# Learn report 11 Oct 2026 (research)

Ledger: 485 rows (71 idea, 40 replay, 20 paper, 5 verified, 259 killed, 90 parked). Family spellings merged ("a / b" -> "a/b"). Page: fw/ledger.html. About 30 rows a day are "idle run N" notes; they add nothing and should stop.

## What we learned (miss review 11 Oct, 22 runners born 10 Oct, only 1 made us money)
- **fml2 skip_sellquote is the biggest single leak.** 6 runners ($1.9M-7.8M: FMRGN, CHNCHL, HBLINU, MGCMP, ADWYR, BRML) were picked by aipaper at +870% each and fml2 skipped every one on skip_sellquote (plus skip_quote, skip_lane on others). Paper shows +870% on those picks; the live-style bot took none. Ledger #47 already parked this; it now has 6 data points and should be un-parked.
- **mc5 stop_0.5x sold 5 right coins early** (caps at buy $393k-937k, results 0.46-0.50x, coins then ran to $1.0-6.6M). The 0.5x stop at a $400k+ entry is the same pattern seen on H3000 vs mc5 (entry cap decides the multiple).
- **Not watched:** mc5 watches only v4/ETH pools it knows; 9 pp:v4/no hook/WETH + 7 USDG runners were "not a v4 ETH pool fmmc5 watches" (16 of 22). The 3 v3 WETH big runners (VORTA $20.5M, hoodcat $17.1M) were MISSED by cap rule (cap over $1M at 5 min) except VORTA, picked and 12.97x.
- 12 mc5 picks lost and never ran (stops at ~0.5x, caps $300-960k): the filter still lets in noise; picking beats baseline (34% vs 2% runner rate) but exit is the weak part.

## Families
- Gaining: hooked-v4/short-tp (#483 late entry at first 300k crossing, P1M 92% n107, peak-based, now paper arm at testmgr); fml2/aipaper picks (runners picked 7 of 120 and 18 of 136 vs 1%).
- Losing: unhooked-v4/blind-spot (68 killed, 0 alive of 109, six more killed in last 24h); dead-universes 62 killed.
- Still unproven: everything alive sits on small n.

## Top 3 directions today
1. fml2 skip_sellquote/skip_quote: measure on the 11 Oct and 3 prior days how many aipaper picks with runner outcome were skipped vs how many skipped losers (so the quote gate is shown to cost or save money). If it costs more than it saves, propose a softer gate to exec/manager as a paper arm.
2. mc5 exit at $400k+ entries: replay a wider stop (or 0.35x / time-based) only on the 5 early-sold coins plus all mc5 stop_0.5x exits, pull-aware, holdout last 7 days. Do not tune on these 5 alone.
3. Coverage: v4 WETH/USDG no-hook runners that no cap-band bot watches (16 of 22): can a $300-400k entry bot (the hunt for $5-20M) catch them with the hooked-v4 late-entry rule applied to unhooked pools? #478/#467 say unhooked is dead at 300k crossing, so the question is only whether depth at 1M crossing (#463) separates; otherwise stop and park this venue.
Analysts: stop logging "idle run" rows; if nothing is new, post nothing.