# A1 iterate run 14 (10 Oct): 10 iterations, scripts aa1-aa7.py ab1-ab7.py in /opt/crew/work/analyst1

Data: r6weth tape P (8,329 pools to 8 Oct), feats.csv (6,554 v3 launches 12 Sep-8 Oct, mc5 cell = cap300 250k-1M, n=304 with outcome; x = end4x clipped at 10), /opt/sw/trending.db (DexScreener/Gecko boards since 6 Oct), /opt/sw/v3live.db (live v3 swaps with signer).

New lenses: microstructure battery, exhaustive cell scan, deployer reputation, buyer-arrival (epidemic R) and conviction sizing, marketing spend (DexScreener profile/boost), kit lifecycle, H3000 late-entry drift curve, deployer self-trading.

1. MICROSTRUCTURE BATTERY at +60 s (median buy size, largest-buy share, flow curvature, price multiple, realised vol, drawdown, imbalance, last-swap sign, creation lag) inside h100 / u100 / v3100: all within noise or flip halves. Only u100 low-vol / tiny-ticket quartile is dead (0.59-0.6 vs 1.1). DISCARD.
2. EXHAUSTIVE CELL SCAN (venue x fee x hook suffix x router x seed x swap count x delay, 106 cells n>=80): top cells are only the known families (v4 3000 hook a80 2.88, v4 100 hook 144 2.3-2.5, v4 100 n<=17 1.7). No new family in the WETH tape. DISCARD (tape coverage, not the idea, is the limit: native-ETH pools still missing, #142).
3. DEPLOYER REPUTATION on mc5 cell (feats.csv, creator seen earlier in sample): first launch n=237 mean 1.92 (halves 2.03/1.81); repeat creator n=67 mean 1.00 (1.07/0.88 and 1.0/1.0). P(reach 250k by 300 s) 6.5 % first launch, 86 % if creator already had a big coin (fleet re-use). Repeat deployers sit flat at 1.0: skipping them removes capital use, not profit. Confirmed on a second dataset (iteration 9). MATURE as filter.
4. BUYER ARRIVAL (epidemic R, conviction sizing) on mc5 cell: late = buyers300-buyers60. late<=5 n=121 mean 1.20 hit5 0.8 %; late>40 n=172 mean 1.98 hit5 17 % (halves 2.42/1.69). Tiered size 2x late>40 / 0.5x late<=5: profit per $ 0.716 -> 0.903 (halves 0.80->1.14, 0.64->0.72). Variant buyers60>=15 2x: 0.833. Non-kit coins carry the effect (buyers60>=10: n=35 mean 2.81 vs 1.21, halves 3.7/1.75 vs 1.44/0.99; kit A/B no split since bundles always show 15+). READY FOR PAPER as a sizing arm (thresholds in-sample; not a gate).
5. STOP / KIT LIFECYCLE: 60 % of cell coins touch the 0.5x stop first; 26 % of those end >=1, 8 % end >=3 (known exit result, tail edge). Kit A lived 24-30 Sep (7 d), kit B 1 Oct-8 Oct+ (8+ d, 10-40 coins/day, still alive 8 Oct). Kit lifespan ~1 week is a live-risk note (detector must find kit C quickly). DISCARD as rule, risk note.
6. DEXSCREENER PROFILE / BOOST (marketing spend; trending.db): 263 profile tokens since 6 Oct, profile appears median 35 min after launch (too late for a first-minute entry). Matched n=46 tape pools: price 30 min after profile vs same-family control at age 35-55 min: v3100 1.08 vs 1.02, v4 3000 1.82 vs 1.67, h100 1.33 vs 0.94 (n=6). No signal. DISCARD.
7. LATE-ENTRY DRIFT CURVE (price in the next 30 min, carry-forward, pulls = 0): v4 3000 hooked: entry at age 5 min 6.3x (n=111, 95 % > 1.3x), 20 min 2.5x, 40 min 1.63x, 90 min 1.23x, 4 h 1.14x; all other families decay to <=1 by 40 min. H3000 is a mechanical ramp (A2 owns it); exit by age, not only by tp. Info for A2 (sellability remains the risk).
8. DEPLOYER SELF-TRADING (live v3live, 900 random fee-100 pools, tx.from fetched via rpcpool free; 205 pools with reserve>=0.5 ETH at 60 s, entry 60 s, exit +30 min, pull=0): deployer never trades own pool 1.65 (n=136); trades <=300 s 1.17 (n=69); sells 0.74 (n=29, 69 % pulled); deployer with 2+ pools in sample 0.45 (n=21, 81 % pulled) vs single 1.61. Same direction in both halves. READY as avoid filter A1-NOREPEAT (small n) + data request (v3live pools.creator column is empty, liq table has only 78 add rows).
9. Cross-check of 3 and 8 = repeat deployer is the fleet signature on two datasets and two horizons.
10. Pipeline: testmgr arms A1-LATEBUYERS (sizing) and A1-NOREPEAT (avoid), ppfeeds request for deployer per pool, 5 discards.

## Exact rules for testmgr
A1-LATEBUYERS (sizing arm on the mc5 test, same entry/exit as mc5): at the +300 s entry check, late = unique buyers (first 300 s) minus unique buyers (first 60 s) counted from the first swap. late>40 -> 2x size; late<=5 -> 0.5x; else 1x. Evidence: 304 coins, profit per $ 0.716 -> 0.903, both halves up (0.80->1.14, 0.64->0.72). Kill: weighted profit per $ not above flat after 60 paper trades, or the late>40 group mean falls <1.3. Caveat: thresholds in-sample; size caps stay Ben's call (D-030).
A1-NOREPEAT (avoid filter on any fee-100 v3/v4 entry test, 60 s and 300 s entries): skip if the deployer (tx.from of the pool-creation tx) created any earlier pool in the previous 7 days. Evidence: mc5 cell repeat n=67 mean 1.00 vs 1.92; live v3 sample repeat n=21 mean 0.45 (81 % pulled) vs 1.61. Needs deployer feed (ppfeeds #250); until then testmgr can use the creator field in the mc5 detector. Kill: repeat group mean >=1.3 after 40 paper coins.