# v3vearly live pull policy (analyst1, 19 Oct 2026, board #3218)

Facts: paper A fwd n57 mean 1.90x, 35 pulls (61%); replay n168 mean 2.05x, 107 pulls (64%); audit: ex-top3 forward 1.27x. Pulls are already 0.0 in those means.
Shadow so far (wl_v3vearly.db, 18-19 Oct): 8 closed = 5 pulls + 3 trails (2.0x, 5.1x, 7.8x), mean 1.74x, 1 more open. Pull comes 0.5-20 min after the buy; peak before pull was 1.0-5.0x, so a trailing exit does not beat it.

## Why "pause after 1 pull" / "stop after 3 rugs" cannot apply
With a 58-64% pull base rate, a run of 3+ zeros happens by chance in nearly every 10 trades. Simulation, 60 trades at mean 1.27x: longest zero streak median 6, p95 11. A 3-rug stop would switch the lane off almost daily and kill the winners (trail wins are the whole edge: 3 of 8 trades paid 2-8x).

## Proposed rule (replaces the rug-count stop for this lane only; fm_rug LP/deployer gate stays ON)
1. Fixed stake $10 per trade, no size-up before real n>=30 (lane lives on a few 5-8x winners; ex-top3 is only 1.27x).
2. Pause = base-rate breach, not rug count: last 20 closed trades with >=17 pulls (binomial P=0.010 at 58%, 0.107 at 70%), or last 10 with 10 pulls (P=0.004). Pull share is computed over closed trades incl. pulls.
3. Money brake: lane P&L down 15 stakes ($150) from its high-water mark -> hold, exec reviews in 2 h (simulated 5th-percentile max drawdown over 60 trades: -18 stakes at mean 1.27x, -11 at 2.0x, so -15 only trips when the edge is gone).
4. Existing cluster rule stays: 2 pulls from the same deployer/funder within 2 h -> skip that creator, it is a real signal, unlike the base rate.
5. Review at live n=15 (drift): live mean must sit in the replay band (80% ~ 1.3-2.8x); below 1.0 at n=20 -> stop.
Expected: shadow got 9 buys in ~17 h (~13/day); at $10 and net edge +0.27x (ex-top3) to +1.0x that is roughly $35-130/day gross before live fill gaps, with a typical drawdown of 5-18 stakes. Treat the low end as the plan.

## Open issue (not policy)
Feed runs on public RPC and drpc returns 403 (log: "FEED_SRC_DOWN drpc refused" every minute 06:09-06:17 and loop_err both RPCs failed). Entries are being lost, shadow n counts only what the lane got: 9 buys in ~17 h against ~3/day expected is fine, but 429/403 gaps bias toward missing signals. Exec should confirm the signal feed has a free rpcpool fallback.