# Analyst 3 idle run 5, 10 Oct 2026 (ledger #144, #155-158; scripts work/analyst3 o1-o13.py)

Hunt: v3 mc5 band (enter +306 s, cap 200k-1.5M at +300 s, 7% slip, stop .5, tp5, 4 h) and the unwatched pool types.

## Result in one line
One candidate: **MOMDD flag** in the mc5 band = price rising since 60 s (cap@300s / cap@60s >= 1.1) AND 10-30% below the early peak (cap@300 / max cap in first 300 s in (0.7, 0.9]). tp5 sl.5 net of 7% slip: 2.35 (n=47, hit5 43%, halves 2.56/1.99) vs 0.94-1.12 for the other 179 dust-guarded pools. Not ready: thresholds chosen in-sample, shifted gates give less (see below). Sent to testmgr as a forward paper arm.

## Iterations
1. Token-level accounting (sibling pools). 59 of 98 real-1M USDG-nohook pools are tokens that also have a v3 pool; a v3 token getting a USDG/other sibling pool is a consequence, not a predictor (P1M at +30 min 0.185 vs 0.184). INFO (#144). Unwatched USDG runners = mostly tokens we already see via v3.
2. Radar census 7-10 Oct. 1M crossings by venue: v3 98, usdg 28, native 23, pons_grad 5, weth 5; the v4 ones are single-print/frozen caps (cap_first = cap, age 0). INFO, nothing new.
3. MOMENTUM + PULLBACK (new lens: closing-line analogy, "has the market already corrected and re-bid"). Dust-guarded (n>=20 swaps, vol>=20k) n=226: mom>=1.1 alone 1.6-1.8; pullback 0.7-0.9 alone 1.78; both 2.35 (n=47). Grid mom 1.0-1.2 x dd (0.6-0.95 .. 0.8-0.97): 1.7-2.35, always halves >1.4. Gate shift: +240 s 1.53, +300 s 2.34, +360 s 1.60, +480 s 1.84 vs rest 1.35/1.07/1.28/1.23. So real but the 300 s number is the tuned peak; expect ~1.6-1.8. Analyst1 found the momentum half (cap60/first) already; the new part is the pullback.
   Live check: the 46 mc5 fills (mc5.db) scored with the flag (features from v3live): flagged 8-11 trades (17-24%), they hold all 3 big winners (29x, 19x, 13x) for mom>=1.0 and 2 of 3 for mom>=1.1; mean mult 4-6 vs rest 0.4-1.3. Caveat: about a quarter of the fit sample is the same period, n tiny. Evidence only; live is Ben's call.
4. Resurrection/migration (v3 pool for a token that already traded elsewhere): 23% real 1M vs 3.6%, but all 16 are born-at-price copies (first 100k=300k=1M in one second). Jump trap. DISCARD (#155).
5. Pons graduated pools (358 in v3live tape, which exists, contrary to my earlier note; 6 reach 1M, 2 reach 3M): enter on cap cross 60k-500k, tp2/3/5: 0.8-1.2x; active at +5 min (>=20 swaps, n=69): tp2 hit 46% but 1.10x net. Consistent with #25/#30. DISCARD (#156).
6. Gate microstructure in the band: entry-vs-gate print gap, turnover (vol/cap), average trade size, LP add/remove events in first 5 min, holder overhang (c5/VWAP): all non-monotone or flat (inverted U, halves flip). DISCARD (#158).
7. Dead air (pool creation to first swap): 93% trade within 2 s, the long lags are backfill artifacts. DISCARD. TRAP: v3hist pools.created_t is NULL for 7550 of 7584 and first_trade_t is the backfill time, not the first swap. Data fix requested (below).
8. Gate-shift robustness of iteration 3 (above). MATURE.
9. Live-fill scoring of iteration 3 (above). MATURE, forward arm.
10. Data fix request to dataeng (v3hist/v3live pools table) plus forward arm to testmgr.

## Forward arm (testmgr)
A3-MOMDD, paper only, on v3 pools at birth +300 s (same dust guard as mc5): cap in 250k-1M, swaps>=20, vol>=20k; flag = cap300/cap60 >= 1.1 and 0.7 < cap300/max(cap in 0-300 s) <= 0.9; enter at +306 s, stop .5, tp5 (also arm hold 4 h, no tp), 7% slip. Judge on the first 30 flagged forward pools: kill if mean < 1.3 or hit5 < 15%; keep if >= 1.6. Expect about 3 flagged pools per day (56 in 18 days of history); 30 pools take about 10 days.