brain.
a snapshot of the linear model living on chain. four input features, four
weights plus a bias. updates land via train(). predictions
are clamped between the configured min and max LP fee.
total swaps observed-
training updates-
obs in buffer0/32
last validation loss-
flow EMA-
volatility EMA-
fee bounds-
learning rate-
weights
red bars = negative weight (feature reduces predicted toxicity). green = positive.
recent observations (latest first)
| # | block | size (f0) | dir (f1) | recency (f2) | flow (f3) | outcome | predicted |
|---|
feature reference
- f0 (size): a normalized log of the WETH side of the swap, in the range 0 to 1. larger trades push f0 up.
- f1 (direction): +1 if the swap is a buy (WETH in, SYN out) and -1 if it is a sell, so the model can pick up side-specific bias.
- f2 (recency): 1 if the swap is back-to-back with the previous one, decaying linearly to 0 as roughly a hundred blocks pass without trades. acts as a cluster detector.
- f3 (flow): an exponentially weighted directional volume over the last several swaps, picking up sustained buy or sell pressure.
- outcome: the surcharge-to-volume ratio actually realized on each swap, scaled, which is the value the model regresses against.