calle-ai/awesome-phone-call-agents/call-ai-disclosure-comprehension-auditor

call-ai-disclosure-comprehension-auditor

Saying "this is an AI" is step one. Knowing it landed is the actual job.

call-review checks that a disclosure exists somewhere in the agent's turns. This skill audits the full sequence the law and the research both point at: disclosure in the first breath, an explicit comprehension question before any pitch, and a captured acknowledgment - and it writes the opening that produces that sequence next time.

When To Use

• after any CALL-E outbound call, to grade its disclosure posture
• before acting on a result where the callee's agreement matters - an

agreement from someone who never knew an AI was speaking is fragile

• before placing calls, to craft an Article-50-style disclosure opening

When Not To Use

• to certify legal compliance for any jurisdiction; the verdicts are

transcript observations, and regulators - not scripts - decide compliance

• to prove the person did not understand; silence after a disclosure is

not evidence of confusion, and the card says so

• during a call; this is strictly post-call analysis plus pre-call goal

crafting

Workflow

Audit a finished call

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

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

• disclosure_present / disclosure_early (first agent turn vs later)
• comprehension_check_present - an explicit question ("do you

understand", "is that okay", "does that make sense") in the disclosure turn or the next agent turn, BEFORE business content: if digits (dates, amounts, times) already appear in the disclosure turn, the pitch came first and the check does not count

• acknowledgment_captured - an affirmative callee reply in the first

callee turn after the check

• evidence: the disclosure / check / acknowledgment turns, masked
• verdict: FULL / PARTIAL_NO_ACK / DISCLOSED_NO_CHECK /

LATE_DISCLOSURE / UNDISCLOSED, plus unclear paths for empty or agent-less transcripts

Only agent turns can disclose; a callee asking "are you a robot?" is not disclosure and does not count.

Craft the disclosure-first goal

python3 scripts/disclosure_comprehension_auditor.py craft --scenario disclosure-first-call

Emits the plan_call inputs JSON whose goal puts the disclosure in the first sentence, pairs it with one comprehension question, and refuses to proceed to business until the person acknowledges.

Scientific and Regulatory Foundation

Citation notes recorded during verification: the two Chen-lineage papers are distinct (a 2024 individual study in CHB with the exact DOI above, and a 2026 IP&M meta-analysis); Article 50 obligations apply from 2 August

1. California's B.O.T. Act (SB 1001, 2018) covers bot disclosure in

the US and is consistent with this design but is not separately cited in the table. This skill implements a lexical approximation and labels every output analysis_mode: "heuristic".

Differences from sibling skills

• call-review checks disclosure presence as one compliance line; this

skill grades the full disclosure-comprehension-acknowledgment sequence and writes the compliant opening for the next call.

• call-sycophancy-guard protects factual integrity under pressure; this

skill protects the conversational precondition - the person knowing who is talking - under which any agreement is meaningful.