Cintia Henriksson
Case 02 • Functional Prototype & Evals

“Where is my money?”: Support Agent with Asymmetric Release Gates

Design and implementation of an in-transit payments support agent governed by a formal behavior specification without out-of-band prompts, validated across a 120-ticket test harness with zero tolerance for false promises.

MODELED DEFLECTION

18.8%

Realism discount

RELEASE GATE

0 Violations

On C2.1–C2.6 & C2.8

COST / 1K TICKETS

$1.80

Economy-C Model

P50 LATENCY

480 ms

Optimized inference

01. Problem Statement & CFPB Regulation

In international remittances, peak friction happens when a transfer is flagged 'in transit'. Standard chatbots placate users with empty empathy or invent delivery timestamps, triggering regulatory complaints and chargeback disputes.

Injection of Statutory Facts (CFPB Remittance Transfer Rule):

  • Mandatory cancellation window: 30 minutes (provided funds have not been picked up).
  • Statutory refund timeline: maximum 3 business days.
  • Dispute windows: 180 days to report an error and 90 days for formal resolution.

Regulatory facts live as structured data injected into the spec, never in stochastic LLM memory.

02. Operating Prototype (Live Simulator)

behavior_spec.v1

Select a test scenario to inspect conversational behavior in real time alongside JSON contract extraction and guardrail enforcement.

UserHi, I sent 200 USD this morning and want to know if it has arrived yet.
Guarded AgentYour transfer TX-1092 is completed and funds have been credited to the recipient account since 14:30.
Status: HTTP 200 OK • p50 Latency: 480ms
PM INSPECTOR (LIVE PAYLOAD)Spec v1.1.0
{
  "intent": "STATUS_INQUIRY",
  "confidence": 0.98,
  "hard_constraints_checked": [
    "C2.1_PASS",
    "C2.2_PASS",
    "C2.8_PASS"
  ],
  "handoff_triggered": false
}

03. Evaluation Test Harness & Release Gate

The release gate is asymmetric: zero violations across critical constraints (C2.1–C2.6 and C2.8) across the 120-ticket golden test set. In cross-border remittances, a single hallucinated arrival date or privacy breach costs far more than ten uncontained tickets.

ModelCost / 1kp50 LatencyPass RateGate ViolationsRelease Decision
Frontier-A14,20 USD1.850 ms94,1%1 violation (C2.1)BLOCKED
Frontier-B11,50 USD1.200 ms91,6%0 violationsDISCARDED (COST)
Economy-C (Guarded)1,80 USD480 ms89,2%0 violationsRELEASE APPROVED

04. Post-Mortem: Failure Gallery & Gate Hardening

Failure T014: Hallucinated arrival timestamp (C2.1)

“Don't worry, it will arrive Monday by 6:00 PM at the latest” The naive model responded with this timestamp when the core banking engine had no confirmation.

Solution: Strict prohibition against outputting time units whenever system_eta == null.

Post-Run Addition: Third-Party Privacy (C2.8)

In the initial run, a recipient with a valid transaction ID prompted the model to disclose the sender's source amount in 3 edge cases.

Solution: The release gate was hardened post-run by adding C2.8 to mandate sender session validation.

05. Economic Impact: Discounted Deflection Modeling

Net Containment Calculation Chain:18.8% Net Deflection
Deflection = (% status tickets: 34%) × (Pass rate: 82%) × (Residual containment: 45%) × (Realism discount: 0.70) ≈ 18.8%

In contrast to the 30–60% deflection routinely marketed by AI vendors, this model enforces an explicit realism discount for multi-turn support noise and reweights against Case 1 corpus mix, projecting ~115,000 USD in sustainable annual direct support savings.

Next Step • First-Party Data Decision

From technical agent specifications and evaluations to digital product roadmap prioritization backed by authentic feedback and RICE.

View Case 3: Student Roadmap