research-gap-call-verifier

Research Gap Call Verifier

Use this skill after research, not instead of it. It separates facts already supported by cited sources from material gaps that a business can answer, produces an exact no-call preview, and reconciles returned evidence with honest status labels.

The bundled scripts use only the Python standard library. They never contact CALL-E, a phone provider, or the network. A host may execute an approved plan with CALL-E, but approval and execution remain outside this skill.

Compatibility

The no-call plan builder and result reconciler run on Python 3.9 or newer in any Agent Skills-compatible host. Live execution supports the CALL-E Python SDK or another CALL-E integration that preserves the frozen recipient, task, and idempotency key. Read references/calle-handoff.md before implementing that adapter. The bundled path is deliberately runnable without a provider account or credentials.

When To Use

Use this skill when all of these are true:

• the user has a concrete research goal and constraints;
• cited research has identified one or more businesses;
• a small number of material facts remain unresolved, such as current availability, a price range, a policy, or a scheduling window;
• each recipient is a published organizational number in E.164 format; and
• the user can review the exact recipient, purpose, and questions before any call.

Do not use it for marketing, lead generation, surveys, political outreach, emergencies, deceptive pretexts, personal or wireless numbers, high-impact eligibility decisions, or requests for credentials, payment-card data, health details, government identifiers, or other sensitive account information.

Workflow

1. Separate evidence from gaps. Keep cited facts in established_facts. Add only unresolved, decision-relevant questions to gaps. Read references/input-contract.md.
2. Apply the safety boundary. Reject prohibited purposes and sensitive questions. Confirm every recipient is a published business line. Read references/safety.md.
3. Build the preview. Run scripts/build_call_plan.py. The output is deterministic, masks numbers for display, binds each recipient and question set to an idempotency key, and always says that no call was placed.
4. Ask for explicit approval. Show the complete preview: organization, masked recipient, opening disclosure, purpose, questions, and total expected calls. Do not infer approval from the original research request.
5. Execute outside this skill. Only an approved host integration may translate the frozen plan into CALL-E calls. Make at most one attempt per call-plan item unless the user separately approves a retry. Do not allow recipient, purpose, or questions to change after approval.
6. Reconcile results. Save the provider-neutral result envelope and run scripts/validate_results.py. A completed call still needs a direct callee quote for a fact to become confirmed_by_phone. Voicemail, refusal, ambiguity, timeout, and provider failure remain unresolved.
7. Report honestly. Present sourced, confirmed_by_phone, not_established, and not_reached separately. Never describe an attempted call as a verified answer.