# a3: v3 early tx.from wallets, walk-forward (#713) 10 Oct
Data: v3tx.db early (a3run 242 runner pools, a3ctl 995 random fee-100 controls; first 60 buys in 180 s). Script v3w.py. Runners fully sampled, controls ~1/7.4 of fee-100 pools -> rates reweighted x7 on controls.
Method: pools in time order; a wallet is "winner-linked" if it was an early buyer in an EARLIER runner pool (and not more in controls). Infra wallets (in >=30 pools, 210) excluded. Scored on last 2/3 of pools.
- Early buyers winner-linked >=1: n274, 117 runners, population-weighted P(runner) 9.6% vs 1.7% (5.6x). >=2: 21% vs 2.0% (10x, n137). >=3: 34% vs 2.1% (16x, n102). >=5: 48.5% vs 2.4% (n76, 66 runners).
- Creator repeat winner: 15 of 16 pools whose creator had an earlier runner were runners (creators 0x7a0c.., 0x2c25..: serial launchers); a creator with earlier controls only: 5.6% (n54) vs 19% base (sample). Tiny n, 2 creators drive it.
Caveats: (1) lift is within first 180 s: entry price at that moment unknown, and entry may be already late (A1/A2 must replay cap at signal time). (2) Weight 7.4 is an estimate; unweighted lift is 2-4x anyway. (3) Wallet lists need to grow forward; sampling-by-outcome bias means the list is built from the runner sample only.
Next: testmgr forward arm "A3-WLINK": maintain set of early buyers of past >=1M v3 fee-100 runners (excl infra); fire when >=3 of them buy within first 180 s; judge cap at signal and tp5. Analyst2 can replay on a2band first (cap at 3rd linked buy).