call-apology-effectiveness-auditor

call-apology-effectiveness-auditor

"I'm sorry you feel that way" is not an apology. It is a complaint about the person who had one.

Apology effectiveness is component-decomposable: research on real conflict apologies shows that a working apology is not one utterance but a stack of parts - acknowledgment of responsibility, concrete repair, brief explanation, expressed regret, forbearance, and (optionally) a request for forgiveness - and that acknowledgment carries the most weight. When a callee raises a grievance on a call, an agent that says only "I'm sorry about that" and moves on to scheduling has produced a rote apology: audible, and empty. This skill detects the grievance and grades the recovery response against that component stack.

When To Use

• after support, collections, or clinic calls where the callee may have raised a grievance (delay, broken promise, repeated contact)
• in QA pipelines, to find calls where the recovery was rote, deflecting, or missing entirely
• before recovery callbacks, to craft an apology-protocol goal that names the miss and offers concrete repair

When Not To Use

• as a measure of sincerity or relationship health; the components are surface-lexical markers - an agent can say "that's on us" with no feeling behind it
• to force apologies on every failure; sometimes the right posture is no apology at all, because the agent cannot own fault it does not have
• during a call; strictly post-call analysis plus pre-call goal crafting
• on non-English transcripts; the lexicons are English-only

Workflow

Audit a finished call

python3 scripts/apology_effectiveness_auditor.py analyze \
  --transcript path/to/call-result.json

Reads the real get_call_run result shape ({status, result: {transcript}}) or the flat fixture shape used by sibling skill fixtures. Emits a card:

• grievance_turn_index / grievances_total / grievance_evidence: where the first grievance was voiced (callee turns only), how many grievance-bearing callee turns exist, and the masked evidence sentence
• apology_detected / apology_evidence: whether any agent turn from the grievance onward contains an apology, and the masked sentence
• components_present: the six-component markers found in the graded window (the apology sentence plus the following agent turn, since explanations often arrive one turn late)
• typology: rote / empathic / explanatory (empathic > explanatory

  rote)

• verdict: EFFECTIVE_APOLOGY (acknowledgment plus repair or explanation) / PARTIAL_APOLOGY / NON_APOLOGY (deflection override) / GRIEVANCE_UNADDRESSED / NO_GRIEVANCE, plus unclear paths (empty transcript, fewer than two turns)