DashboardSign inStart your trial

Modes & Confidence

Approval modes, how confidence scoring works, and the spot checks and safety holds that route cards to you

Two things decide whether a proposal reaches your queue or clears on its own: the mode a Human Review node runs in, and the confidence score each run earns. This page covers both, plus the spot checks and safety holds that can route a card to you regardless.

#Approval modes

Each Human Review node runs in one of two modes:

  • Always ask: every run pauses for review. Use for high-stakes actions: invoices, deletions, anything where the cost of getting it wrong is high.
  • Confidence gated: high-confidence runs approve themselves; anything below your confidence bar routes to your queue. The default bar is 85%. Lower it as you build trust.

Modes are per-node. A single workflow can have one step on Always ask (the email send) and another on Confidence gated (the routing decision).

Changing the mode on a node doesn’t affect proposals already in your queue; those stay in whatever mode was active when they were created.

#Confidence scoring

Confidence is calculated per run, not stored as a workflow property. The same step can score 95% on one run and 62% on the next: the score reflects the input data and the system’s pattern history, not the step itself.

The score is built up per input. Every context field on the card is scored on its own:

  • Deterministic inputs (a fixed value, a field passed straight through) score a perfect 1.0.
  • Non-deterministic inputs (an AI-generated draft, a classified value) get a 0–1 score with a short reasoning note you can read on the card.

Deterministic checks run as gates on top of the scores: validation (does the value match its expected shape), anomaly (is it in the normal range for this step), and containment (is it within allowed bounds). A failed gate trumps the score and holds the card no matter how high it scored.

The card’s overall score is a weighted combination of its input scores. In Confidence gated mode, the card also gets a final coherence sanity check across all its inputs. If the card scores at or above your confidence bar and clears every gate, it auto-approves; otherwise it comes to your queue.

The system needs some history before scores are meaningful. Brand-new steps lean on the deterministic checks; once you’ve made a handful of decisions, scoring reflects your preferences more closely.

#Spot checks and safety holds

Even when a step is auto-approving reliably, Rills routes the occasional above-the-bar card to your queue anyway as a spot check, about 1 in 10, so you keep eyes on autonomous decisions without reviewing every one. These aren’t configurable; they’re a built-in audit.

Two other safety holds can route a card to you even in Confidence gated mode:

  • Circuit breaker: after you reject a card on a step, auto-approve pauses on that step and the next runs come to you for review, until the step re-earns trust.
  • Scoring errors: if the system can’t score a card for any reason, it holds it for a human rather than auto-approving.

Each held card shows a short chip explaining why it came to you.

Modes and confidence decide whether a card reaches you. Who it goes to, how long it waits, and how you’re notified live on the same node. See Routing, Timeouts & Notifications .

#Related