Daniel du Plessis proposes using Jev to interpret evidence and explicit rules to combine the results. His essay connects decision models to truth maintenance: keeping track of which facts and assumptions support a conclusion, then revisiting it when evidence changes.

Give each judgment a smaller job

The essay’s family-tree example separates recorded relationships from the rules that connect them. A model can assess whether a passage says one person is another’s parent. Code can then establish the resulting ancestry.

  • Keep domain rules available for review, alongside the definition of each fact the model assesses.
  • Record which source passage supports or denies each fact.
  • Trace conclusions through their supporting rules and evidence.
  • Revisit conclusions when supporting evidence is corrected or withdrawn.

This gives reviewers a concrete procedure to inspect. A convincing explanation generated after a decision offers a different kind of evidence.

Factbag has a real Jev integration

The linked Factbag project implements document classification through TypeSafe’s SDK. Its current code asks Jev whether a passage supports, denies or is irrelevant to each candidate fact.

  • Application permissions determine which predicates, or named claims, Jev may assert or deny.
  • Admission rules decide which probability scores become recorded evidence.
  • Source identifiers stay attached to that evidence for later inspection.
  • Explicit rules combine admitted facts; queries can expose a proof tree linking the result to its premises.

Use an explicit permission guard when restricting model writes. The inspected classifier treats a bare knowledge base as unrestricted access to its known predicates.

Keep disagreement visible

Factbag distinguishes true, false, conflicted and unknown. Conflicted means evidence exists on both sides; unknown means evidence is missing. Those states can direct a reviewer toward resolving disagreement or collecting another source.

Its logic uses a custom ordering of those states. The author explicitly distinguishes it from standard Belnap logic, the four-valued system that inspired it. Applications need to review what combined rules actually mean.

A trace still depends on the evidence

Typed output checks the answer’s shape. The passage interpretation, admission threshold and domain rules still need correctness tests. A deterministic rule engine can faithfully propagate a mistaken model judgment.

The essay’s speed and cost statements cite TypeSafe’s launch claims. It supplies no measured end-to-end evaluation of this architecture, and Factbag remains an early alpha.

“The workflow remains fully agentic, but leaves an inspectable procedural trail for identifying flawed rules, gaps in evidence, or errors in application.”

Verification: read the complete essay and its companion explanation; inspected Factbag’s README, Jev adapter, permission handling, admission policy and logic documentation. No package execution, live Jev calls or performance reproduction was performed.