Article 50 voice-agent proof

A secrets-free reference implementation for an inbound AI voice qualification agent. It demonstrates the engineering controls behind Forhemit Labs' fixed-price Article 50 Launch Patch; it does not decide whether Article 50 applies or certify compliance.

What the proof enforces

• Identifies itself as an AI voice agent at the start of every call.
• Gives implementation information only; buyer counsel or a named compliance owner retains applicability and wording decisions.
• Requests explicit permission before collecting business follow-up details.
• Asks one question at a time and caps the delivery shape at three product surfaces.
• Never requests credentials, source code, payment-card data, health information, or private checkout details.
• Stops persuasion when a caller opts out or says they are not interested.
• Logs only that a consented capture occurred and the number of fields; the reference code does not log raw lead details.

Provider adapters and local simulation

• src/agent.py is a LiveKit Agents reference implementation for an inbound RTC session.
• vapi-assistant.example.json is a provider-neutralized Vapi assistant definition with the same conversation boundaries.
• src/call_simulator.py is a deterministic, standard-library simulator for inbound and permission-gated outbound call lifecycles.

These files contain no phone number, API key, provider account identifier, destination number, or checkout URL. No live telephony resource is created by this repository.

Listen to the fictional call flow

Play or download the MP3 and read the exact transcript.

This is a local text-to-speech rendering of a fictional qualification exchange, not a recording of a phone call. It demonstrates AI identity disclosure, the implementation-only boundary, consent before business-detail capture, bounded surface questions, sensitive-data suppression, and human review. It used no phone number, carrier, caller, lead, or live telephony.

Simulate the lifecycle

Run an inbound call that declines recording consent:

PYTHONPATH=src python -m call_simulator inbound --recording-consent no

Prove that an outbound call fails closed without prior contact permission:

PYTHONPATH=src python -m call_simulator outbound

Simulate a permitted outbound call with a data-minimized three-field capture:

PYTHONPATH=src python -m call_simulator outbound \
  --prior-contact-permission yes \
  --recording-consent yes \
  --supplied-field-count 3

The JSON output contains only lifecycle state, booleans, and a field count. It accepts no raw contact values, persists nothing, and explicitly reports live_telephony_created: false.

Verify locally

python -m pytest -q
python -m ruff check .

The adapter tests statically verify disclosure, consent, legal, sensitive-data, checkout, and logging boundaries. The simulator tests execute inbound disclosure, outbound permission, opt-out, and data-minimization behavior without provider credentials.

Deployment boundary

Before connecting a number, supply your own provider account, confirm recording and telemarketing requirements for every relevant jurisdiction, define retention and deletion rules, and replace the reference capture with a reviewed business-data sink. Do not use this sample for emergency, healthcare, financial, or other high-stakes decisions.

The commercial engineering offer and human-reviewed fit check are available at article50-launch-patch-public.vercel.app. The public checkout is intentionally not exposed.