the reflex next to one coin

I am not going to pretend this is a large protocol.

It is a small program next to one coin.

Most token fee systems decide what every trade should do before the first trade even happens. A quiet buy, a large sell and a market under stress can all produce exactly the same result.

I wanted the decision to happen closer to the trade itself.

SAYA watches the swap, builds a compact description of the market around it, and passes that state to Laya running locally.

The model is not asked to explain anything. It only has four answers available:

keep.
shield.
return.
cut.

The answer becomes the action. KEEP leaves the fee in reserve. SHIELD directs it toward liquidity. RETURN directs it toward holders. CUT directs it toward supply reduction.

Laya is useful here because it can return one constrained choice, a score and a confidence without writing commentary token by token. It is the small local brain inside the mechanism, not the project itself.

Running the read locally keeps the question close to the event. The state and answer can be recorded together. The result can be inspected without pretending that a paragraph of generated text is economic logic.

There is no timer. A read begins with a swap. If nobody trades, nothing moves. If the market changes, the next answer may change with it.

The reserve is where KEEP waits. SHIELD, RETURN and CUT account for the other possible destinations. Until execution is connected, those destinations remain modeled and are never presented as completed transfers.

Every read stays public: the trade, the compact state, the typed answer, confidence, latency and resulting reserve action.

If the model stops or returns something outside the four allowed answers, SAYA records a miss and does nothing. It does not guess quietly.

SAYA is an independent experiment using Laya as a local typed-decision model. There is no partnership or endorsement implied.