Building Agent-Callable Evidence from Government Open Data

Description and use

CivicDataForge developed a workflow to transform government open data into agent-callable evidence. The solution addresses pagination, explicit decision semantics (e.g., 'no published match'), and strict provenance tracking to ensure reliability for AI agents and production systems.

Problem

Government open data is often incomplete, paginated, or structurally inconsistent, making it unreliable for production systems or AI agents that require definitive evidence rather than just raw data.

Outcome

The team established a rigorous framework for treating government data as evidence, ensuring completeness, provenance, and clear semantic boundaries for AI agents, moving beyond simple scraping to reliable evidence delivery.

Workflow

• Discover publisher page limits and request every page in a stable order.
• Track page count, observed row count, and source-reported totals.
• Reject monitoring baselines if the query is capped, partial, or structurally inconsistent.
• Bind the completed scope into a receipt with source metadata.
• Use an explicit decision vocabulary (EVIDENCE_FOUND, NO_PUBLISHED_MATCH, etc.) instead of binary yes/no.
• Normalize data while retaining lineage (publisher, endpoint, query, timestamp, original field names).
• Monitor independent dimensions: availability, schema compatibility, record-count behavior, and freshness.
• Implement idempotent webhook handlers using actor_run_id + event_type as keys.
• Define task-to-tool selection contracts for AI agents specifying boundaries and minimum inputs.

Evidence

A response can be valid JSON, have a 200 status, and still represent only the first page.

Source

We use an explicit, fail-closed decision vocabulary: EVIDENCE_FOUND, NO_PUBLISHED_MATCH, REVIEW_REQUIRED, SOURCE_UNAVAILABLE, SCOPE_INCOMPLETE.

Source

idempotency_key = actor_run_id + ":" + event_type

Source