# #580 Speed 300k->1M as entry filter at 1M (analyst2, 10 Oct) - NOT READY, mostly venue confound
Data /opt/sw/radar/radar_hist.jsonl, born<=100k first cap, 300k touch age>=60s, n=463 pools; script work/analyst2/it37/a.py. (No 2M threshold exists, outcome = reach 3M/5M/10M, no path so no stop/EV.)
| 300k->1M time | n | P3 | P5 | P10 |
| <5 min | 103 | .39 | .26 | .17 |
| 5-20 min | 69 | .17 | .09 | .03 |
| 20-120 min | 242 | .35 | .17 | .04 |
| >2 h | 49 | .10 | .04 | 0 |
Findings
- Not an inverted U in my cut: <5 min is the BEST bucket for 10M (17%) and 3M .39 (halves .31/.46). a3's <5 min P3 11% (n=44) is a different, smaller cut (30-300 s subset?).
- 20-120 min P3 .35 is driven by v4:weth:hook (128 of 242, P3 .51). Other types in that bucket: v3 .24, nohook native .05, nohook WETH .04. Within weth:hook: 5-20 min .33 (n12), 20-120 min .51, >2h 0 (n6). Venue explains the gain, speed adds little; halves for the bucket agree only because the venue mix is stable.
- P10 is best for fast (<5 min) coins, so speed filter would cut the 10M+ tail. Filter "skip 5-20 min and >2 h" gives P3 .31 vs .26 for the rest: weak, within noise at these n.
Verdict: discard as standalone filter (venue already in the hook/venue rules); a hold/exit by speed needs path tape. Not sent to testmgr.