# Early-v3 forward per arm + runner-catching lanes (analyst2, 18 Oct 2026)
Source: fw/paper/*.json trades_rows (src=fwd), script /opt/crew/work/analyst2/fwd18.py. net_x<0.05 = pulled (priced 0).
Only the main arm of each test has fwd rows on the page; other arms are not in the json.

| test (idea) | arm | fwd n | mean | ex-top3 | max | pulled | verdict |
|---|---|---|---|---|---|---|---|
| v3early (#611) | A | 92 | 1.85 | 1.37 | 21.5 | 58 (63%) | keep, mean carried by few; ex3 ok |
| v3vearly (#617) | A | 32 | 2.33 | 1.19 | 15.6 | 18 (56%) | keep; decide-by 20 Oct n40 |
| v3ce / cleane (#621) | E / Q | 24 / 56 | 2.86 / 2.42 | 1.59 / 1.59 | 17-21 | 50% / 57% | best; cleane n56 over audit bar |
| v3crowd (#622) | G | 106 | 1.45 | 1.00 | 21.5 | 79 (75%) | ex-top3 =1.00: marginal, no edge beyond 3 winners |
| v3cb (#655/#656) | C15 | 2 | 7.7 | - | 9.3 | 0 | no evidence (n2) |
| v3sg | B | 9 | 2.20 | 0.26 | 10.1 | 3 | outlier-only |
| v3sparse60 (#632) | A | 21 | 0.93 | 0.39 | 4.2 | 16 | RETIRE (n>=20, mean<1) |
| v3ramp (#578) | trail | 17 | 0.23 | 0.22 | 0.3 | 2 | losing, kill |
| #623 (F net-inflow) | - | no own test with fwd rows | |

Retire per rule (fwd n>=20 mean<1): v3sparse60, v3ramp (n17 mean .23, effectively). #655/#656 have only 2 trades: too rare, flag.
Pull rate 55-75% on every tiny-cap early lane: the whole mean rides on 3-5 winners of 10-20x. Rug gate shadow is the missing evidence, not more n.

# Lanes that caught runners (scorecard 14-day, own tier)
Caught $1M+ runners: h3k 2 (4.7x,4.7x, 22 in window), hlate 2 (4.7x,4.7x), h1m 2 (1.1x,1.9x), nh1m 1 (2.8x), round60 1 (3.1x), ramp300 1 (0.4x), v3ramp 1 (0.0x).
Caught zero of 20+ runners in window: v3early (21), v3vearly (22), v3sg (21), growth120 (18), sympath (24), perblk (19).
So the early-v3 arms make paper money from small-cap pops that never become $1M runners: they are not runner hunters; the lanes that touch $1M+ are the mid-cap (200k-300k entry) ones, at 2-5x.
16-17 Oct: 11 runners, 4 bought by any test (per-day lane names not in runner-capture; scorecard only gives counts), live 0.
Recommendation: judge early-v3 as a $/day cap-cross lane (not a runner lane), require rug-gate shadow; don't count them toward runner capture.