Study 1: observed optimization signals on cited pages
CC BY 4.0 for released data and prose. Analysis programs: MIT.

The public scalar capture omits response bodies, redirect-chain/location metadata
and free-text errors. One failed final destination with a sensitive query-key
name is blanked; the analysis already blanks those public URLs. No signal
used to assign a classification or substantive aggregate is removed. The file
manifest distinguishes original capture/table hashes from public LF-normalized
copies. The frozen collection instruments retain their original bytes.

Offline reproduction, Python standard library:
  python reproduce_study1.py .
This checks each classification and every numeric aggregate table against the
released scalar classification and citation-membership tables, including
Wilson limits, citation order, schema types and host-file observations.

Rebuild from the captured scalar JSONL, without fetching any page:
1. Obtain answers-2026-09.jsonl, release-2026-09.json and the control-question
   input from version DOI 10.5281/zenodo.22907695. Place answers and release in
   inputs/22907695/ beside the frozen receipt. Preserve the deposited bytes.
   The controls input is the separately published control-questions.json here.
2. Keep collect_page_signals.py, page-signal-rules.json, study1-news-hosts.csv
   and study1-precollection-freeze.json beside the analysis program, unchanged.
3. Create your own local excluded-hosts.txt for domains outside your collection
   permissions. The published capture has no client-domain URL examples. A
   private exclusion list is not distributed. An empty file reproduces the
   already released, public URL projection; this is not permission to fetch.
4. Run:
   python analyze_page_signals.py --answers inputs/22907695/answers-2026-09.jsonl --release inputs/22907695/release-2026-09.json --controls control-questions.json --signals study1-page-signals.jsonl --rules page-signal-rules.json --freeze study1-precollection-freeze.json --exclude-hosts excluded-hosts.txt --out reproduced
5. Compare CSV cells, not platform-dependent CRLF versus LF bytes. The summary
   records the supplied scalar projection hash; original_capture_sha256 in the
   provenance file identifies the fuller captured metadata record.

The collector is supplied for inspection as the exact frozen research instrument.
A new live collection measures a new date and is not a reproduction of this
snapshot. It requires a locally supplied host exclusion policy, the public OSF
availability check, robots compliance and per-host delays. Never distribute
the collected response bodies. The exact .py fingerprint in the frozen receipt
is checked before analysis alongside the news list, rules and answer inputs.

Interpretation: static signatures are not evidence of optimization intent or
that a signal caused an AI citation. Failed fetches remain separate. Intervals
describe this dependent convenience archive, not a population of buyer searches.