14.8 Application probes

The broader experimental program asks whether this pattern survives outside planted chains.

10-K workflows.

Staged filing pipelines provide typed evidence, mechanism, and narrative artifacts. LASKO measures whether schema, evidence, citation, and validator repairs commute.

Democritus and causal PSR.

Trace-derived causal-claim workflows expose ordered edits to extraction, qualification, and validation. A predictive-state-representation (PSR) audit reaches the exhaustive served score with 72 cheap bracket probes and 19 served validations, versus 73 exhaustive validations.

SearchQA.

The gated sanity check finds that a failure-derived edit can carry a stronger local interaction signal without winning hard accuracy. This is evidence for retaining soft and fiber-level audits.

ALFWorld.

A micro-benchmark closes one live ordered repair edge from a failed rollout. It establishes executability of the workflow construction, not broad ALFWorld performance.

Experiment: the retained empirical hierarchy

The controlled studies establish validation savings under known sparse interaction structure. The PSR and workflow studies show that the machinery can run against served evaluators. SearchQA and ALFWorld remain scoped probes, not general agent-performance claims.