Prompt for retro-classifying pre-incident drain reports

Wicked announces that he is running a language model over publicly reported COLDCARD losses that predate the disclosure, to estimate how many were seed collisions rather than ordinary theft, and posts four screenshots of the prompt itself rather than a conclusion. The prompt sets firmware inclusion criteria at roughly 2^50 to 2^63 for Mk3 and Mk4 entropy, tells the model to exclude losses attributable to phishing, supply-chain compromise, malware or physical access, and asks for a per-report classification, a probability comparison against accidental collision, a sensitivity analysis and a confidence score. It is held for the method: the archive already carries contested claims that owners were drained before July 2026, in btctherapist-prior-drain-report and reddit-ledgerwallet-drain-comment, and this is a dated attempt to sort them systematically with the criteria published in advance. No result is contained in this post, and the estimates the prompt asks for would be the model's output, not a finding verified here.