what does lulu ablate do?
Asks whether a replica round tracks the live surface closely enough to reason from, and returns three answers rather than two, because a design that cannot tell has not passed.
how do you run it?
In the parser’s own words, whether a replica round may stand in for a live one, and why not. Every flag this build accepts, on one line:
lulu ablate --live LIVE --replica REPLICA --brand BRAND [--margin MARGIN] [--confidence CONFIDENCE] [--brands BRANDS] [--per-brand] [--ledger LEDGER]
lulu ablate --live ROUND --replica ROUND --brand BRAND --margin 5 ± the gate, at five points
what is a replica?
A live engine chooses its own sources, so what it retrieves and what it says cannot be held apart. A replica is the same model asked the same questions with the source list supplied, which makes the list something you can change. That is only worth anything if the replica behaves like the surface it stands in for, and this command is the gate the causal half of the product rests on, so it runs before anything is concluded rather than after.
why an equivalence test rather than overlap?
It asks whether the difference is provably smaller than the margin you plan at. Overlap would reward not looking, since two rates measured with too few calls overlap with everything.
So there are three outcomes, and this design cannot tell is one of them, printed with the number of calls that would settle it and a non-zero exit code. It is a real answer with a call count attached rather than a failure to produce one.
which rounds does it refuse?
A round whose answers did not all come from one engine. A round is one engine and one surface, so a file holding two was not written by one round: keyed by the question alone one engine's answers replace the other's, and keyed by the sample the two sides pair on nothing they share. Both are numbers about the collector, so such a round is refused and the engines it holds are named.