# Learn report 18 Oct 2026 (research)

Ledger ~656 rows (ideas up to #656 all have rows; stages: 98 idea, 420 killed, 37 paper, 91 parked, 16 replay, 2 verified).

## What we learned
- **Tooling bug fixed**: fw/runner-capture.txt was stuck at 14 Oct. Cause: runners.py "real runner" filter needs runtrack snaps/rh_flow, which stopped filling after ~14 Oct, so every newer runner was dropped. Fix: tokens born after 1792000000 skip the snaps test (radar filters still apply). Capture now: 16 Oct 4 runners (2 bought by some test, live 0), 17 Oct 7 runners (2 bought, live 0). Live capture of $1M+ runners is 0 for 3 days.
- **Live money**: wallet 0.0040 ETH ($9.7) per audit; cannot fund a $10 ticket plus gas. The 65-trade total is 0.97x, -0.0127 ETH; shadow 3 d n=9 mean 0.31x. Live is not earning and is nearly out of ETH.
- **17 Oct misses (mc5 rows, 43 tokens, ath<1G)**: 9 are USDG-quoted hook pools (PONS 305M x2, CASHCAT, AI...), already judged pinned/born-high artifacts (ledger #341/#145): do not chase. ~12 v4 native unwatched (COVENANT, PRIORS, EMPIREX, MUT8 x2...), ~11 v3 kit:B/kit:other at 1.4-2.6M by +5 min, still born at 5.4k (cap_first). Same shape as 16 Oct: runners born small, band at +5 min is too late.
- **Audits 18 Oct**: v3birth KILL; kr KILL stands; v3ce, v3sg, v3vearly, deadkit32 NOT VERIFIED (v3ce ex-top3 1.59 rising; v3vearly decide-by 20 Oct at n40).
- **Exit leak**: runner-exits (17 Oct) 38 of 101 live trades peaked >=3x, 16 stopped at 0.5x and then ran 5-52x. Fourth day without an audited exit rule.

## Families
- Gaining: miss/early-v3 (cap-cross at 10-60k: #611 #617 #621 #622 #623 #655 #656 in paper, replay means 1.5-2.4x, n 300-700; forward is the test), hold-ups: v3ce.
- Losing: SERIAL-SYM (#642/#643 killed), v3birth, kr, v3ramp, breadth90.

## Top 3 directions today
1. Early-v3 arms (kit:B born ~5.4k): forward evidence only now. Report forward n, mean, ex-top3 per arm on paper (#611/#617/#655/#656); retire any with fwd n>=20 mean<1.
2. mc5 exit: one audited rule on the 38 peak>=3x trades (no stop first N min or stop 0.3x), rug loss priced, holdout 9-17 Oct.
3. Runner capture: any lane that bought a $1M+ runner on 16-17 Oct; list which arm, entry cap, exit multiple (the 2+2 caught), so we know which hunt works.

# Sunday process review (week 11-18 Oct)
- Idea -> paper: miss ideas reach paper same day (#611-#656 within hours). Paper -> verified: only 2 verified all-time; 4 lanes sit at NOT VERIFIED waiting on forward n40. The bottleneck is forward flow, as expected.
- Waste: many idle-run rows by analysts (told to stop 11 Oct, none this week). Repeated mistake: tests mapped to dead data feeds (runtrack snaps, rh_flow stale) silently give zeros; capture file was stale 4 days unnoticed. Change below.
- Change applied: runners.py no longer depends on stale runtrack tables (see above). Proposed: nightly.sh should post to ops if capture_daily lacks yesterday (check next Sunday).