Bands

Confidence bands → deterministic control actions

Bands are policy-friendly semantics over a continuous confidence signal. They prevent accidental “threshold roulette” and make governance stable across model upgrades and provider churn.

Signals Governance

Why bands exist

A single confidence number is not a control policy. Bands translate confidence into operational semantics: safe to automate, warn, route to review, fallback, or block.

Bands are stable; thresholds are policy

Qubanta keeps band meanings stable; organizations choose thresholds per context (task type, risk class, environment). Stability beats “per-model tuning.”

Bands do not imply correctness probability

Bands describe reliability posture, not truth probability. They are designed for routing and safety, not epistemic claims.

Default band map (v1 baseline)

These defaults can be tightened for clinical / regulated workflows and loosened for low-risk exploration. The semantics remain unchanged.

high       C ≥ 0.80
moderate   0.55 ≤ C < 0.80
low        0.35 ≤ C < 0.55
critical   C < 0.35

Action table (deterministic)

Bands map to actions, then policy refines behavior using state + flags.

band=high:
  action = accept

band=moderate:
  action = accept or warn (if non-critical flags exist)

band=low:
  action = review_recommended or fallback (task-dependent)

band=critical:
  action = blocked or fallback (never silent accept)

Collapse rule (non-negotiable)

If constraint_integrity == 0, then confidence collapses to 0 and action becomes blocked regardless of other signals.

State refinement

If state is unstable or drift_suspected, policy can override “moderate” into review routing. State is an operational risk class.

Flag preservation

Critical flags must remain explicit (schema_invalid, policy_violation, missing_required_field). Policies can escalate actions, but cannot erase flags into generic “low confidence.”

Example governed events

(1) Extraction with missing field (collapse)
signals.constraint_integrity = 0
confidence = 0.00
band = critical
state = constraint_invalid
flags = ["missing_required_field:id"]
action = blocked

(2) Retrieval QA with weak grounding
constraint_integrity = 1
grounding_strength = 0.31
stability_index = 0.88
system_health = 0.92
confidence ≈ 0.28
band = critical
state = weak_grounding
flags = ["retrieval_gap"]
action = fallback  (or review_recommended per policy)

(3) Tool routing under instability
constraint_integrity = 1
stability_index = 0.42  (high variance)
grounding_strength = 0.79
system_health = 0.90
confidence ≈ 0.30
band = critical
state = unstable
flags = ["high_variance"]
action = review_recommended

Operator playbook (v1)

This is the minimal incident-response posture for governed inference.

If action=blocked:
  - Treat as a contract failure (schema/policy). Fix constraints or input envelope.
  - Do not retry blindly; ensure required fields, schemas, and invariants are correct.

If action=fallback:
  - Switch to safer mode: smaller scope, stricter constraints, alternate provider, or human review.
  - Log: request_id, flags, and signal vector. Compare to baseline runs.

If action=review_recommended:
  - Route to human verification; preserve flags and evidence pointers.
  - If repeated: inspect stability and system_health; investigate drift or dependency degradation.

If action=warn:
  - Permit controlled automation only in low-risk contexts.
  - Monitor for drift_suspected; tighten thresholds if warnings cluster.

If action=accept:
  - Still log the governed event; acceptance is not proof of correctness.

Semantics Developer contract