# A2.19: the sell cascade, is there a warning? (6 Oct 2026, 16:01 Bangkok)

## Read this first
- **No warning works. The cascades are a fixed cost; keep them in every number.**
  - I tested 9 warnings that a live bot can see before the cascade block, on all 70 wide-cell trades at $1k with real depth.
  - None raises $ a day in both halves. The best one (net sells in the last 60 s ≥ 50% of the pool's ETH) gives −$49 a day in the first half and +$53 in the second. That is noise.
  - The warnings that catch the cascades also cut 6 to 26 winners.
- **Under the live 0.3x stop the cascades cost little.**
  - Pricing every 0.3x exit at the END of its block (what live gets) moves the book from 2.06x / 2.10x to 2.06x / 2.09x at $1k. That is +$12 a day in the first half and −$57 a day in the second.
  - 4 of 33 stops cascade: the Coder's 3 plus 0xfe0f309cee96912d9926b8f0786fe3347364e678 (block 79896014). They end at 0.07x to 0.14x instead of 0.28x to 0.29x. Together that costs 0.69x of stake, about −$77 a day at $1k in the second half.
  - The other 29 stops end their block within 0.04x of the trigger.
- **All 4 cascades are in the second half (29 Sep to 5 Oct), and none in the first.** This may be new behaviour. A2.11 should count them live; if more than 1 in 8 stops cascade, re-check.

## What a cascade is (from chain)
| Coin | Block | Sells in the block | Who sold | Their history |
|---|---|---|---|---|
| 0x2a6efa703d8f1f06da07d1960711f02b1ba66f10 | 75017642 + 75017643 | 208 (two blocks, 0.1 s apart) | 208 different wallets, all through SwapRouter02 0xcaf681a66d020601342297493863e78c959e5cb2 | each bought once from the pool, before and after the gate; **none ever appears in any of the other 299 coins** |
| 0xf5a9da666eb257c0859b47efe8c2cb6745baff94 | 80503986 | 119 | 119 wallets, same router | same: one buy each, never seen in any other coin |
| 0xfe0f309cee96912d9926b8f0786fe3347364e678 | 79896014 | 122 | 122 wallets, same router | same |
| 0x89262fb9d6b05de797c99a927994f7c2411469d4 | 75156154 | 18 | ONE wallet, router 0xf2a277490c6e1d511e236c64895a0fef39aa8098 | it collected the tokens through about 200 transfers from 0x7dde40ad95…, a wallet active in 114 other coins |

- Three cascades are **fresh-wallet farms**: 120 to 200 single-use wallets bought one at a time and then sold in the same 0.1 s. The fourth is one wallet that gathered the tokens and sold them in 18 chunks.
- No seller wallet is shared between the cascades. So no wallet memory can catch the next one.
- The flow before the cascade looks normal. 0xf5a9da666eb257c0859b47efe8c2cb6745baff94 had 2 sells and 2 buys in the 60 s before, at 0.93x of entry. 0x2a6efa703d8f1f06da07d1960711f02b1ba66f10 and 0xfe0f309cee96912d9926b8f0786fe3347364e678 were busy two-way markets, with buys about equal to sells.

## Gate-time check (Coder's idea: skip farm coins at the gate)
- Measured at the gate on all 70 coins:
  - share of buyers that are fresh single-buy wallets;
  - the supply they hold;
  - how evenly sized their buys are.
- **The 4 cascade coins sit in the middle of every one of them** (fresh single-buy share 0.25 to 0.45; all coins median 0.28; 5x winners median 0.33). Winners look the same. No gate filter is possible from token data.
- One test was not possible: whether the 120 to 200 wallets share one ETH funder. The blockscout API blocks the container, and a normal RPC has no per-address history. That is the only lead left. It needs an address-history source (the PP data layer or a paid indexer).

## Warnings tested (all trades, $1k, real depth, sell in the block after the warning)
| Warning | Fired on trades | Cascades caught | Winners cut | $ a day first / second half |
|---|---|---|---|---|
| none (live 0.3x stop, block-end fills) | | | | +2,730 / +4,099 |
| net sells 60 s ≥ 10% of pool ETH | 65 | 4 | 26 | −202 / +1,241 |
| net sells 60 s ≥ 20% | 50 | 2 | 17 | +1,226 / +2,298 |
| net sells 60 s ≥ 30% | 34 | 1 | 9 | +2,571 / +2,490 |
| net sells 60 s ≥ 50% | 11 | 1 | 1 | +2,681 / +4,153 |
| ≥10 sells in 60 s and 2x the buys | 33 | 2 | 11 | +2,346 / +2,753 |
| ≥20 sells, 2x buys | 24 | 1 | 6 | +2,932 / +3,459 |
| ≥40 sells, 2x buys | 12 | 1 | 3 | +2,940 / +3,524 |
| price under 0.5x of in-trade peak and under entry | 44 | 3 | 10 | +1,927 / +2,859 |
| price under 0.4x of peak and under entry | 37 | 2 | 8 | +2,311 / +2,729 |

- First principles: a cascade is many sellers acting inside one block. Nothing on chain has to happen before it, so a price or flow warning can only guess. The guesses fire mostly on normal dips that recover.
- Table: http://159.223.88.168/hx-c9ois/sw/sw-a2-19-warning-grid.csv

## For the Coder / Supervisor
- **Keep the cascade cost in the numbers:** price stop exits at the end of the trigger block, as the Coder's real-swap pass does. At $1k the wide + 0.3x book is then 2.06x / 2.09x.
- **0xa2508b15ab16826c720631b9afa0a90c453ae8a4** (the Coder called A2 the outlier at 0.027x).
  - In my events, ONE swap (block 70225887, log 99) takes the cap from 1.91x of entry to 0.028x and the in-range ETH to 0.17x of entry, with nothing in between. No exit can beat that.
  - Coder or Analyst 3: please list your events between blocks 70225845 and 70225887 for pool 0xaf69d9d9d0dbeaaa2e03eddbd7971d5b536627ba. If yours show a swap I don't have, my data has a hole.

## How
- Events: every Swap, Mint and Burn of the 70 pools (sw3), with the router (Swap sender) and the tx. Cascade sellers come from the token's Transfer logs in the cascade blocks, read from chain (465 transfers). Their history comes from the pre-gate and post-gate transfers of all 300 coins.
- Exits: 0.3x instant stop, 5x, in-range pull, 12 h, $1k tick-walk fills (A2.14). Warnings use only data up to the block they fire in; the sale is priced at the end of the next block.
- Scripts: container /tmp/w2/a219 (ev.py, build.py, blockend.py, casc_tr.py, gfeat.py, warn.py, wgrid.py).

## First principles score
8/10. Every cascade was traced to its sellers on chain, and every warning was priced on real depth in both halves. Weak point: the ETH funding of the farm wallets was not read, because no address-history source was reachable.

Version 1, 6 Oct 2026 16:01 Bangkok. Latest: http://159.223.88.168/hx-c9ois/sw/sw-a2-19-cascade.md