Citelift consumer citation repeatability evidence
=================================================

This standalone bundle contains the public-safe records used to calculate the
four dated ChatGPT and Claude summary rows in Citelift's consumer-browser study.
It does not contain raw answer text, excluded attempts, customer data,
credentials, or expanded destinations hidden behind collapsed citation groups.

Requirements
------------

Python 3.9 or newer. The checker uses only Python's standard library and makes
no network calls.

Verify the published aggregate
------------------------------

After extracting the ZIP, change into the extracted directory and run exactly:

  python3 resources/consumer-citation-repeatability-study/reproduce.py --root . --check

Expected output:

  verified 4 separate wave/interface rows from 96 public run records

Run the same command without --check to print the independently recalculated
CSV. Pair-level Jaccard values are recomputed from shared_urls / union_urls in
the included repeat-overlap.csv files; the checker does not average the rounded
jaccard display column.

Included records
----------------

- resources/consumer-citation-pilot-2026-09-09/{runs,visible-links,repeat-overlap}.csv
- resources/consumer-citation-wave-2026-09-14/{runs,visible-links,repeat-overlap}.csv
- resources/consumer-citation-repeatability-study/study-summary.csv
- resources/consumer-citation-repeatability-study/reproduce.py

Scope
-----

The 9 September and 14 September collections remain separate dated convenience
samples. ChatGPT and Claude remain separate consumer interfaces. Do not pool the
four rows or treat differences as a controlled trend or platform ranking.

Citelift funded and published the study. Siddesh Patil was the single observer;
Codex assisted with offline analysis.
