# A3.3: early fail exit and pull warning (6 Oct 2026, Bangkok)

## Read this first
- **No early exit is worth switching to.** The best one adds about +0.03x and touches only 4 to 6 trades. That is noise-sized. Keep the plain rule (5x / -40% / range-aware pull / 12 h).
- **Why it doesn't work: losers look like winners in the first hour.**
  - At 15 min the median loser is at 1.07x of entry and the median winner at 1.21x.
  - Losers do not fade early. They rise first: their median high by 30 to 60 min is 1.84x. Then they crash through the stop.
  - So "exit early if weak" catches very few of them. "Lock in profit after a rise" (move the stop up after 1.5x to 3x) cuts winners and loses money in every setting (1.23x to 1.96x).
- **Pull warning: there is a real tell, but trading on it costs money.**
  - A sell of 5% or more of the pool's depth comes in the 10 minutes before 69% of pulls. In ordinary active moments it shows up only 11% of the time.
  - Used as an exit on the 58 trades, it loses money in every setting (1.32x to 1.95x vs 2.00x), because winners get the same big sells on the way up.
- Base for everything below: my range-aware rebuild of the 58 trades (A3.1), **1.998x** (first half 2.112x, second half 1.897x).

## 1. What separates the 19 to 20 winners from the stops (median at each checkpoint after entry)
| Min after entry | Losers: price vs entry, high so far | Winners: price vs entry, high so far | Buyers so far (L / W) |
|---|---|---|---|
| 5 | 1.02x, 1.23x | 1.10x, 1.26x | 107 / 139 |
| 15 | 1.07x, 1.64x | 1.21x, 1.67x | 213 / 284 |
| 30 | 1.18x, 1.84x | 1.63x, 1.89x | 359 / 425 |
| 60 | 1.22x, 1.84x | 2.40x, 2.93x | 663 / 821 |
- Sell share is about 50% of volume for both groups at every checkpoint. The flow is mostly bots churning both ways, so "sell pressure" carries no information here.
- Winners and losers only pull apart after 30 to 60 min, and by then most losers have already stopped out.

## 2. Early exit grids, money result on the 58 (first half / second half / all)
| Rule | A | B | All | Winners cut | Losers cut |
|---|---|---|---|---|---|
| Plain rule | 2.112 | 1.897 | 1.998 | | |
| Exit at 15 min if price is 10% or more under entry | 2.138 | 1.925 | 2.024 | 0 | 4 |
| Exit at 15 min if it never went above 1.2x | 2.120 | 1.948 | 2.028 | 0 | 6 |
| Exit at 15 min if under entry and net selling in last 10 min | 2.151 | 1.925 | 2.030 | 0 | 5 |
| Exit at 5 min if 10% under | 2.124 | 1.784 | 1.942 | 1 | 7 |
| Exit at 15 min if never above 1.5x | 1.578 | 1.645 | 1.614 | 7 | 10 |
| Lock: after 2x, stop moves to entry | 1.918 | 1.822 | 1.867 | | |
| Lock: after 3x, stop moves to 1.2x | 2.149 | 1.798 | 1.961 | | |
- The three best rows are each +0.03x and each depends on 4 to 6 trades. Paper variant at most: **"exit at minute 15 if 10% or more under entry"** (`w2_v3_e15`). Do not change the main book.

## 3. Pull warning
- **What a pull is here:** one in-range Burn that removes 50% or more of active depth. 178 of the 296 study coins were pulled. The median pull comes 2.8 h after the first swap (p10 0.3 h, p90 59 h).
- **Who pulls:** the pull tx sender is a different wallet for every one of the 178 pulls. No wallet pulled 3 or more coins. Wallet memory cannot predict pullers.
- **Signals in the 60 min before the pull, against the same check at random active moments** (20+ swaps in the hour, at least 2 h before any pull):
  | Signal | Before pulls | Random active moments |
  |---|---|---|
  | A sell of 5% or more of depth in the last 10 min | 69% | 11% |
  | Same, in the last 60 min | 90% | 23% |
  | A small "test" removal in the last 60 min | 3% | 2.5% |
  | Buys in the last 10 min under 30% of the earlier pace | 23% | 39% |
- **Lead time:** when the big sell comes, it is a median 8.5 min before the pull (p25 5.4 min).
- **Money test, sell when it fires** (5% / 8% / 12% / 20% of depth):
  - 1.32x / 1.60x / 1.80x / 1.95x;
  - only when price is under entry: 1.57x / 1.81x / 1.85x;
  - all worse than 1.998x.
- **Why it fails:** insiders dumping a big bag before the pull looks the same as a big holder taking profit on the way to 5x.
- The pulls in our own 58 are only 2 trades (after the A3.1 range fix), so the most a perfect warning could save is about +0.03x.

## 4. What would change this verdict
- A sell tell that is specific to the LP owner (the Mint owner's own token sells or ETH moves) instead of any big seller. That needs the LP NFT owner per pool. That is a PP data job (NFT Transfer of the position plus the owner's wallet tape). It is not worth it for +0.03x unless the book grows.

## First principles score
- 8/10. The checkpoints use only data up to that minute, and exits are priced from range-aware depth.
- The sample is 58 trades, so any gain under about 0.1x is inside the noise. That is the reason for "no change".

Version 1, 6 Oct 2026 14:55 Bangkok, Analyst 3. Latest: http://159.223.88.168/hx-c9ois/sw/sw-a3-3-early-exit.md