Skip to content
Model cards

Agent model cards

A model card describes one of FACE’s agents in plain language: what it is for, what it reads, what it hands a person, who has to approve it, which model it runs on, and where it stops. There is one card for each kind of work you meet in FACE. What an agent finds shows on Ask, its options on Options, and what it drafts for approval on Actionable Twins. Several cards cover more than one agent, because those agents do one job together.

These cards are written for buyers, evaluators and operators. The mechanics behind them are on the Analysis & evidence pages, which these cards link to rather than repeat.

One rule for every card

FACE drafts. A person decides. Every claim, short-pay, filing, order, quarantine, remediation and repair that an agent proposes is a recommendation for a named approver. Nothing is sent, filed, paid or settled until someone with the authority approves it, and the server records who did. The one exception is FACE’s own operational self-healing: restarting, isolating, rolling back or scaling FACE itself runs automatically from a fixed list, and an operator can switch it off. The approval tiers and the hash-chained decision log are described in Auditability & lineage and Agents & oversight.

The cards

CardWhat it helps withBase model
Claims & recoveryShortage, OS&D and SLA claims, filing windows and the letter for approval.General
Freight invoice auditCarrier invoices against the rate confirmation, SLA credits and spend.General
Customs & trade documentsCustoms holds, document cross-checks and draft trade documents.General
Compliance reviewPolicy against practice, framework checks, personal data and remediation.General, plus code for code checks
Insurance claim filesAssembling the cargo claim file for an adjuster.General (documents use no model)
Cold chain & telemetryTemperature excursions, sensor faults, site thresholds and custody.General
MaintenanceEquipment trending towards a limit its feed states.General
Stock & demandReorder decisions, forecasts and purchase orders for approval.General (forecasts use no model)
ReturnsReturn windows, quantities and disposition by your policy.General (dock triage uses no model)
Yard & transportYard capacity, border dwell, demurrage, custody and route ranking.General
Options & approvalsThe options for a decision, scored, and the approval tier for each action.General (option scoring and approval policy use no model)
Field readings (edge)Writing up a handheld reader’s site walk.General
Data postureGrouping your tables into domains and grading their readiness.General
Voice dispatchShort spoken answers for drivers, with the call logged.General, plus the voice tier
Documents & plans (vision)Camera frames, photographed forms and floor-plan images.Vision model, starts on demand
Self-healingFACE’s own health: restart, isolate, roll back or scale, automatically.General
Claims leakageRecoverable losses, credits and short-pays that never became a claim, handed to claims recovery.General
Fraud indicatorsPatterns worth a reviewer’s look, with evidence and a likelihood band; never an accusation.General
Freight market analystEach lane against its own history and public fuel prices; re-bid and consolidation candidates.General
Counterparty background checkPublic registry and sanctions findings on a carrier, shipper or broker, for review. Businesses only.General
Customer analysisCost to serve, service performance, claims, payment behaviour and concentration per business account.General

How to read a card

Every card has the same sections:

  1. What it does: one sentence, then the jobs it handles.
  2. What it reads: the records, documents and systems it takes in.
  3. What it produces: what you get back, and in what form.
  4. Human oversight: what a person approves, and what the agent can never do by itself.
  5. Model: the base model it runs on.
  6. Where it runs and data handling: where inference happens.
  7. Guardrails: the policies it follows, in plain words.
  8. How to read its results: what a result shows, what stays with a person, and when it declines or says data is missing.

To see what each agent produces on real data, load the quickstart estate and ask about it on Ask.

Facts shared by every card

Models. FACE sends each model call to one of three tiers on the deployment’s model plane, chosen by the kind of task (see Sovereign models):

  • General reasoning: most agents.
  • Code: rule checks written as code.
  • Vision: documents and images; it starts on demand.

Each tier serves the deployment’s own models.

Not everything is a model. Many of the figures on these cards are fixed arithmetic over cited records: evidence documents, the dock triage of returns, forecasts, the approval policy. A model may phrase a result only after it has been determined. See Evidence documents.

Where inference runs. Every agent calls the model plane of its own FACE deployment: Runink’s shared machines on a Trial, the customer’s own cloud on Cloud and their own servers on On-premises. FACE calls no outside AI service on any plan. Records an agent reads are passed to the model as fenced, untrusted data, never as instructions (see AI safety). The connections you configure still reach their own systems, and an approved action, such as an email sent through a connected account, goes where that action says.

Guardrails. Each agent’s requests are checked against its OpenBias guardrail rules before a prompt reaches the model, and each proposed action passes the judging ladder. Both are described in Agents & oversight and AI safety, including the scope of each.