# Seed-size gate, all lanes (research, 16 Oct, #2603)
Data: 10 days only (6-16 Oct; v3live/mc5 history starts 6 Oct, not 30). 2515 mc5-checked pools (103 bought: live+watch), seed = PULL burn units (any fm_mirror log) else liquidity() now. Winner = bought mult>=2 or max mcap >=2x cap at check. Scripts: research/rug3/seed.py, an.py.
- 2138 of 2515 pools (85%) end pulled. 290 of 351 winners (83%) pulled later too. Pulls are the norm, not an exception.
- Gate "seed equals a size pulled >=2 times" (38 sizes): blocks 92% of pulls BUT 70% of winners (245/351) and 75% of winners that never pulled (46/61). Holdout (table from 1st half, tested on 2nd): pulls 84%, winners 58%. FAILS the <15% winners criterion. Sizes are launcher-kit signatures (63.2e21 x537 pools, 31.6e21 x336, 12.27e21 x232), not rug signatures.
- Bought lanes: 96 of 103 buys sit in a "pulled size": mean 1.53x, 7 winners; the 7 outside the table: mean 0.52x, 0 winners. So the size table would have removed almost nothing bad and the only profit. Do NOT hand to exec as a gate.
- Honeypot question: yes. Pull minus first outside buy (n=1966): p10 9s, p25 16s, median 44s, p75 340s, p90 4939s. Only 1 pool pulled before any outside buy. A quarter of pulls land within 16 s of the first outside buy, so any entry delay beyond ~10 s mostly sees a pool the deployer is about to pull; shadow with rugs=0x overstates live unless it models pulls at those timings (wlink 3 live: pulls 12/43/82 s after our buy).
- Better lever than a size table: exit/entry timing (sell side must beat ~16-44 s) and per-pool facts at buy time (LP owner, deployer history) - already tested negatively in rug-12oct.md.