Shared the outputs from initial analysis for questions to @Clement_Goubert for review.

Spreadsheet (internal): https://docs.google.com/spreadsheets/d/1bylF4s2FOo-W5F-Dhfb-RuiOk-dI-sAN5gvJtjEmBLo/edit?usp=sharing

It has the following outputs:

• Overall clients and requests that'd be blocked
• Blocks by residential proxy presence, and counts by individual proxy networks.
• Blocks by spider vs. user categorization.
• Blocks by file (image, image_thumbnail, image_3d, video, audio and documents)
• Blocks by browser likelines
  • bucketed scores: negative < 0, low < 30, medium < 100, and high >= 100
  • individual x-is-browser score
• Clients that would have been most hit by the limits (top 20)

For each breakdown, we have the three proposed limits: 300K kbytes/minute, 600K kbytes/minute, and 1200K kbytes/minute with how many clients would have been blocked and how many requests would have been blocked.

Data period: 13 July to 9 August 2026 (4 weeks) of webrequest logs to upload.wikimedia.org and thumb.wikimedia.org

Let me know if there are any following up questions or there are other dimensions of interest.

Additional parameters for a new run:

• Correlate with KAPOW scores to see if interesting information pops up
• Correlate with known_client to see if we need to implement known_client ratelimit bypass
• Estimate bandwidth savings avg(response_size)*blocked_requests, possibly estimate over X-Is-Browser and top clients