Muster — Directory Ghost-Rate Auditor (CALL-E Agent Skill)

Audits directories before patients and consumers suffer from ghost networks.

Muster automates offline batch phone verification of contact directories that institutions claim are active:

• Insurer in-network provider rolls (primary care, therapists, clinics, specialists)
• Marketplace verified merchants (vendors, retail store fronts, inventory fulfillment)

By placing non-invasive, consent-first qualifying calls via CALL-E, Muster asks ONE benign verification question, extracts structured operational signals, cites verbatim transcript quotes, and computes the verified Ghost Rate (e.g. "7 of 12 audited = GHOST, 58.3%").

---

Prerequisites & Engine Setup

This skill orchestrates the open-source Muster directory auditor engine. To run audits locally:

# Clone and install the Muster engine
git clone https://github.com/omshukla24/Muster.git
cd Muster
pip install -e .

Note: For lightweight workflow exploration, the skill operates in mock mode by default, allowing zero-cost simulated audits with bundled sample schemas without needing live telephony credits.

---

When to Use This Skill

Activate Muster when an agent or auditor needs to:

1. Audit a directory file (.csv or .json) to flag defunct, disconnected, or left-network listings.
2. Calculate headline Ghost Rate and real-world consumer access metrics (e.g. "A patient must call ~2.4 listings to reach 1 real provider").
3. Generate cited evidence reports containing verbatim spoken receptionist quotes for compliance or enforcement.
4. Invoke via FastMCP tool audit_directory from any agent workflow.

---

Inputs & Parameters

Parameter	Type	Default	Description
directory	File Path	Required	Path to CSV or JSON directory file (must contain entity name and phone number).
sector	String	US_INSURER	Sector archetype: US_INSURER (healthcare in-network) or MARKETPLACE_SELLER (merchants).
mode	String	mock	mock (zero-cost simulated rehearsal) or live (places real phone calls via CALL-E).
goal	String	Auto-selected	Custom verification question. If omitted, uses sector default.
concurrency	Integer	2	Number of simultaneous verification calls (max recommended: 3).

Input Schema Examples

CSV Format (sample_data/us_insurer_network.csv)

id,name,phone,claimed_status,category,address
PROV-101,"Dr. Sarah Jenkins, MD",+15552340101,In-Network - BlueCross PPO,Primary Care,"Austin, TX"
PROV-102,Summit Psychological Associates,+15552340102,In-Network - Aetna Choice,Mental Health,"Denver, CO"

JSON Format (sample_data/us_insurer_network.json)

[
  {
    "id": "PROV-101",
    "name": "Dr. Sarah Jenkins, MD (Family Medicine)",
    "phone": "+15552340101",
    "claimed_status": "In-Network - BlueCross PPO",
    "category": "Primary Care",
    "address": "Austin, TX"
  }
]

---

Outputs & Artifacts

Muster produces experimental, advisory classifications intended for human review and operational triage (not definitive legal or regulatory compliance verdicts):

Verdict Definitions

• PRESENT: Authorized respondent affirmed in-network participation and intake availability.
• GHOST: Dead/unallocated number, wrong person/business reached, or respondent explicitly stated provider left network / closed.
• UNREACHABLE: Call timed out after multiple rings, busy signal, automated IVR voicemail loop without human answer, or carrier delivery failure.
• UNCERTAIN: Inconclusive dialogue or premature disconnection.

Generated Artifacts

1. reports/muster_audit_<job_id>.csv: Tabular dataset with entry IDs, phone numbers, verdicts, confidence scores, and cited evidence quotes.
2. reports/muster_audit_<job_id>.json: Machine-readable audit package including verbatim transcripts and full schema extractions.
3. reports/muster_audit_<job_id>.md: Executive Audit Report (Advisory) summarizing findings, confidence scores, and patient access ratios for human audit review.

---

Usage Examples

1. Zero-Cost Rehearsal (Mock Mode)

Run a deterministic audit simulating realistic reception dialogues without spending CALL-E credits:

python skills/muster/scripts/run_audit.py sample_data/us_insurer_network.json --mode mock

2. Live Phone Verification (CALL-E Runtime)

Run real phone verification calls through authenticated CALL-E CLI:

python skills/muster/scripts/run_audit.py sample_data/us_insurer_network.json \
  --mode live \
  --concurrency 2 \
  --goal "Hello, I am calling to verify directory network status. Are you currently in-network and accepting new patients?"

(Note: Live runs require interactive confirmation or --confirm-live to protect against unintentional credit spend).

3. Invocation via FastMCP Tool

Any agent or MCP client can invoke Muster directly:

from muster.mcp_server import audit_directory

report = await audit_directory(
    file_path="sample_data/us_insurer_network.json",
    sector="US_INSURER",
    mode="mock"
)
print(report)

4. Interactive Live Audit Ledger UI

Launch the full-bleed responsive web dashboard:

python -m muster.cli ui --port 8000

Open http://127.0.0.1:8000 to inspect the live switchboard, telegraph typing, and "The Reveal" finale.

---

Safety & Telephony Contract

All live operations are strictly governed by our Safety Contract, enforcing authorized E.164 recipients, per-run human preview/approval, privacy phone masking, no auto-retries on ambiguity, and honest cancellation limits.

• Benign Qualifying Questions Only: Muster strictly verifies administrative network status. Never request patient health information, medical history, or payment details.
• Rate-Limited Gateways: Concurrency defaults to 2 simultaneous calls to respect cellular gateways and medical front desks.
• Transparent Identification: Verification inquiries explicitly introduce themselves as directory verification checks.