call-right-party-gatekeeper

Verify who is listening before saying what matters.

In collections, healthcare, and any sensitive outreach, disclosing account or medical details to the wrong listener is a compliance failure even when the call otherwise goes well. This skill audits finished CALL-E calls for the ordering question: did the agent confirm the recipient before the first sensitive disclosure keyword? This is an advisory English phrase check, not identity verification or legal-compliance certification.

When To Use

• after an authorized CALL-E call, to flag recognized confirmation/disclosure ordering for human review, not to prove identity or permission to disclose
• to detect wrong-party answers ("wrong number", "can I take a message") and third-party answers ("who is this calling?", "he is busy right now")
• to generate a verification-first goal for the next plan_call that parks politely and reveals nothing until identity is confirmed

When Not To Use

• to detect fraud or scam patterns in the conversation; use call-fraud-shield
• to audit general call quality, disclosure of being an AI agent, or result-vs-transcript consistency; use call-review
• during a call; CALL-E exposes transcripts, not live audio, so this is a post-call audit plus a pre-call goal template
• as proof of who actually answered; regex-level signals miss many phrasings and the card says so

Workflow

Analyze a finished call

python3 scripts/gatekeeper.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:

• right_party_status: CONFIRMED / WRONG_PARTY / THIRD_PARTY_PRESENT / UNVERIFIED
• verification_before_disclosure: ordering flag (false when disclosure preceded the first recognized callee confirmation or no confirmation was recognized; only detected phrases are compared)
• evidence: turn index, masked span, signal type for every detected verification question, identity confirmation, wrong-party signal, third-party signal, and sensitive disclosure
• gate_assessment: "unclear" with a reason when one side never spoke
• recommended_action: proceed, stop_and_retry_with_script (with the verification-first goal text), or human_review (with review guidance)

All action labels are review suggestions. proceed never authorizes sensitive disclosure or a consequential action; stop_and_retry_with_script never authorizes another call. The operator must independently verify identity, decide what can be disclosed and approve any new call.

Craft the verification-first goal

python3 scripts/gatekeeper.py craft --scenario sensitive-outreach --language en

Emits the plan_call inputs JSON whose goal is the same template the card recommends on stop_and_retry_with_script, so audit and next call stay consistent.

Practice grounding

Right-party contact verification and third-party disclosure limits are long-standing practice in collections and healthcare outreach (for example FDCPA third-party disclosure rules and HIPAA minimum-necessary handling). This skill illustrates a narrow phrase-ordering review for CALL-E transcripts; it is not a legal assessment and claims no research results.

Differences from sibling skills

• call-review audits call quality and compliance generally (including sensitive readbacks); this skill specializes in the single ordering question - was the recipient verified before content - and produces a ready-to-use retry script.
• call-fraud-shield detects scams; this skill assumes the caller is legitimate and audits its own disclosure hygiene.