# analyst1 iterate run 13 (10 Oct): 10 lenses, 1 testable tag, 1 exit arm, 1 data request

Scripts: z1-z7.py in /opt/crew/work/analyst1 (data: r6weth P2.pkl, win/feats.csv 6,554 launches 12 Sep-8 Oct, tokens.db, live RPC sample). New lenses: router identity of buyers, chain "attention" hot hand, template meta, hub-token-quoted pools, sequencer speed, wallet profile.

| # | lens | hypothesis | numbers | verdict |
|---|------|-----------|---------|---------|
| 1 | Who is on the other side: router of first-60 s buyers (retail Telegram-bot routers vs custom) | pools whose buyers use many routers (more retail) do better | h100 n=411 router 0x887678 only 2.24, 2 routers 2.45; u100 1 router 0.84, 3-5 routers 1.12; v3100 1.32 (n=56) / 1.02 / 1.11. Router mix flat | DISCARD |
| 2 | Market regime as chain attention: number of coins with >=40 buyers (by +300 s) launched 10 min-3 h earlier | hot chain lifts hit rate of mc5 coins | 2,600 entered coins: hit3 17.4 / 20.9 / 21.5 / 25.1 % by quartile (650 each), same sign inside each 6 h block and in both halves; payoff proxy (stop 0.5 else 4 h end) 0.89 / 1.18 / 1.06 / 1.15; bottom quartile (<=4 big coins in 3 h) 0.89 vs rest 1.13, bootstrap P(diff<0)=99.5 %, but only 14 of 25 days agree and half 1 bottom quartile is 1.10 | MATURE as a paper TAG (testmgr), not a size rule yet |
| 3 | Inversion: what rugs inside 60 s (all 6,554 launches, 38 % rugged by +60 s, 56 % by +300 s) | skin in the game: creator buys in the launch block | creator in launch block: rug5 25 % vs 67 % (n=1,714 / 4,840); kit B/A 0-8 %; fee 10000 9 %, fee 100 69 %. Among coins that pass mc5 entry it does NOT help (hit3 21.7 vs 20.9 %; payoff 1.16 vs 1.00, halves 0.97 vs 0.99 / 1.32 vs 1.02 flip) | DISCARD for mc5 (note: useful only for any unfiltered fee-100 buyer) |
| 4 | Meta of operator templates: trade a template only after its last K resolved pools averaged >= thr (rise and death of a template) | momentum in templates beats the pool average | 8,324 templated pools, walk-forward. All pools 1.25. K=5 thr 1.6: 1.68 (n=766) but ALL of it is the hooked family (hooked 1.72 vs hooked base 1.37). Non-hooked: n=110 1.25 vs base 1.05, n=53 K=5 1.16. Nothing new | DISCARD |
| 5 | Exits: partial TP and pyramiding on h100 (n=414, pull = 0) | a ladder or adding on confirmation beats tp3 | tp3 2.39 (14 % total loss), tp5 2.52, tp10 2.66; half tp3 + half tp10 2.52; ladder 1/3 at 2x, 5x, 15x then 20 min 2.37 with only 5 % total loss (vs 14 %); pyramid add at 1.3x/1.5x/2x: 2.23/2.10/1.89 (worse) | pyramid DISCARD; ladder = same mean, 1/3 the ruin risk: ARM for testmgr |
| 6 | Chain-native meta: pools quoted in hub tokens (not WETH/stable) | a blind spot like the native-ETH gap | tokens.db: 3 hub tokens (0xd0601c, 0x2e8c31, 0x4a0e65) carry 12.9k/13.5k/11.3k dynamic-fee hooked pools (hook 0x4e3468) since July. Live sample of 50 pools (free RPC, 12,000 blocks each): 72-76 % never get 3 swaps, n=11 tradable, mean 0.92, fee 0.1-3 %. Too thin to matter | DISCARD (no request) |
| 7 | Speed lens: sequencer feed / mempool (Arbitrum has no public mempool, the sequencer feed saves < 1 s) | pre-confirmation view helps | hooked entry delay 0/10/20/60 s tp3 2.58/2.55/2.50/2.37 (iterate run 6); kitfast already 0.77 s. Gain <= 0.03 | DISCARD |
| 8 | Follow the wallets: wallet age, nonce and funding source of first-60 s buyers | young single-use wallets = fleet, old multi-use = retail; retail presence changes the return | only 867 distinct tx.from in v3from.db; no age data anywhere | DATA REQUEST to dataeng (below) |
| 9 | Family decay: 3-day blocks of h100 / v3100 tp5 | families die quickly, rotate capital | h100 1.76 1.93 2.32 2.52 2.16 1.86 (n 28-129), v3100 2.1 1.22 1.42 0.90 1.20 1.23 1.60. No decay | DISCARD |
| 10 | Hold longer than 4 h for runners (mc5 universe, 231 coins still >=3x at 4 h) | tail continues | pk24x/end4x median 1.34-1.46, only 28-48 % exceed 1.5x; peak is not exit price, earlier mc5 hold tests already KILLED longer holds | DISCARD |

## Data request: wallet profile (dataeng)
- Wallets: all distinct tx.from in /opt/ppcopy/txfrom_598.db (5,325 wallets) plus /opt/ppcopy/v3from.db swap_from.frm (867).
- Fields per wallet: first tx block/time on the chain (age), current nonce, funding source of the first incoming ETH (address + whether that address funded >=20 other wallets), ETH balance now. Source: Robinhood chain explorer API or free RPC; no paid dRPC.
- Format: sqlite /opt/ppcopy/walletprof.db table wallet(addr, first_block, first_t, nonce, funder, funder_fanout, bal_eth).
- Test: per pool, share of first-60 s buyers that are young (<1 day) with fanout funder; compare tp3 return of pools with low vs high share (h100 n=302 refined cohort, v3100 n=1,075 where v3from covers). Kill: spread < 0.2x in return or opposite sign in the two halves.

## For testmgr (#241), exact rules
1. TAG (no new trades): on every mc5/kit paper trade store `big3h` = number of pools created 10 min to 3 h before the trade's pool that had >= 40 distinct buyers by their +300 s. Report mean payoff by big3h (<=4, 5-7, 8-10, 11+) when n >= 100 trades. Kill: lowest bucket not below the rest by >= 0.15x in both halves. If it holds, the rule is half size at big3h <= 4 (Ben decides, D-030).
2. ARM on the H100 test: same entry (+60 s, hooked v4 fee 100), exit 1/3 at 2x, 1/3 at 5x, 1/3 at 15x or 20 min or pull. Sim 2.37 vs tp3 2.39, total-loss share 5 % vs 14 %. Compare PnL and worst day, not just mean.
Evidence limits: in-sample thresholds, 2,600 coins 12 Sep-8 Oct, 25 days.