Tiesībsarga biroja iesnieguma izskatīšanas process

UAPF package: lv.tiesibsargs.iesnieguma-izskatisana@0.1.0

An algorithmated complaint-intake and triage process for the Office of the Ombudsman of Latvia (Tiesībsarga birojs). Fetches an incoming citizen complaint (iesniegums), redacts PII before AI processing, extracts structured facts with an AI capability, runs three deterministic DMN decisions to classify the topic / determine priority / route to the appropriate department, persists the classification record, and emits a domain event. Human assignment, response drafting, and case closure remain in the host document management system.

This package is the worked example for the AI-Assisted Case Management profile of UAPF-IP v0.1.

Scope and constraints

This package conforms to the UAPF-IP v0.1 Orchestrated Process profile and the v0.1 reference engine's BPMN walker — which means linear flows only, no gateways, no user tasks inside the process itself. The deliberate consequence: the UAPF process produces a recommendation (classification + priority + routing); the actual human assignment and response work happens in OpenDMS workflows outside the UAPF execution boundary. This is the right shape for a v0.1 demo and the right separation of concerns regardless — UAPF holds the algorithm; the DMS holds the human workflow.

Flow

Start
  └─> [service] document.fetch        — retrieve the iesniegums from the DMS
      └─> [service] ai.redact         — strip PII before any AI invocation
          └─> [service] ai.extract    — produce structured facets per the guardrails-allowed schema
              └─> [rules]  classify-topic         — DMN: facets → canonical topic
                  └─> [rules]  determine-priority — DMN: topic + indicators → priority + SLA
                      └─> [rules]  route-to-department — DMN: topic → department + reviewer role
                          └─> [service] data.write — persist the classification record
                              └─> [service] event.emit — publish "iesniegums.classified"
                                  └─> End

Nine sequence flows, ten nodes, linear. Three service tasks invoke host capabilities. Three business-rule tasks invoke embedded DMN. Two final service tasks record and emit.

Capabilities required

The host must implement at least these to load and execute this package:

Capability	Used at	Purpose
document.fetch@1+	step FetchDocument	Return the iesniegums body + metadata by DID
ai.redact@1+	step RedactPii	Detect and redact personas_kods, addresses, names, financial, health
ai.extract@1+	step ExtractFacets	Return the boolean facet schema below
data.write@1+	step RecordClassification	Persist the classification + routing record
event.emit@1+	step EmitClassifiedEvent	Publish "iesniegums.classified" to the host's event bus

Facet schema (what ai.extract must return)

{
  "mentionsChildren": true,
  "mentionsDiscrimination": false,
  "mentionsPrisons": false,
  "mentionsPolice": false,
  "mentionsHealth": false,
  "mentionsSocialServices": true,
  "mentionsPrivacy": false,
  "mentionsPublicAdministration": false,
  "vulnerablePerson": true,
  "urgency": false,
  "ongoingHarm": false,
  "languageDetected": "lv"
}

The DMN tables consume those booleans plus a topic string flowing between them. Output schema after the full process completes:

{
  "topic": "child-rights",
  "topicConfidence": 0.92,
  "priority": "high",
  "slaHours": 168,
  "department": "Bērna tiesību nodaļa",
  "reviewerRole": "berna-tiesibu-jurists"
}

Guardrails

See resources/guardrails.yaml. Highlights:

• AI is advisory only. decisions_ai_may_make_unattended is empty by design.
• PII redaction is required before any AI invocation. personas_kods, bank accounts, exact addresses, children's names, and health record numbers are forbidden from AI input.
• Allowed models: Claude Opus 4.7, Claude Sonnet 4.6 (Extended Thinking).
• Human oversight required for final classification acceptance, response dispatch, and case closure.
• Audit retention: 10 years for the VC chain, 2 years for AI invocation logs.

Files

lv.tiesibsargs.iesnieguma-izskatisana/
├── manifest.json
├── README.md                              (this file)
├── bpmn/
│   └── iesnieguma-izskatisana.bpmn.xml    Process flow (linear, 10 nodes)
├── dmn/
│   ├── classify-topic.dmn.xml             9 rules, FIRST hit policy
│   ├── determine-priority.dmn.xml         6 rules, FIRST hit policy
│   └── route-to-department.dmn.xml        9 rules, UNIQUE hit policy
├── resources/
│   ├── guardrails.yaml                    Policy snapshot for session
│   ├── tiesibsargs-taxonomy.yaml          Canonical topic list + legal anchors
│   └── department-roster.yaml             Department-to-topic-to-role mapping
├── fixtures/
│   ├── sample-good-governance-complaint.json
│   ├── sample-child-rights-complaint.json
│   └── sample-discrimination-complaint.json
├── tests/
│   └── expected-classifications.json      Expected DMN outputs per fixture
└── docs/
    ├── 00-overview.md
    ├── 01-legal-basis.md
    └── 02-ai-act-classification.md

Building and loading

# Package as a UAPF zip
cd lv.tiesibsargs.iesnieguma-izskatisana
zip -r ../lv.tiesibsargs.iesnieguma-izskatisana-0.1.0.uapf .

# Drop into a uapf-engine packages directory
cp lv.tiesibsargs.iesnieguma-izskatisana-0.1.0.uapf /path/to/uapf-engine/packages/

# Or push to ProcessGit for governed distribution
git push https://USER:TOKEN@processgit.org/Tiesibsargs/iesnieguma-izskatisana.git main

Running against the engine

Once loaded into a uapf-engine instance and a UAPF-IP-conformant host is reachable, invoke with:

curl -X POST http://uapf-engine:4000/uapf/start-session \
  -H "Content-Type: application/json" \
  -d '{
    "packageId": "lv.tiesibsargs.iesnieguma-izskatisana",
    "packageVersion": "0.1.0",
    "processId": "iesnieguma-izskatisana",
    "input": {
      "documentDid": "did:doc:tiesibsargs:2026-0518"
    },
    "hostManifest": {
      "hostDid": "did:web:tiesibsargs.opendms.dev",
      "hostBaseUrl": "https://tiesibsargs.opendms.dev",
      "profiles": ["uapf-ip-orchestrated"],
      "capabilities": [
        { "namespace": "document", "operation": "fetch",  "version": 1 },
        { "namespace": "ai",       "operation": "redact", "version": 1 },
        { "namespace": "ai",       "operation": "extract","version": 1 },
        { "namespace": "data",     "operation": "write",  "version": 1 },
        { "namespace": "event",    "operation": "emit",   "version": 1 }
      ]
    }
  }'

Status

• v0.1 draft. Ready for first integration against OpenDMS under the MIC AI regulatory sandbox.
• DMN tables hand-tuned from the Tiesībsarga 2025 annual report category distribution; should be refined with real complaint data once the sandbox protocol allows.
• The ai.extract facet schema is the spot most likely to evolve. Treat it as a contract between this package and whatever AI model serves the capability.

License

MIT.