The cycle
Every five minutes the watcher pulls freshly surfaced pairs from public market data and measures each one: age, pooled liquidity, hourly volume, swap count, the split between buys and sells, hourly volume as a multiple of liquidity, how much of a run-up has been handed back, and whether the token publishes a site or a social at all.
Those measurements run through the gates listed opposite. A gate is a threshold in code. It returns a boolean and a sentence explaining itself, and nothing downstream can override it — not the model, not a good story, not a chart that looks ready.
Grades and size
A pair that clears every gate with room to spare grades A. Clearing every gate on smaller size grades B. One failed soft gate grades C, which is probe-only. A failed liquidity or age check is fatal and the pair grades below the bar, whatever else it has going for it. The grade sets the maximum position, and the maximum position is the only number the model is not allowed to influence.
What a call is
When a candidate grades A or B and the desk has an open slot, LUX publishes a call: the symbol, the entry price at that instant, the size the rails allow, a thesis and the single observation that would prove it wrong. The call is then hashed. That hash — the seal — is published with it.
The seal covers the timestamp, symbol, pair address, entry price, grade and thesis. Anyone can recompute it from the public record. If any of those fields were edited afterwards, the seal would stop matching, so the record cannot be quietly improved.
Marking
At one hour, six hours and twenty-four hours after publication, the desk fetches the pair's current price and records the move. A move beyond three percent in either direction is a decision: right or wrong. Anything inside that band is flat, and flat counts as neither. Once the twenty-four hour mark lands, the call settles and its verdict is final.
If a pair stops reporting before it can be marked, the call is voided rather than scored. A dead pair is a real outcome and it is shown as one, but it is not counted as a win.
What the model does and does not do
The narrator receives the measurements as plain text and writes the reasoning you see in the log. It never sees a private key, never chooses a size, never moves a threshold and never learns what happened after a call. If no model is configured, a deterministic writer takes over and the log says so on every line, because a desk that hides which mouth is talking is not transparent.
Execution
Execution is off. Calls are published, not placed. That is a deliberate starting position: a record has to be worth something before capital follows it, and this desk would rather be judged on whether it was right than on whether it was lucky with size. When execution is armed, this page will say so, and the caps opposite will be the caps that apply.
Where the numbers come from
Pair data comes from public decentralised-exchange market feeds. The desk holds no private data source and no privileged access. Every figure in the log can be checked against the same public endpoints the watcher reads.