# Analyst 2, idle run 38 (10 Oct 2026): 10 iterations, scripts it38/
Hunt: the live 5-Min Cap band (v3 fee100, 250k-1M at +5 min, entry +345 s, stop 0.5, 4 h) and the pool types we do not watch. Tape: /opt/sw/v3live.db (6-10 Oct, fresh out-of-sample), v3hist (12 Sep-6 Oct), live trades /opt/fm_mirror/mc5.db, runner list /opt/ppcopy/e1x-runners-v2.csv.

| # | Lens (new = never used) | Result | Verdict |
|---|---|---|---|
| 1 | Live vs replay on the same 4 days (6-10 Oct, n=54 band pools, 7% slip, glitch prints removed) | replay mean 0.91x, 4% reach 5x before the stop, 31% stop inside 3 min (live: 40%, 7% winners). tp5 1.16, tp10 1.38. The 2.2x seen on Sep data does not show on the newest 4 days outside 2 runners. | INFO to Ben/Manager: edge is 2-3 trades a week |
| 2 | **Counterfactual audit of every live decision (new)**: which band pools did the bot not buy, and why | 49 band pools, 47 traded. The 2 missing: a **72x runner lost to HTTP 429** (both RPCs rate limited, `entry_err`) and a 22x runner lost to "Swap confirmation required" (buy_fail). Posted to exec #593. Details: fw/a2-mc5-techmiss-10oct.md | READY (ops fix, exec) |
| 3 | Runner scoreboard 6-9 Oct (real = >=10 buyers, wash<0.5, best pool of token): 42 real $1M+ runners | v3 fee100 31, H3000-type hooked WETH 9, v4 no-hook (WETH/native/USDG/SUSD) 5, all born big or thin. Of the 7 v3 runners that went 17x+: 2 bought and held (29x, 19.5x), 2 lost to errors (#2), 2 bought and stopped at 0.39-0.44x before running 19-30x, 1 cap_out at 1.008M. The unwatched types hold no catchable real runner in these days. | INFO |
| 4 | Exit: pull-only exit instead of price stop (liq removal table) | no stop: 0.59, stop 0.5: 0.91 (median final value without stop 0.01: most band pools die by 4 h). Pulls are atomic, cannot be front-run. | DISCARD (= #196) |
| 5 | Exit: hold past 4 h only for survivors at >=3x (28 pools), trail 30% | mean at 4 h 14.4x, with trail 14.2x, hold to end 8.4x (peak 29.9x not capturable) | DISCARD |
| 6 | **Holder graph from real Transfer logs (self-served via free RPC, 169 band pools, +5 min)**: top-1 / top-5 share, holders, creator share | top-1 low tercile clip10 mean 0.91 vs high 0.64, but hits 3/2/1 of 6, halves 0.77 vs 0.53 and 1.03 vs 1.03. Flat. | DISCARD. Ask dataeng to close #562 and #553 (holder tables) |
| 7 | **Bundle at creation (bundle_fund.db, "gold" set, new on mc5 band)**: 86 of 222 band pools bundled (6-22 wallets, $27-57k) | bundled: 1 hit of 4x in 86, clip10 mean 0.55; unbundled 6 in 136, 0.84. Same sign in both halves, but unbundled set includes pools the scan never covered (scan ends 8 Oct, only 'gold'). Bundle sold >1% by 300 s: 1.33, nothing sold: 0.40 (liveness proxy). | PARKED, data request below |
| 8 | H3000 (hooked WETH fee 3000) timing re-check, 47 real since 21 Sep | deterministic ramp: 10-14 buyers, 100k at ~5 min, 300k ~15 min, 1M ~55 min, 3M ~2 h for half of them. Watched by h3k; sellability still untested (#64, needs Ben's $5 sell). | KNOWN, nothing new |
| 9 | Stop-then-run: 2 of 7 big runners hit the 0.5 stop (0.39, 0.44 at 104 s and 840 s) and then went 19x and 30x | The exit grid (#196) says wider stops lose on the whole universe; not re-opened. Counted for the record. | INFO |
| 10 | **Data request (below)** | bundle scan with full coverage | REQUEST |

## Request to dataeng3 (owner of bundle_fund.db)
What: `bundle_launch` + `bundle_sold` for EVERY v3 fee100 pool born 12 Sep to now (not only the 160 'gold' pools), refreshed daily; add cumulative pct sold at 60 s and 300 s after birth, and number of distinct outside buyers in the first 300 s.
Format: same sqlite tables as /opt/ppcopy/bundle_fund.db (pool key), copy to /opt/ppcopy/.
Test I run: mc5 band pools (cap 250k-1M at +5 min, about 400+ over 30 days) bundled vs unbundled, hit rate of 4x before the 0.5 stop and clip-10 mean, both halves. Pass: bundled clip10 mean <= 0.7x of unbundled in both halves and hit rate lower at p<0.1 -> skip flag for the live gate. Kill: ratio > 0.85 or halves disagree.