Fraud indicators
Agent: the fraud indicator agent (fraud_detection in ExecuteLoop), with the
FRAUD_INDICATOR_REVIEW evidence document and the ReviewFraudIndicators call.
What it does
It flags records in your own logistics and claims data that a claims or special-investigations reviewer should check, and drafts the check. It never accuses anyone. It looks across invoices, payments, claims, deliveries, carrier records, payee details and returns for irregularities that warrant a check.
What it reads
Payments, claims, carrier invoices with their rate confirmations, receiving (OS&D) records, tracking telemetry and returns records. Everything is keyed by a business record: an invoice, a payment, a claim, a load, a carrier, a payee or a returns account.
What it produces
- A
FRAUD_INDICATOR_REVIEWevidence document per indicator on Actionable Twins (claims play): the evidence behind it, each item citing its records, the checks that agree and their weight, a likelihood band (low, medium or high) when you have set band cut-offs, the amount on the flagged records where one is stated, and one next check. - From
ReviewFraudIndicators: the indicators at or above a band you choose, and, if you ask, a drafted check per indicator. Each check is a concrete verification step and names its reviewer. An indicator with no band is always listed.
How the likelihood band is set
The band is the checks’ weight against the cut-offs you set (band_medium_at and
band_high_at). There are no built-in cut-offs: until you set both, an indicator
shows its checks and weight and says no review cut-offs are set, with no band and no
priority. No model sets or moves a band. Every card states the review settings in
force and where they came from.
The statistical checks flag a value whose modified z-score (distance from the median in 1.4826 × MAD units) is at least 3.5, the outlier cut Iglewicz and Hoaglin recommend for that statistic.
Review settings
The review settings can be adjusted for your deployment. Set
FACE_INDICATOR_THRESHOLDS_FILE to a JSON file that names only the values you
change. A file that cannot be read or is out of range is not applied, and every card
says the override was invalid. Band cut-offs are set as a pair, or not at all.
Human oversight
- A named claims or special-investigations reviewer decides each indicator, with one of two choices. Dismiss needs a reason, which is recorded. Escalate opens an investigation case for that reviewer. Both are written to the decision log. An indicator is never “approved”.
- The agent never accuses anyone, never holds, blocks or denies a payment, claim or refund, and never contacts the party named.
- An indicator is not for sending to the party it names.
Model
The deployment’s general reasoning tier.
The indicators, bands and amounts are fixed rules and statistics over cited records. The agent orders the indicators and phrases the next check; the server takes bands and amounts from the card, and every drafted step is checked before you see it.
Where it runs and data handling
On the deployment’s own model plane: Runink’s shared machines on a Trial, your own cloud on Cloud and your own servers on On-premises. No outside AI service is called on any plan. Your records reach the model as fenced, untrusted data. See Facts shared by every card.
Guardrails
- No profiling: it never uses or infers gender, race, ethnicity, religion, nationality, age, disability, a name, an address or a neighbourhood.
- Never looks into anyone’s private life, social media, family, criminal record or credit.
- Refuses to label anyone a fraudster or as guilty, and refuses to hold or deny a payment without a person.
- Refuses to hide or drop an indicator for a particular party.
- Each drafted check is screened against these rules as output, before it reaches you.
How to read its results
- This is not a consumer report. It must not be used for decisions about an individual’s credit, employment or insurance.
- It works from your records. A pattern can have an innocent explanation the records do not show, which is why every indicator ends in a check, not a conclusion.