Keeping the model out of the verdict path
Measure which online communities actually produce paying customers, and withhold a verdict until the evidence supports one. *Madrugar* is Portuguese for rising before dawn.
The whole product rests on one rule: the verdict engine is deterministic and no
LLM may reach it. A verdict is withheld until three gates clear, three weeks
since entry, thirty cumulative clicks, real recorded effort. A kill
recommendation needs the same stage failing on two consecutive reads rather than
one bad week. Verdicts::Compute is the only writer of the verdicts table.
That last constraint is what makes the gates real instead of advisory. If a model can talk its way to a conclusion then the confidence gating is decorative and the product is a confident guesser with extra steps. Single-writer discipline is the enforcement.
It is spec-driven, and every threshold traces back to a written rule. That reads as overhead until someone asks why the click gate is thirty and not fifty, and the answer exists in writing.