of the underlying page, not a primary one, and is recorded as such rather than presented as equivalent to a direct fetch.