Narrative flow profile

Support response agent

Evaluator maturity pipeline for support replies.

Abstract pattern

Evaluator maturity pipeline

Shows how an expensive evaluator can be replaced by a validated classifier while the system producing replies stays the same.

  1. Manual Instrumented boundary; human performs the governed effect.
  2. HITL Automation proposes or prepares the effect and a human authorizes it before execution.
  3. HOTL The effect executes with asynchronous human oversight.
  4. Autonomous The effect executes under declared evaluator and demotion policy.

Narrative task profile

Support response agent

Shows how an expensive evaluator can be replaced by a validated classifier while the system producing replies stays the same.

Boundary
Input is the ticket, customer context, and knowledge base. Output is a response draft.
Evidence log
Ticket, draft, judge verdict, human edit or approval, and customer reaction such as resolved, reopened, or escalated.
Evaluator
Property
response_quality
Target
Output
Technique
LLM judge first, then a classifier distilled from the judge's labels once enough examples accumulate.
Evaluation basis
human_gold: support_agent_edit_and_approval_human_gold
Epistemic status
ground_truth
Version
1
Authority
blocking
Placement notes
hitl
blocking, every draft is held for human approval until judge agreement and reopen rate are proven stable
hotl
sampling, sampled monitoring replaces blocking review once the judge meets the bar; novel and sensitive tickets remain gated
Promotion rule
Judge agreement with human approvals stays above threshold and reopen rate stays low, which allows sampled monitoring instead of blocking every send.
Demotion
Rule
Reopen or escalation rate rises, or the distilled classifier drifts from fresh human gold, which demotes both the task and the evaluator.
Fallback
Human approval stays in place for anything outside the routine ticket classes.

Teaching point

What this flow proves

Eval-the-eval is the point here, because the evaluator is validated, promoted, demoted, and eventually made cheap enough to own the gate.

Teaching context

Why this flow is governed this way

Worker context
The LLM that drafts the response.
Without PAA
Responses go out without a systematic quality gate; the model's quality is invisible until a customer complains; there is no audit trail distinguishing what the model produced from what a human approved.
Never full-auto
Anything outside the routine ticket classes, novel, sensitive, or escalating tickets stay human-gated as a permanent design constraint, not a classifier gap.

This page is linked from the canonical card set on the flows index.