# Hero 3 - AI Honesty-Audit Agent

**Stack:** Python + Anthropic Claude (Sonnet 4.6) + GHL email/SMS read API + roofer self-report intake.

## Flow

1. Weekly: pull last 7 days of closed-deal email + SMS for each roofer (via GHL).
2. Extract signals: claimed close rate, claimed avg ticket, claimed lead-to-close days.
3. Cross-reference against self-reported numbers in roofer dashboard.
4. Use Claude to flag discrepancies (semantic, not just numeric).
5. Log trust score per roofer, alert Storm Sniper ops if score drops.

## Files

- `audit.py` - main script. Uses `ANTHROPIC_API_KEY` from `.env`.
- `prompts/honesty-system.md` - Claude system prompt for the audit
- `sample-data/` - fake roofer data + comms for `--dry-run`

## Run (dry run)

```bash
python3 audit.py --dry-run
```

Reads `sample-data/`, runs Claude on a fake roofer, prints trust score + flagged discrepancies. Uses real `ANTHROPIC_API_KEY` to make one Claude call (cost: <$0.01) unless `--no-api` is passed.

## Run (live)

```bash
python3 audit.py --live --week-ending 2026-05-02
```

## Env vars

- `ANTHROPIC_API_KEY` (existing, in `agency-operator/.env`)
- `GHL_API_KEY` (existing)
- `STORM_SNIPER_OPS_WEBHOOK` (TODO: provision)

## Status

Stub. `--dry-run` works against fake data. Live mode requires GHL pull patterns to be confirmed.