research-document

Research TODO Backfill

Research TODO Backfill

Decision

Represent the repository's actionable research backlog as proposed ROS mission artifacts under missions/backlog/. Do not create active ROS transitions for work that has not started.

The canonical research/frontier/ already deduplicates most roadmap checkboxes, open questions, and workstream language into ten RFRs. Each RFR therefore maps to one backlog mission. NX-005 remains separate because its audit/no-audit counterfactual is not represented by a dedicated frontier record.

Mapping

Mission Canonical source Consolidated TODOs
MS-AIRES-2026-0001 RFR-001 / NX-001 independent pilot reviews and blinded review-label comparison
MS-AIRES-2026-0002 RFR-002 / NX-002 contained repeated baseline and exploratory variance
MS-AIRES-2026-0003 RFR-003 / NX-001 unseen-outcome and natural-outcome grader validation
MS-AIRES-2026-0004 RFR-004 / NX-004 representative suite and independent code/automation replication
MS-AIRES-2026-0005 RFR-005 context, state, memory, and retrieval-compiled policy comparison
MS-AIRES-2026-0006 RFR-006 total verified-cost and reviewer-burden measurement
MS-AIRES-2026-0007 RFR-007 fixed versus risk-triggered human escalation
MS-AIRES-2026-0008 RFR-008 / NX-003 containment, gold separation, monitoring, pause, and rollback validation
MS-AIRES-2026-0009 RFR-009 bounded metadata adoption and real supersession-edge calibration
MS-AIRES-2026-0010 RFR-010 knowledge-graph value experiment
MS-AIRES-2026-0011 NX-005 audit/no-audit decision counterfactual

Grouping rules

  • Duplicate checklist language is consolidated under the matching RFR mission.
  • A mission may cover multiple experiment aliases when they share one decision gate, but its body preserves the component evidence and success requirements.
  • Dependency order follows the frontier graph and evaluation critical path.
  • All new missions are proposed; beginning one requires a future authorized ROS work transition and prerequisite check.
  • Publishing deployment confirmation and Time Entry contract authoring remain engineering/product work and are not mislabeled as research missions.
  • Governance template checkboxes are completion criteria, not repository TODOs.
  • Illustrative uses of “TODO” inside research prompts and reports are examples, not actionable repository work.

Deferred broad themes

Model routing, interface-agent testing, orchestration variants, and general retrieval benchmarking remain roadmap themes rather than separately actionable missions because the current frontier makes them dependent on RFR-004 and later evidence. They should receive new RFRs and missions only when their prerequisites are met and a bounded decision question is defined.

Validation

The ROS mission registry must be generated with ./ros registry build. The backlog is valid only when ./ros registry check, ./ros validate, and the project test suite pass.