# PP books leak audit and re-price (FW Analyst 1, sup-to-a1 v41)

## READ THIS FIRST
- 7 Oct 2026 19:40 Bangkok. Forward paper books since 5 Oct 14:36 Bangkok (about 2 days). Re-priced from the chain: Uniswap v4 Swap tapes (PoolManager 0x8366a39cc670b4001a1121b8f6a443a643e40951), buy filled 2 s, 6 s or 12 s after the decision, pool depth impact for each size, 1% fee, $0.10 gas round trip, sell price from ppexit v1 at $200 (so smaller tickets are a bit conservative).
- **The edge survives the re-price.** Fill delay costs little (aipaper real lane: +4.09 per $1 at 2 s, +3.77 at 6 s, +3.57 at 12 s). Ticket size $10 to $50 changes nothing (pools are deep enough).
- **Leak 1 (real): the board counts badhook and pond pools.** Those can block sells. Real lane only is the honest book. aipaper "all" is +2.21 per $1; the other lanes average only about +1.25 and may not be sellable at all.
- **Leak 2 (small): no fill delay on the board.** Worth about 15% to 20% of the return at 6 s to 12 s (aipaper real lane: board +4.54, 6 s +3.77, 12 s +3.57). Not a deal breaker.
- **Leak 3 (open): best of 27 books picked after the fact.** Partly answered: aipaper is positive on every day and every lane cut, and random is negative on every day. Only 2 days of data, so this stays open until live.
- **Leak 4 (open, unmeasured): the decision clock.** The pos clock uses a fixed seconds per block constant. If the real decision is minutes later than the block I price, the paper buys too cheap. The 12 s test covers small drift only. Live fills are the test.
- **Recommended live lane: aipaper (60 s AI), real lane only, $50 a ticket.** Paper says about 16 trades a day, median +2.06 per $1, mean +3.77, win rate 76%, top 3 coins only 23% of profit. That is roughly **$1,600 to $3,000 a day at $50** on paper. Expect live to be lower.
- **Kill bar:** stop after 20 live trades if the mean is under +0.5 per $1 or the median is 0 or below; stop on 2 losing days in a row; stop if the median live buy price is more than 20% worse than the paper buy price for the same coin (that would mean Leak 4 is real).
- Paper only. The live switch is Ben's.

## Re-priced books (mean return per $1, one row per coin)

| book | lane | n | board $200 | $10 at 2 s | $50 at 2 s | $50 at 6 s | $50 at 12 s | $50 at 2 s, best coin removed |
|---|---|---|---|---|---|---|---|---|
| aipaper (60 s AI) | all | 113 | +2.21 | +2.13 | +2.13 | +2.01 | +1.89 | +2.06 |
| aipaper (60 s AI) | real lane | 33 | +4.54 | +4.09 | +4.09 | +3.77 | +3.57 | +3.90 |
| aipaper30 (30 s AI) | all | 107 | +1.65 | +1.52 | +1.53 | +1.43 | +1.29 | +1.44 |
| aipaper30 (30 s AI) | real lane | 27 | +4.46 | +3.98 | +3.98 | +3.83 | +3.49 | +3.72 |
| v5_top5 | all | 138 | +1.19 | +1.19 | +1.20 | +1.11 | +1.04 | +1.13 |
| v5_top5 | real lane | 31 | +1.80 | +1.85 | +1.85 | +1.77 | +1.63 | +1.55 |
| f5_20s_top5 | all | 119 | +0.86 | +0.89 | +0.88 | +0.78 | +0.65 | +0.81 |
| f5_20s_top5 | real lane | 34 | +1.66 | +1.72 | +1.72 | +1.53 | +1.40 | +1.49 |
| money book 10 s (m10) | all | 11 | +2.66 | +2.47 | +2.47 | +2.61 | +2.46 | +1.37 |
| money book 10 s (m10) | real lane | 3 | +7.42 | +6.69 | +6.69 | +8.09 | +8.25 | +3.29 |
| old 30 s money book | all | 134 | +0.85 | +0.87 | +0.87 | +0.80 | +0.73 | +0.81 |
| old 30 s money book | real lane | 51 | +1.13 | +1.20 | +1.20 | +1.12 | +1.06 | +1.04 |
| rand_30s (random) | real lane | 167 | -0.54 | -0.56 | -0.55 | -0.57 | -0.60 | -0.61 |
| rand_gated (random) | real lane | 81 | -0.41 | -0.43 | -0.42 | -0.43 | -0.46 | -0.54 |

$20 and $200 columns match $10 and $50 within 0.03 (see raw file). m10 has only 3 real lane coins: too few to trade.

## Per day, real lane, $50 ticket, 6 s fill (mean per $1, n)

| book | 5 Oct | 6 Oct | 7 Oct (to 16:40) |
|---|---|---|---|
| aipaper | +4.09 (6) | +2.64 (18) | +5.83 (9) |
| aipaper30 | +1.80 (10) | +5.42 (10) | +4.47 (7) |
| rand_30s | -0.80 (14) | -0.41 (100) | -0.82 (53) |
| rand_gated | -0.43 (4) | -0.16 (44) | -0.78 (33) |

## Why (first principles)
- Profit per trade = (sell price × tokens) minus cost. Tokens bought depend on the price at the fill block plus my own price impact. Going from $10 to $50 moved returns by under 0.01 per $1, so my own price impact at these sizes is tiny. Delay matters only if the price runs in the first seconds after the pick; it drifts up a little (2 s to 12 s costs about 13%), so the AI is early, not instant.
- Badhook pools can stop sells by contract. A paper sell there is not money. That is why only the real lane counts.

## Method
- research-box /root/fwa1/v41/v41b.py (v41.py plus a bad print guard: any tape price more than 100x away from entry is replaced by the last good price). Rows in v41b_rows.json, raw table v41b.md, tapes cached in tape.json.
- Books from main /opt/hist/analysis/{aipaper,aipaper30,aipaper5}.db and ppexit.db (sellsys v1.3, rule v1: 10x take profit, dead 60 s, 30 min). Lanes from main crew.db.
- Orientation check: 29 of 33 aipaper real lane rows matched the stored entry price within a factor of 2.7; the other 4 are kept as is.

Ledger: FW-025. Version: v1, 7 Oct 2026 19:40 Bangkok, FW Analyst 1. Latest: http://159.223.88.168/hx-c9ois/fw/lab/pp-books-audit.md