Verify By Phone

Verify By Phone

Use this skill when an agent must establish whether stored information about an organization is still true, and a phone call to the organization's published line is the way to find out.

verify-by-phone is a verification workflow skill. It places exactly one disclosed outbound CALL-E call per record, extracts the answer from the transcript with a verbatim supporting span, reconciles it against the stored record with transparent match-weight arithmetic, and returns either a calibrated confidence or an explicit abstention. The design goal is an agent that never converts an uncertain phone answer into a confident database write.

When To Use

Use this skill for:

• verifying a provider directory listing: is this practice real, reachable, accepting new patients, taking a given insurance plan
• checking whether a stored business record (hours, services, availability) still matches reality
• refreshing any dataset where the phone is the source of truth and the record is suspected stale
• workflows where an "unknown" outcome must be recorded honestly instead of guessed
• one record at a time, with a human-authorized recipient list

When Not To Use

Do not use this skill to:

• place undisclosed or pretext calls; every call announces that it is an automated assistant and why it is calling, in the same breath
• call wireless or personal numbers; verification targets published organizational lines
• run marketing, sales, lead generation, or any promotional outreach
• batch-dial a list without per-record human authorization
• overwrite a stored record directly from a call result; the output is a verdict with evidence, and the write decision stays with the operator
• guess an answer the respondent did not give; if the call yields no usable answer, the result is an abstention, and that is the correct output

Consent And Disclosure Rules

These are load-bearing, not boilerplate:

1. The call opens by stating the AI identity and the purpose together, before anything else, and announces that the call may be recorded.
2. If the respondent objects to speaking with an automated caller, the call thanks them and ends immediately. The result records the refusal as an unverifiable outcome.
3. Hold time is capped. If the respondent asks the caller to wait, it waits briefly, then reports back rather than waiting indefinitely.
4. The agent never invents information it was not given, on the call or in the result. Concretely: a clinic will often ask the caller for a name, a date of birth, an insurance member or card number, or a reason for the visit. The agent says plainly that it does not have that information, because this is a directory verification call and not an appointment request, and then repeats the question it called to ask. A placeholder is an invention.
5. Voicemail is not an answer. On reaching voicemail or an answering machine the call ends immediately without leaving a message: a directory fact cannot be established from a recording, and nobody should find a robot message on their line.
6. Calls are informational verification, never promotional. Treat every call as recorded with all-party consent requirements in mind.

Verification Workflow

1. Calibrate the abstention threshold first, with scripts/calibrate.py on labeled scenario data, so "confident" means something measurable: at the default level, the true answer falls inside the prediction set at least 90 percent of the time on held-out data. It prints a qhat that step 6 consumes. This runs with no credentials and no calls, so it can be done once, ahead of any dialing.
2. Collect the record to verify: organization name, published phone number in E.164, and the claims to check (for example accepting new patients, accepts a named plan).
3. Confirm the operator authorizes this specific call to this specific number.
4. Build the call task with scripts/place_verify_call.py. The script defaults to a dry run that prints the exact task and recipient without dialing; pass --live only after the dry run looks right.
5. Poll with scripts/poll_result.py until result_status is no longer pending, even if execution already says completed. It saves the full Calls V2 payload to a private local file. An unavailable result is a normal outcome; extraction still reads any recorded transcript.
6. Extract the answer with scripts/extract_answer.py --qhat from step 1. Every extracted field carries the verbatim transcript span and character offsets that support it. Hedged answers ("I think so") keep their polarity at a dampened trust score. Non-responsive turns (wrong number, refusal, "call back later") never count as answers. The answer is served only when the calibrated prediction set is a single value and that value is not "unknown"; otherwise the result is an abstention. Omitting --qhat abstains on everything and labels the output uncalibrated, because a threshold with no calibration behind it guarantees nothing.
7. Reconcile against the stored record with scripts/reconcile_record.py: each agreeing field adds documented bits of evidence, each disagreeing field subtracts them, and the verdict is verified, contradicted, or unverifiable.

Maintenance Note

The scripts in this skill are self-contained copies of the reference implementation in the Attest backend (backend/app in the source repository). They are kept small on purpose so the skill installs with no dependencies on that repository.

That copy is enforced, not promised. tests/test_skill_parity.py in the source repository runs this skill's extractor and the backend's over the same transcripts and fails if they disagree on the answer, on the character offsets of the cited span, or on the cue lexicons themselves. An earlier version of this note asked a human to re-sync by hand, and the copy drifted anyway: the backend learned to trust the last cue in a turn and to read "no problem" as agreement while this script still trusted the first, so a plain "No, we are not" abstained here and answered there.