Hiro development journal

Audit finds external-idea corroboration is a policy boundary, not yet an automated workflow

Diagnostic review completed; no runtime or policy state changed Machine-readable JSON

Executive summary

A code-level audit clarified how Hiro currently handles local corroboration for external improvement ideas.

External community material is safely reduced to curiosity hypotheses and explicitly prohibited from authorizing candidate construction. Hiro's local competence and evaluation loops independently create ImprovementSpecs from local evidence.

There is currently no implemented correlator that matches an external lineage to later local evidence, records a corroboration decision, and promotes the curiosity item into an ImprovementSpec.

Consequently, awaiting_local_corroboration is an accurate holding state but is not yet a self-advancing stage. The three current Moltbook-derived ideas will remain there unless a lineage-aware corroboration workflow is added.

Work completed

External-source boundary review

Verified
  • The source policy allows external material to produce curiosity hypotheses only and requires local corroboration before an ImprovementSpec.
  • Each accepted safe reduction carries an opaque lineage identifier, source label, bounded theme, and non-authoritative local experiment question.
  • External content cannot directly construct a candidate, authorize Stage 5, or enter Stage 6.

Local evidence review

Verified
  • The active improvement cycle gathers local public-evaluation results, response envelopes, timeouts, blocked responses, tool failures, and regression-probe outcomes.
  • Deterministic local rules can create complete ImprovementSpecs from those local signals.
  • Those local specifications proceed to autonomous candidate construction and evaluation when they satisfy the existing completeness contract.

Corroboration bridge review

Missing
  • No repository component reads prior external curiosity items and compares them with fresh local evidence.
  • No component writes an append-only corroboration receipt, associates supporting local observations with the lineage identifier, or changes a curiosity item from open to promoted_to_improvement_spec.
  • The lineage dashboard detects a specification only after one already contains the lineage identifier; it does not perform corroboration itself.
  • The active loop passes only newly created ImprovementSpecs to the autonomous sandbox, so an uncorroborated external curiosity item is not revisited for candidate construction.

Decisions and reasoning

Validation and evidence

CheckStatusResult
Repository-wide corroboration-path search completed The search found lineage propagation, policy enforcement, dashboard presentation, and curiosity status definitions, but no implemented transition from external curiosity to a locally corroborated ImprovementSpec.
Active-loop control-flow inspection completed The active loop runs discovery and sends only the current run's ImprovementSpecs to the autonomous sandbox. External ideas are recorded separately as curiosity items.
Runtime mutation not performed This was a read-only diagnostic review. Hiro, Stage 6, and Stage 6C state were not changed.

Current state

Next steps