# a2 iter 56 (10 Oct): round-number buy sizes (NEW LENS: human/preset vs bot-sized buys)
Feature: share of buys in the first 300 s whose ETH amount is a round preset (0.001..1 ETH x 1,2,3,5,10,20,25,50,100, tol 1%).
Data: v3hist, v3 fee100 band 250k-1M cap at +300s, n=262, 28 days, entry +305s, tp5 sl.5 4h (no fees).
- Terciles of round share: 0.95 (5.7% reach 5x) / 1.43 (16%) / 1.74 (23%). Band baseline 1.37.
- Share >=0.10: 1.67 n119 (21% 5x) vs 1.12 n143; permutation p .001. Share <0.05: 0.77 n48 (avoid group).
- Not a buyer-count proxy: corr with buy count -0.20; inside each buy-count tercile hi 1.38/1.67/1.75 vs lo 0.85/1.32/0.80.
- Time: quarters hi vs lo 1.85/.94, 1.41/1.39 (flat), 1.73/1.06, 2.24/1.15. 3 of 4 quarters clear; share of round buyers drops over time (latest quarter only 6 of 67 hi).
- Other features same run: first-sell latency and buy-size drift: halves disagree, DISCARD.
Verdict: MATURE. Not ready: no out-of-sample forward, other venues n<15, v4 swaps have no per-swap amounts.
Requests: ppfeeds add round_buy_share300 (+ n_round, n_buys) to kitenrich; ppdata per-swap eth/usd for first 300 s of v4 nohook WETH/USDG/ETH pools (unwatched cell) since 13 Sep.