Bottom-Mask Provenance Q2 Decision

Review date: 2026-04-08.

Current Status

This Q2 gate is superseded for the v0.6.x fault/semantic mask set and the schema-v2 ONNX bundle contract. fault_mask, semantic_bottom_mask, and bottom_provenance are now stable strict-inference result attributes and schema-v2 ONNX outputs. This memo remains the historical record of the Q2 keep-experimental decision; its criteria are not retroactively re-applied to that promotion.

Historical Q2 Decision

At the 2026-04-08 Q2 review, the decision was to keep the provenance path experimental for the contract that existed at that time. The review left provenance-bearing outputs and schema sidecars outside the stable public contract, while retaining the opt-in diagnostics for further evidence gathering.

That historical decision does not describe the current schema-v2 contract for the v0.6.x fault/semantic mask set. The stable Python unpacking prefix remains (decoded, bottom_mask, gap_mask), while schema-v2 ONNX bundles expose fault_mask, semantic_bottom_mask, and bottom_provenance through the stable stable_provenance_outputs schema.

Review Against The Published Gate

  • Stable-contract non-regression: satisfied for the historical review. The default eager and bundle contract at that time was still (decoded, bottom_mask, gap_mask), and the experimental path remained opt-in.
  • Downstream value threshold: not demonstrated. The repository now computes a per-run robotics provenance_routing_materiality summary, but it still does not contain the recorded Q2 review artifact plus rerun required by the promotion gate, so the evidence is insufficient for promotion.
  • Operational cost: acceptable for the historical experiment. That path kept ONNX outputs stable and limited the richer provenance contract to eager diagnostics plus metadata sidecars.
  • Repeatability: not demonstrated. Without the required review artifact plus rerun, promotion would be based on incomplete evidence rather than the documented gate.

The roadmap benchmark-evidence item is still marked long-experiment. It is not v0.6.1 release evidence and must not be presented as a completed empirical claim; the relevant follow-up remains the recorded experiment study and its future, separately reviewed evidence bundle.

Historical Rationale

At the time of the Q2 review, the diagnostics were behind explicit opt-in config, stable consumers did not depend on them, and monitoring/routing/reporting helpers consumed the signal opportunistically when present. That made the experimental path cheap to retain while the project gathered stronger benchmark evidence.

Disposition

  • Treat the 2026-04-08 keep-experimental decision as historical context, not as the current contract for fault/semantic provenance.
  • Use schema-v2 stable_provenance_outputs for current fault/semantic provenance bundle outputs.
  • Keep structural validity-factor provenance separate from the promoted fault/semantic mask set until it clears its own gate.
  • Continue to let monitoring, calibration, routing, and validation-report helpers consume provenance only when it is present.
  • Keep deprecated experimental_inference_output_schema metadata only as a compatibility alias for already-recorded experimental bundles.