calle-ai/awesome-phone-call-agents/call-answer-evasion-auditor

call-answer-evasion-auditor

A callee who asks a straight question deserves a straight answer.

On a voice call, a dodged question is louder than a wrong one. A caller who asks "are you a robot?" and hears "great question!" stops trusting every sentence that follows; a caller who asks "how did you get my number?" and hears a counter-question hangs up. Evasion destroys more calls than imperfect information does, because it is audible as evasion. This skill audits the agent's answers to every direct question the callee asked.

Grading is lexical, derived from the QEvasion response-clarity taxonomy: each callee question (identity, yes-no, or information/wh) is graded against the next two agent turns as CLEAR, PARTIALLY_CLEAR, or EVASIVE, with a mechanism label for every non-clear answer. It is wording analysis, not intent detection.

When To Use

• after any CALL-E call where the callee asked direct questions and the

transcript should show candid answers

• when reviewing whether identity ("are you a robot?"), provenance

("how did you get my number?"), or commitment questions were answered or dodged

• before placing calls, to craft an answer-first goal template with a

canned truthful identity answer and a provenance line

• alongside call-repair-sequence-auditor, which owns bare repair

initiators ("Sorry, what was that?") that this skill skips

What It Checks

Questions asked by the callee only; agent questions are never graded.

Question kinds: identity (mentions robot/AI/automated/real person), yes_no, wh (who/what/when/where/why/how/which/whose/whom - also the fallback when ASR drops the "?").

Grading Rules

The answer window is the next TWO agent turns after the question turn, stopping early if the callee asks a new question; no window at all grades EVASIVE (no_response). Identity questions are CLEAR only on a truthful AI/self disclosure; "I don't know" never clears them. Explicit not-knowing ("I am not sure", "I don't know") clears any non-identity question; identity questions are never cleared by not-knowing. Yes/no questions are CLEAR only when the yes/no token (an explicit negative such as "I am not" or "we won't" counts) leads the first sentence of the first window turn; a token found later in the window grades PARTIALLY_CLEAR. Information (wh) questions are CLEAR on provenance wording, a plain not-knowing statement, or a concrete amount. When one callee turn asks several questions, each question sentence is graded individually: a clear answer to one never lifts an evasive answer to another. All lexicons match uncontracted forms too ("we will", "i am").

Research Grounding

• Thomas, Filandrianos, Lymperaiou, Zerva, Stamou - "'I Never Said That': A dataset, taxonomy and baselines on response clarity classification" - Findings of the ACL: EMNLP 2024 - https://aclanthology.org/2024.findings-emnlp.300 - the QEvasion dataset and its two-level clarity taxonomy, which our CLEAR/PARTIALLY_CLEAR/EVASIVE grades derive from.
• Thomas, Filandrianos, Lymperaiou, Zerva, Stamou - "SemEval-2026 Task 6: CLARITY - Unmasking Political Question Evasions" - arXiv:2603.14027 (2026) - a shared task validating evasion grading as a benchmark problem.
• Ma, Lin, Yang - "EvasionBench: A Large-Scale Benchmark for Detecting Managerial Evasion in Earnings Call Q&A" - arXiv:2601.09142 (2026) - three-level evasion grading in a second domain.
• Kim, Jeong, Kwak - "A Shaky Voice Is Not Always a Dodge: Benchmarking Textual and Vocal Evasion Detection in Earnings Calls" - arXiv:2608.28040 (2026) - DualEvasion; text cues, not vocal ones, carry the signal, which grounds our text-only posture.

Limitations

• English-only lexicons; wording-based, so an EVASIVE grade is evidence

of wording, never of intent to deceive.

• ASR frequently drops the "?" token; the wh-word start fallback

catches most information questions, but yes/no questions reduced to statement word order may be missed.

• Bare repair initiators ("What?", "Sorry, what was that?") are

delegated to call-repair-sequence-auditor and skipped here, whether they form the whole turn or one sentence inside a longer questioning turn; the remaining question sentences in that turn are still graded.

• Amount detection reads $<digits> and "<digits> dollars" only;

number-word amounts ("twenty dollars") are not parsed. An amount anywhere in the window clears a wh question even when it does not answer the question asked.

• Social wh openers like "What's up?" are not in the repair-initiator

skip list and are graded like information questions.

• Provenance matching is deliberately generous: any record-attribution

phrase such as "your reservation shows..." clears a wh question.

• turn_index refers to the normalized turn list (after non-dict

entries are dropped), not the raw transcript array.

Usage

python3 skills/call-answer-evasion-auditor/scripts/answer_evasion_auditor.py analyze --transcript <call-result.json>
python3 skills/call-answer-evasion-auditor/scripts/answer_evasion_auditor.py craft --scenario booking-candid

analyze accepts the real get_call_run result shape ({status, result: {transcript}}), the flat fixture shape, or a bare string transcript, and exits 2 on missing/invalid input. See references/examples.md for real outputs against the shipped fixtures and references/safety.md for scope and data-handling notes.