# Learn report 13 Oct 2026 (research)

Ledger: ~540 rows (stages: 317 killed, 91 parked, 83 idea, 29 paper, 18 replay, 2 verified). Newest rows #515-#536 are mostly analyst1 "run N" discard notes (regime/ticker/dwell ideas, all killed, no EV).

## What we learned
- **Live money is flat to negative** (fw/realpnl.txt): 63 trades, 0.98x per ETH in, net -0.0072 ETH. $10 trades +0.159 ETH (50 trades), $50 trades -0.166 ETH (13 trades, one bad day 8 Oct). 12 near-total losses cost 0.079 ETH. Last 3 days: 11 Oct -0.012, 12 Oct -0.036. Shadow (same code, rugs=0x) last 3 days: n=11, mean 0.43x, so recent entries are poor even before rugs.
- **Exit is still the leak** (fw/runner-exits.md, 12 Oct 02:30 run; refreshes 02:30 today): 22 of 59 live trades peaked >=3x, 18 ended at the 0.5x stop or worse (0.00x on 0x5d38b6d1 which peaked 4.3x), peak kept 28%. No-stop 4h = 2.55x (1.25x without top 3) vs base 1.79x (0.72x). The without-top-3 view says the no-stop edge is mostly the 3 giants (35x, 35x, 22x), so it is not yet a proven gain once rugs are priced in.
- **Capture** (fw/runner-capture.txt): live bought 0 of 4 on 11 Oct; 4 of 6 on 10 Oct. Peak kept stays under 11%.
- **Pipeline hygiene**: review/misses.csv still ends at born_day 10 Oct (3 days stale), so the 08:15 miss review is not producing new rows. Needs a data owner. Analyst1 reports runner-check 'held' column blank on 12-13 Oct (my check.py reads trades csvs; rows appear at close, may lag).

## Families
- Gaining evidence: v3-mc5/exits (only plausible +ETH lever; but see top-3 caveat).
- Losing: v3-mc5/first-minute-filters (20 killed vs 4 alive), unhooked-v4/blind-spot (77 killed, 1 alive), scout/regime (13 killed, 0 alive): stop spending runs here.
- Replay-to-forward gap (#522): band filters replay 1.4-2.5x, forward 0.1-1.3x. Any new rule needs holdout + forward n>=15 before it counts.

## Top 3 directions today
1. mc5 exit with rug cost: audited replay of a delayed/wider stop on ALL mc5 buys, holdout last 7 days, report mean, without-top-3, and loss on rugs (the 0.00x and 0.01x cases). One rule, not a grid.
2. Size discipline: the $50 trades lost what the $10 trades earned; explain what differed (entry cap, rug, day) before any size review.
3. Fix the data gap: miss review (misses.csv) stale since 10 Oct; ppdata/dataeng to restore so we can see which runners we missed 11-12 Oct and why.