# Analyst 3 idle run 16 (10 Oct 2026 night): breadth, and a hard look at our own filters
Scripts: h4.py h6.py h10.py h11.py h14.py (work/analyst3). Base: v3 mc5 band sample (277 pools, /tmp/a3_c9.pkl, 12 Sep-10 Oct), outcome = tp5 sl.5 multiple (winsorised at 8). 8 iterations (long queries).

1. Born-high v4 nohook (unwatched, canon_pools, cap_first>=500k): median max/first = 1.00 (native 1.04, USDG 1.00, WETH 1.00); only 3-6% ever do 5x from the first print; creators 0x50e8 (20 native), 0x3f5d (20 USDG) = serial kits. Jump/pinned set, NOT a hunt target. DISCARD (re-confirms the earlier census with canon).
2. Unwatched born-low v4 (cap_first<60k): P1M native 1.4-1.9%, USDG 1.0%, WETH fee100 0.9%; P1M by 5-min cap bin USDG 150-400k only 10% (P5M 1%), native 150-400k 12-25% on n=16. No pocket. DISCARD.
3. Lens: our own filters (selection bias). 7 binary flags, best-of-7 permutation test (2000 shuffles): max |t| observed 4.46 vs 95th pct of the null max 2.79 (p=.001). Survive: SEED (t 3.7), VOL5>=49k (4.4), SERIAL skip (4.5), BORN-HIGH skip (4.0); all keep sign in the last 40% (since 3 Oct). Do NOT survive: CAPS (t 0.9, last 40%: 1.00 vs .97), COPY (t 1.1, last 40% 1.10 vs .90), HOT. VERDICT: CAPS and COPY are noise; testmgr may retire both arms early rather than wait for 30.
4. Bootstrap of the stack arm (seed&vol49&!serial&!born_hi, n128, mean 1.49, 25 winners): 90% CI 1.25-1.73, in-sample. Whole band 1.12 (CI .99-1.24). In-sample Kelly 29% is meaningless (grid edge); the real number is the forward one. Do not size from this.
5. NEW LENS: who is buying = breadth (analogy: IPO oversubscription / number of distinct bidders). first5.db: distinct tx_from BUY wallets in first 300 s on v3 band pools (247/277 covered; uncovered 30 have mean .74).
   - router mix (share of buys not via 0xcaf6, 86% of buys) : flat. proxy-wallet share: flat. DISCARD.
   - distinct tx_from buyers >=30: n111 mean 1.57 vs .83 (t 4.1 after scanning 4 thresholds); first 60% 86 pools 1.61 vs .76, last 40% 25 pools 1.40 vs .86. With seed: n80 1.77 vs .89; seed&vol49&nfrom>=30 n79 1.78 (1.82/1.72) vs 0.94 below. hit5 22-25%. 80% of the 30 hit5 winners have >=30 buyers (median 48 vs 17 for losers).
   - Inside vol49 the lift is breadth: nfrom~1 (one wallet makes the volume) 0.92, ~34 buyers 1.34, ~54 buyers 1.59.
   - tx_from (origin EOA) needed: recipient `wallet` count (corr .4) is weaker (1.46/1.03 halves).
   Verdict: READY as forward arm A3-BREADTH (needs live tx.from count, see request to ppfeeds). Threshold chosen in-sample; expect lower forward.
6. Breadth on unwatched v4: first5 rows exist for 100% of v4 pools, but tx_from is almost always NULL/1 for native/USDG (nfrom<=2 on 99%): untestable (= analyst2 #323 ask). WETH fee100: >40 buyers 3.0% P1M vs <=1% (ceiling 1M, y3 = 0). H3000 hooked (127): breadth no lift (lo 84%, hi 79% P1M). So breadth works on v3 where the real runners are; not shown elsewhere.
7. Router mix on band: see 5.
8. Requests: ppfeeds live distinct-buyer count; ppdata tx_from for v4 native/USDG first 5 min (exists as #323, not duplicated).