The Evidence Ladder

miStudio can discover candidate circuits, but discovery is not proof. A cross-layer edge can be a strong statistical association and still not be a mechanism the model actually uses. So every circuit and every edge carries an evidence rung — a single, honest grade of how strong the evidence for it is — and the interface uses server-rendered language for each rung so nothing ever gets described more confidently than the evidence allows.

The four rungs

Rung	Name	What earned it
0	mined	Surfaced by statistical mining (PMI, support, null model, held-out replication). An association, disclosed as such.
1	attribution-supported	A gradient-attribution pass agrees in sign and magnitude. A second independent signal — still not causal.
2	causally validated	Intervention (suppress the upstream, measure the downstream) shows a real effect beyond a matched null (Feature 017).
3	faithfulness-tested	The circuit's members are shown necessary and sufficient for the behavior, not just individually implicated (Feature 017).

A circuit's rung is the minimum over its edges — a circuit is only as validated as its weakest link, and a circuit with no edges sits at rung 0.

Rules that keep it honest

• The word "causal" never appears below rung 2. Mined and attribution-supported edges are described as candidates and associations. Only intervention earns causal language, and the language comes from the server, not the client.
• A failed test never demotes. If an edge is tested and the effect doesn't hold, that tested_and_failed fact is recorded as history — it doesn't erase a rung the edge earned another way, and it's visible so you know what was tried.
• Promotion is a badge, not a gate. You can promote an unvalidated circuit into a steering profile — it just carries its rung visibly everywhere, so a rung-0 circuit is never mistaken for a proven one.

The ladder is why you can trust the Circuits panel: it tells you exactly how much is known, and it is structurally prevented from telling you more.