Adding trust, one signal at a time

Every signal below is optional and additive. A pipeline with none of them turned on is a legitimate, supported way to run VectorStep — that’s rung 0, and plenty of pipelines never leave it. Climb a rung when the cost of the agent being wrong goes up, not because a rung exists. Most pipelines should stop climbing well before rung 6: once the risk of the action no longer justifies the next rung, stop.

This page is a ladder, not a reference. Each rung explains why you’d bother before it says how — the technical detail lives on the linked page.

Rung 0 — Nothing

Turn on: nothing.

The problem it solves: none, deliberately. This is the baseline — a working pipeline that triggers on a webhook, runs a step, and does whatever the step’s executor does. No confidence check, no verifier, no grounding, no calibration, no readiness gate.

If the action a step takes is low-stakes or already reviewed by a human downstream, this is where you should stay. Read the pipeline schema to see the shape of a pipeline with nothing turned on.

Rung 1 — A confidence floor

Turn on: confidence_threshold and on_low_confidence.

The problem it solves: the agent already reports a confidence number on every response, whether you use it or not. Rung 1 just stops ignoring it — below the threshold, the step escalates, aborts, or proceeds anyway, depending on on_low_confidence, instead of the response being trusted unconditionally.

This is the cheapest rung to climb — no new agent call, no new config block, just a bar and a threshold. See how confidence and calibration work for what the number means and why it’s a starting point, not a verdict.

Rung 2 — A second opinion

Turn on: verifier.

The problem it solves: the agent grading its own work is not evidence — a model that got the task wrong will often say so confidently. A verifier is a second agent call that sanity-checks the primary response before its confidence is trusted, and it can only lower that confidence, never raise it.

Climb this rung when a wrong answer would be expensive enough that a second, independent look is worth the extra call. Read Verifiers for the trigger bands and combination strategies.

Rung 3 — Grounding

Turn on: grounding.

The problem it solves: plausible-sounding prose isn’t evidence. Grounding checks whether every load-bearing claim in a step’s output is actually backed by a tool call or tool result in its own trace — not just phrased confidently.

Climb this rung when the step’s output makes factual claims you’d want challenged in an incident review — “the deploy succeeded”, “the ticket was created” — rather than claims that are inherently a judgement call. See Grounding.

Rung 4 — Deterministic checks

Turn on: deterministic checks.

The problem it solves: some facts aren’t a judgement call at all, and don’t need a model in the loop to settle. A deterministic check runs a real command or query and treats the result as ground truth — one failed check forces trust to zero, no averaging, no partial credit.

Climb this rung when there’s a fact you could check with a shell command or an API call instead of asking a model — “is the alert still firing?”, “does this record still exist?”. Deterministic checks are documented alongside grounding: see Grounding.

Rung 5 — Calibration

Turn on: calibration.

The problem it solves: an agent saying “90% confident” doesn’t tell you whether its 90% has historically meant 90%. Calibration bins every marked outcome per agent, model, provider and prompt version, and measures what each confidence band was actually worth — in advisory mode with zero effect on behaviour, or enforced once you trust the bins.

Climb this rung once a step has enough run history to calibrate against, and you want the measured track record to matter more than the agent’s own self-report. See Calibration.

Rung 6 — Readiness

Turn on: readiness.

The problem it solves: promoting a pipeline from testing to production is usually a gut call. Readiness replaces the gut call with owner-defined criteria across operational, confidence, accuracy and calibration tiers, judged against real evidence before a pipeline earns production traffic.

This is the top of the ladder — climb it for pipelines whose production promotion should be earned, not asserted. See Promotion readiness.