Description
Knowledge Verification Workflows test cited evidence, probe links safely, detect opposing claims, and decay confidence when material goes unreviewed.
Separate behaviors handle source support, URL health, contradiction candidates, and time-based confidence signals so each concern has its own configuration and evidence. They return rationale-bearing effects or events through the automation engine. Trusted verification remains a human-controlled state: a model or network probe can identify work, but it cannot silently certify a claim.
Selling Points
- Separates verification proposals from trusted human stamps.
- Distinguishes blocked URLs from genuinely dead links.
- Produces rationale-bearing contradiction signals.
- Turns several kinds of knowledge risk into reviewable, bounded work.
- Keeps transient provider and network failures distinct from negative evidence.
User Story
As a knowledge steward, I can prioritize unsupported, contradictory, or aging claims before they mislead the team. I bind the checks to the fields and thresholds used by my Dotabase. Scheduled or row-driven runs gather evidence and surface specific reasons for review. I inspect the cited source, blocked link, possible contradiction, or elapsed interval and decide what to correct or verify. I walk away with a risk queue, not an automated claim of truth.
Extension Surface
Configurable: owners select source fields, thresholds, decay policy, and behavior bindings.
packages/dotabases/src/behaviors/source-verification/index.ts demonstrates the binding and provider boundary. Owners choose eligible fields and policy, while hardened output validation and the rule that automation cannot grant a trusted stamp remain fixed.
Capabilities & Limits
- Capability: Executes the scoped workflow described above through the cited live implementation paths.
- Capability: Preserves typed validation and applicable tenant, schema, lifecycle, or service boundaries.
- Limit: Source judgment and contradiction checks need provider seams; verification never auto-stamps a row as trusted. The workflows can test configured citations, classify link outcomes, propose contradictory pairs with rationale, and emit staleness or confidence-decay signals. They cannot prove truth from unavailable evidence, bypass network safety, or replace a reviewer. Provider, budget, and access failures must remain observable non-verdicts.
Implementation Map
packages/dotabases/src/behaviors/source-verification/index.tspackages/dotabases/src/behaviors/link-health/index.tspackages/dotabases/src/behaviors/contradiction-detection/index.ts