# Learn report 15 Oct 2026 (research)

Ledger ~593 rows; every new idea (#587-#593) has a row. #593 (keep cap_out at 1M) killed as a non-change.

## What we learned
- **Live money flat** (realpnl 15 Oct): 63 trades, 0.98x, net -0.0072 ETH. No new live trades since 12 Oct's day total; $50 lane -0.166 ETH vs $10 lane +0.159 ETH is still unexplained. Shadow last 3 days n=30, mean 1.70x (steady vs yesterday's 1.82x at n=19).
- **Misses 14 Oct**: a quiet day, one runner (SEALS, 2.7M at +5 min). The V3-R300 paper bot bought the 5-min top and lost 0.35x: loosening mc5's cap_out would buy spike tops. Keep cap_out. A catch needs entry before ~60 s, which belongs to the early-entry lanes.
- **Capture**: 13 Oct had 11 runners of $1M+, live bought 0, any bot bought 7. Live has bought 0 on 3 of the last 4 days.
- **Exit leak (runner-exits.md)**: 30 of 79 live trades peaked 3x or more. 8 of the 10 biggest (35x, 35x, 31x, 29x, 22x ...) ended at the 0.5x stop, keeping 1% of the peak. The ones with no stop hit kept 46-79%. This is the largest ETH lever we have, and the open question is rug cost.

## Families
- Gaining: v3-mc5/exits (the lever), miss/native-kit (blocked on tape).
- Losing: unhooked-v4/blind-spot (12 killed, 0 alive), v3-mc5/wallets (24 killed), first-minute-filters (21 killed). No more runs there.

## Top 3 directions today
1. mc5 exit: one audited rule (wider or delayed stop, e.g. no stop for the first N minutes) on all mc5 buys, with rug loss priced, mean, without-top-3, and a holdout of 9-14 Oct.
2. Early ramp (SEALS, 13 Oct cluster 2): entry at +90-120 s on buyer growth, including the V3-R300 loser leg. Replay first, then decide.
3. Why live bought 0 of 11 runners on 13 Oct: for each, give the reason (cap_out, not watched, no signal). Fix whichever reason repeats.