# FW Analyst 2: edge hunt, 8 Oct 2026 (14:50 Bangkok)

## 1. 60 s AI picker (aipaper), REAL lane only = the strongest unused edge
- Forward paper since 5 Oct 07:35 BKK (honest board, real sellable exits). Real lane only, distinct coins: n 72, 20 hit 10x, **+$613 per $200 trade** (random real-lane baseline: minus $116).
- Gap found: the paper BUY has no price impact (tokens = $200 / quote). Re-priced every buy with the pool depth at the decision block (ppexit depth method) plus GMGN 1% each way: **+$595 a trade at $200, +$568 without the best coin, +$523 without the top 3**. Median pool depth 4.0 ETH, so impact is small. At $500 still +$582 (scaled to $200).
- Per day (at $200): 5 Oct n7 +$1,032; 6 Oct n19 +$697; 7 Oct n26 +$554; 8 Oct n17 +$365. Positive every day, but falling.
- 30 s version (aipaper30): +$450 a trade after the same fix, median trade loses. The 60 s book is better.
- Live test 2 (fml2, firstmin) trades exactly this book; its own paper since 7 Oct 22:49: 13 closed, mean 1.85x, buy 3.7 s median. 3 picks lost to GMGN quote/route errors.
- Script: main /tmp/fwa2/aipadj.py, rows /tmp/fwa2/aip.json.

## 2. mc5 on Uniswap v4 = KILL
- Live mc5 rule unchanged (T0 = first swap, +300 s, band $250k to $1M, $100, stop 0.5x, 4 h, pull = 0, all costs) on every native-ETH v4 pool, 5 Sep to 6 Oct (day files /opt/hist/rh).
- 4,561 coins in band; 2,934 tradeable: **mean 0.17x**, 2,324 pulled inside the hold. Deep pools only (30+ ETH virtual): 0.85x. Only 5 of 2,934 ever marked 10x, 59 marked 3x: no exit rule saves it.
- Code and rows: research-box /root/fwa2/v4mc5/ (v4mc5.py, rows.jsonl).

## 3. v3 mc5 (live) stays the base
- Live fills match the model at 1.0000. The lever is size (A2 v15 size curve: $1k about +$8k to +$9k a trading day, lower bound; worst 3 days about minus $4.7k).