The ZeruAI article “Any bot detector can flag any percentage you like” presents a powerful test: ask whether the same detector returns the same percentage on a group of accounts with no suspicion of bot activity. If it does, you’re reading a ranking, not a count. I applied this test to @zerufinance’s own data: Their example: A group of 1,981 wallets, with 618 (31.2%) flagged as hard bots, median transactions of 1,886. Manual review of 5 wallets: 3 were genuine bots, 2 were humans (one power user, one non-violating airdrop farmer). Limitations of the example: A sample of 5 wallets is sufficient for illustration but insufficient to estimate false positive rates. The article only reports the median for flagged wallets, not for the unflagged group, so the claim “highest activity” requires a comparative baseline. Campaign metrics from the whitepaper: 56% reduction in Sybil allocation, 49% increase in high-quality wallet participation, 50% reduction in post-distribution selling pressure. Allocation to low-quality wallets dropped from 21.0% to 13.5%; high-quality wallets increased from 41.7% to 54.7%; average holding time rose from 9.8 to 12.6 days. Applying the test: The footnote states this is internal analysis, with customer names and parameters possibly obscured, and compares against “equivalent” campaigns without countermeasures. Who labeled accounts as “Sybil” or “low quality,” using what methodology, and how were the control campaigns selected? If labels are assigned by the filter itself, then improvements are measured by the same filter. The metrics also use multiple measures (percentage changes, weightings, days), so denominators are essential. Strengths: The whitepaper acknowledges its own limitations: insufficient data coverage, model drift over time, adversary adaptation, and clustering by on-chain funding source being circumvented via multi-chain funding. Penalty severity is tiered, so a single false flag only forfeits a portion of rewards, and decision thresholds reside with the allocator, not the oracle. A scoring system that includes its own validation test is a good sign. The rest requires publishing the labels and control groups so outsiders can reproduce the test.