Payment dispute handling with human decisions

Classifies a payment dispute, assesses the evidence against policy and prepares a recommendation that a case worker decides.

router risk high v1 financial services seed

Why this topology

high-risk-no-autonomy: autonomous topology removed (risk is high)
prefer-workflow: agentic topologies removed (the task is not open-ended)
autonomous-needs-consent: autonomous topology removed (autonomy is act_with_approval)
multiple-task-kinds: router (2 request kinds)

Mandatory controls: human_gate input_filter audit_log output_filter retrieval

Flow

intake → Dispute intake · 1200 in / 200 out
assess → Evidence analyst · 6000 in / 900 out

Diagram source (Mermaid)

flowchart LR
    io_in(["request"])
    io_out(["response"])
    c_intake_model(["Dispute intake"])
    c_analyst_model(["Evidence analyst"])
    c_policy_kb[("Policy and precedent index")]
    c_case_tool[["Case system"]]
    c_ledger_tool[["Ledger"]]
    c_decision_gate{{"Case decision"}}
    g_redact>"input_filter"]
    g_disclosure_check>"output_filter"]
    g_audit>"audit_log"]
    io_in --> c_intake_model
    c_intake_model --> c_analyst_model
    c_analyst_model --> io_out
    g_redact -.-> c_intake_model
    g_redact -.-> c_analyst_model
    g_disclosure_check -.-> c_analyst_model
    g_audit -.-> c_case_tool
    g_audit -.-> c_ledger_tool

Components

id	name	kind	role	model / tool / framework
intake-model	Dispute intake	model	classify the dispute type and extract the claimed facts	claude-haiku-4-5
analyst-model	Evidence analyst	model	assess the evidence against policy and draft a recommendation	claude-opus-5-5
policy-kb	Policy and precedent index	retrieval	dispute rules and past decisions
case-tool	Case system	tool	read and update the dispute case
ledger-tool	Ledger	tool	post a refund or a chargeback to the ledger
decision-gate	Case decision	human_gate	a case worker decides every outcome

Guardrails

id	kind	protects	what it does
redact	input_filter	intake-model, analyst-model	Account and card numbers are tokenised before any model sees the case.
disclosure-check	output_filter	analyst-model	Blocks recommendations that quote personal data or internal fraud rules.
audit	audit_log	case-tool, ledger-tool	Every read, update and posting is logged with the case id and the decider.

Failure modes

failure	component	mitigated by
A dispute is classified under the wrong rule set.	Dispute intake	decision-gate
The assessment ignores a document that changes the outcome.	Evidence analyst	decision-gate
A recommendation exposes personal data or internal rules.	Evidence analyst	disclosure-check
A posting cannot be traced to a person and a case.	Ledger	audit

Threats

OWASP	component	mitigated by	note
LLM01 Prompt Injection	Dispute intake	redact	customer-submitted text is untrusted
LLM02 Sensitive Information Disclosure	Evidence analyst	disclosure-check	sensitive data in outputs
LLM06 Excessive Agency	Ledger	decision-gate	money movement
LLM09 Misinformation	Evidence analyst	decision-gate	a person checks every recommendation

Eval plan

metric	method	threshold
classification accuracy	500 labelled historical disputes	>= 0.95
recommendation agreement	case workers compare recommendations with their own decision on a blind sample	>= 0.85
gate coverage	replay; no posting without a decision record	100%

Observability and deployment

traces per case
gate decisions and overrides
disclosure-check blocks
time to decision

service inside the regulated network · No outbound calls except to the model gateway.

Estimated cost

$0.0442 per request · $1060.80 / month at 800 requests/day · ~18.03 s end to end

Prices as of 2026-10-01 · source · latency is assumed.

token counts per step are estimates written with the blueprint and approved by the owner, not measurements
steps run one after another, so parallel branches make the latency an overestimate
latency uses an assumed time to first token and output speed per model
30 days per month and the stated requests per day

Volume Tokens

Runnable starter

No starter is offered: no API data for langgraph from Radar yet.

Changelog

v1: seed