Phone Claim Verifier

Phone Claim Verifier

Use this skill when the user has a table of records - each with a phone number and one or more claims attached - and wants those claims checked by calling the number and asking.

Examples: "call each clinic in this CSV and confirm they still take the Northline Health plan", "phone these suppliers and check the contact name on file is still right", "verify this community-services directory is still accurate".

The skill is a thin wrapper around CALL-E's one-off call workflow. It does not add backend APIs, provider-side batches, a daemon, or new MCP tools. Two small standard-library scripts do the deterministic work; you (the agent) place the calls and read the transcripts.

The one rule

A MATCH or MISMATCH verdict is only produced when a verbatim quote from a recipient turn supports it. If you cannot copy such a quote out of the transcript, the verdict is UNCLEAR - stated plainly, not as an error. This is enforced in scripts/verdict.py, in code. Read references/evidence-rules.md before resolving any call.

When To Use

• checking phone-reachable records against claims: does this office accept this plan, is this still the contact, are these hours current, is this address right
• batch verification where each record has its own phone number and its own asserted facts
• producing an evidence-backed corrections file a team can act on

When Not To Use

• placing calls to numbers that are not on a record in the user's own table
• inferring a phone number, region, timezone, or language - ask the user, do not guess
• calling third-party numbers the user has not confirmed they are authorised to contact
• adding a CALL-E backend verification API or a provider-side recurring job
• resolving a verdict from the recipient's tone or your own inference rather than a quote
• installing calle globally without the user's explicit approval

Setup

1. Confirm the user wants real outbound calls placed to the numbers in their table.
2. Get the records. A CSV with record_id,name,phone,address,region plus one column per claim (values true/false), or a JSON object per record. See examples/providers.csv.
3. Get or write a claim pack (a JSON file listing the questions to ask and how to read each answer). See references/claim-packs.md and examples/healthcare.json.
4. Check CALL-E auth. Resolve a CALL-E command or the MCP route the same way other skills in this repository do; the MCP route is https://seleven-mcp-sg.airudder.com/mcp/openagent_oauth.

CALL-E CLI parameters and command flags are documented in cli-reference.md.

Core Workflow

1. Build the plans:

   python3 scripts/plan.py records.csv --pack path/to/pack.json

   This prints a phone-masked preview of dial, goal, result_schema, and idempotency_key. The preview is not executable. After approval, run the same command with --private-payload for private machine input to the executor: it preserves the exact number from the record and the full goal. Never show that private output to the user or paste it into logs. Neither mode places a call.

2. For each plan, run the CALL-E one-off call workflow:

   auth status  ->  call plan  ->  (show the plan to the user)  ->  call run  ->  call status

   • Use the private payload for execution, not the masked preview. Plan exactly one call, to plan.dial. Do not substitute any other number.
   • Inspect the returned plan. Run it only if it targets plan.dial and carries the goal text.
   • Preserve any returned plan_id and confirm_token exactly.
   • Poll call status until the call reaches a terminal state and a transcript is available.
   • If submission or status times out or acceptance is unknown, stop this batch. Do not redial, advance to the next record, or report no contact. Reconcile the existing intent with the provider and obtain a new explicit operator decision before resuming.

3. Read the transcript. For each claim in the pack, decide:

   • answer: yes, no, or unknown
   • evidence_span: a substring copied verbatim from a single recipient turn that proves the answer, or null if no such substring exists
   • conflicting: true only if the recipient stated both a yes and a no Follow references/evidence-rules.md. Do not paraphrase the quote.

4. Resolve each claim:

   python3 scripts/verdict.py \
     --record '<record JSON>' --pack path/to/pack.json --claim <claim_id> \
     --transcript <call output or [{speaker,text}] JSON> \
     --extraction '{"answer":"...","evidence_span":"...","source_role":"...","conflicting":false}'

   It prints a masked attestation (verdict, quote, provenance, expiry) and appends a private row to corrections.csv only for an evidence-backed MISMATCH. An UNCLEAR never becomes a correction. Exact quote matching and the private corrections file remain unchanged; --private-payload is available only for a private machine consumer. Verdict meanings: references/verdict-taxonomy.md.

5. Report per record: the verdict for each claim, a phone-masked display of the quote behind any MATCH/MISMATCH, and the path to corrections.csv. Do not present an UNCLEAR as a failure - it is the skill declining to guess.