A promotion ladder that never demotes

We extracted claims from conversations and needed to say how confident we were in each one.

The ladder sits in the middle of a three-layer model: learnings, per-conversation insights pulled out of a transcript with evidence rows pointing back at the exact segments; signals, the same insight once it recurs across conversations, clustered by cosine similarity; and artifacts, thirty-one generated documents that only exist once their prerequisite signals do. Everything downstream of transcription exists to move a fact from the first layer to the third.

The ladder is four stages — candidate, emerging, validated, decision-grade — with evidence thresholds between them: three pieces to reach emerging, five across two distinct conversations to reach validated, ten across three for decision-grade.

The constraint that made it work: signals never demote. There is no path back down.

That sounds like a limitation. It is the feature. A stage becomes a fact about evidence accumulated, not a value that flaps between runs as clustering shifts underneath it, so every downstream consumer — gating, tools, dashboards — can cache against it and trust it. It also collapsed a graph problem into arithmetic: once stages only ever increase, "are this artifact's prerequisites met" is a threshold comparison answerable in two queries, not a recursive traversal over a prerequisite chain.

The thresholds live in a pure-logic package with no database and no environment, shared by the worker, the app and the tests — one place to change a number.

The cost cuts the same way the benefit does: a signal promoted on evidence that later turns out wrong has no path back down either. We accepted that because the thresholds are deliberately conservative — a false positive misleads someone, a false negative just means something hasn't generated yet, and those costs aren't symmetric. Instead of a demotion path we kept an audit log recording every promotion and its triggering evidence, so a wrong signal can at least be explained. Clustering isn't perfectly deterministic either — nearest-neighbour ties break on row order — tolerable only because the thresholds absorb a shift that small.