CallProof

CallProof is a closed-loop phone agent system built for the CALL-E: Your Code Is Calling hackathon.

CALL-E executes real calls, Call Analyzer verifies outcomes against explicit policies, and AgentKit Rails orchestrates memory, human review, telemetry, and continuous improvement.

Development status and milestones are tracked in IMPLEMENTATION_PLAN.md.

apps/web/                 Rails application and AgentKit host
services/call_analyzer/   Python analysis service
contracts/                Versioned integration schemas
docs/                     Architecture and submission documentation
hackathon/                Upstream contribution preparation

The project defaults to fake adapters and must not place a real phone call without an explicit live-mode opt-in.

The Rails application now supports two deterministic scenarios:

• a compliant simulated call that AgentKit marks as verified;
• a simulated policy violation that Call Analyzer supports with transcript evidence and AgentKit suspends at a persisted HITL gate.

It also includes the first real Rails ↔ Call Analyzer boundary:

• FastAPI transcript ingestion with persisted received → analyzing → completed states;
• idempotent analysis requests and a recovery endpoint;
• HMAC-signed completion webhooks with timestamp and replay protection;
• a second resumable AgentKit flow that routes remote policy exceptions to HITL.

The analyzer contract is now version 2.0 and fail-closed. Each success condition is a typed claim with an expected provider value and a closed protocol identifier from which Rails and the analyzer derive the same exact confirmation statement. The offline evaluator only verifies that finite protocol; free-form language and semantic-only rules are explicitly unevaluated and go to human review. Provider results are claims to corroborate, never transcript evidence by themselves.

The official CALL-E HTTP adapter is implemented behind explicit live-call safety gates. The demo UI cannot enable it. Terminal result polling is normalized into the same transcript contract and then sent through Call Analyzer and AgentKit.

The optional live workflow is deliberately two-step: it first creates a no-side-effect contract preview, masks the recipient number, and only exposes confirmation after checking provider, global live switch, and API-key readiness. A typed phrase and browser confirmation remain required after those checks.

cd apps/web
rvm use 3.4.1
bundle install
bin/rails db:prepare
bin/rails server

Open http://localhost:3000. The default configuration uses fictional data, a fake phone provider, and a fake LLM adapter.

Run the analyzer independently with:

cd services/call_analyzer
python3 -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'
CALL_ANALYZER_ENV=development CALLPROOF_WEBHOOK_SECRET=replace-me uvicorn call_analyzer.main:app --reload --port 8001

docker compose up --build starts the complete safe stack: Rails at http://localhost:3000, PostgreSQL with pgvector, the analyzer API at http://localhost:8001, Redis, and the RQ worker. The phone provider and LLM adapter remain fake, and external webhook delivery remains disabled. A production deployment must provide a public HTTPS callback URL and explicit live-call switches.

See docs/calle-integration.md before enabling any CALL-E environment variables.