Skip to content

Run Analysis Evidence Model

explain_run(...) produces a canonical analysis artifact for a validation run.

Canonical Inputs

The analysis pipeline resolves exactly one run at a time by run object or run_id. It then assembles a read-only evidence bundle from:

  • persisted run metadata
  • deterministic failed-check extraction
  • deterministic top-column extraction
  • history window summary
  • baseline summary when unambiguous
  • docs artifact reference when available

Deterministic vs Model-Owned Fields

Truthound intentionally splits analysis output into deterministic and model authored parts.

Deterministic fields:

  • run_id
  • failed_checks
  • top_columns
  • history_window
  • input_refs

Model-authored fields:

  • summary
  • recommended_next_actions
  • evidence_refs

This division keeps operational facts reviewable even when provider behavior changes.

Canonical One-Per-Run Rule

Phase 2 fixed analysis persistence to one canonical artifact per run:

  • artifact id: run-analysis-<run_id>
  • rerun behavior: overwrite the same canonical path
  • no ad-hoc question variants in the public contract yet

That choice simplifies dashboards, indexing, and approval-adjacent review surfaces.