# A3.14 Early exit for SW losers: KILL

## READ THIS FIRST
- **Verdict: KILL (clean, by the pre-registered rule).** No candidate signal beats the current exit in half A, so nothing gets frozen. Pre-registration: http://159.223.88.168/hx-c9ois/sw/sw-analyst3-status.md (entry 07 Oct 14:48 Bangkok).
- **Why, in one line:** losers do not die by LP leaving early; they die by selling, and a sell burst before the stop looks the same on coins that later go 5x, so every sell-burst exit costs more on winners than it saves on losers.
- **S1 (in-range LP removal, one Burn of 20%+): fired 0 times in 283 trades.** In the window before the stop or 5x there are only 11 in-range Burns in the whole data, all under 0.3% of in-range liquidity. When LP does leave, it leaves in one go (50%+), and the live pull exit already sells in that same block. Nothing to catch early.
- **S3 (3-block net sell burst over X% of in-range quote reserve): every X loses in half A** (both books together): 5% -$4,977 a day, 10% -$4,344, 20% -$199, 30% -$658.
- The rule picks the best half A cell; that is S1 at +$0 (it never fires), which is not above the current exit, so KILL. Half B priced for the record below.
- Prior kills this agrees with: A3.3 (time and price early exit, +0.03x, noise), V14 net-buy exit, FW A1.28 drip-seller exit.
- cost: Claude credits about 2 h of one session, API $0 (cached public RPC data, no paid calls), 1.8 hours (14:46 to 16:35 Bangkok; promised 16:10, the Mac link dropped for about 20 min).

## Result table ($1k tickets, $ a day change vs the current exit; tick walk fills, signal fill at end of block b+1)
| Cell | Half A live | Half A kit300k | **Half A both** | A fires | A winners cut (cost) | Half B both | B fires | B winners cut (cost) |
|---|---|---|---|---|---|---|---|---|
| S1 Burn 20%+ | 0 | 0 | **0** | 0 | 0 | 0 | 0 | 0 |
| S3 5% | -2,706 | -2,271 | **-4,977** | 133 | 40 (-$158,331) | -14,286 | 130 | 27 (-$149,227) |
| S3 10% | -1,520 | -2,824 | **-4,344** | 105 | 30 (-$127,005) | -1,025 | 69 | 10 (-$25,685) |
| S3 20% | +74 | -273 | **-199** | 42 | 7 (-$18,314) | -2,134 | 16 | 4 (-$15,958) |
| S3 30% | -156 | -502 | **-658** | 14 | 5 (-$14,363) | +181 | 6 | 0 |

- Losers saved vs winners cut, S3 5% half A: +$67,180 saved on losers, -$158,331 lost on 40 coins that would have hit 5x. Same shape at every X.
- S3 30% in half B (+$181 a day, 6 fires, no winner cut) is the only positive cell, and it lost in half A; with 6 fires it is noise. Not chosen, not a lead.
- Base (current exit) on this engine: live A $2,605 / B $9,708 a day; kit300k A $2,153 / B $11,317 a day. Days: A 18.31, live B 5.57, kit300k B 6.40 (kit300k now runs to the data end).
- Random early exits: not run as a gate (no frozen cell); the chosen cell has 0 fires.

## Method (as pre-registered)
- Trades: my live wide-gate book (49 A, 22 B) and the A3.12 kit300k book (91 A, 121 B). Entry block as before, $1k, every fill by a full tick walk of in-range Uniswap v3 positions (tickwalk.py).
- Current exit: 0.3x stop, 5x target then KIT HOLD (50% trail, pull, 48 h), in-range half-pool pull, 12 h. Signals only act between entry and the current exit or the 5x; if the current exit fires by block b+1, nothing changes.
- S2 (bundle wallet sends to pool or router) not run: bundle tags exist only for the 41 coins that hit 5x (A3.13).
- Scripts on the Mac: ~/a3/a314.py (engine), ~/a3/a314e.py (eval), rows in ~/a3/data/a314_rows.pkl.

## Limits
- A slower real fill than one block only makes S3 worse (it sells further down the burst).
- One data set, two halves; kit300k is a paper book.

Version: v1, 07 Oct 2026 16:35 Bangkok, Analyst 3. Latest: http://159.223.88.168/hx-c9ois/sw/sw-a3-14-early-exit.md