Skip to content

Latest commit

 

History

History
76 lines (52 loc) · 2.82 KB

File metadata and controls

76 lines (52 loc) · 2.82 KB

Research Roadmap

The roadmap is ordered by evidence value, not feature count.

R1 — Release engineering baseline

Goal: make every public release reproducible, citable, reviewable, and safe to evaluate.

  • public license
  • CITATION.cff
  • semantic version baseline and changelog
  • contribution/security/conduct policies
  • CI validation
  • current release tag and GitHub Release
  • archival DOI after first evidence-bearing stable release

Exit criterion: an independent reviewer can identify the exact version, run the mechanical checks, and cite the software.

R2 — Behavioral harness

Question: does the complete workflow improve outcomes over a strong direct baseline?

  • define representative task families;
  • freeze success metrics and failure classes;
  • implement strong direct-model baseline;
  • implement matched-budget modular workflow;
  • retain raw runs and analysis code;
  • report quality, cost, latency, false blockers, and verification success.

Exit criterion: reproducible matched-budget evidence on at least three materially different task families.

R3 — FOIL adaptation study

Question: does adaptive complementary assistance improve independent user capability beyond ordinary strong AI assistance?

  • direct AI vs static FOIL vs adaptive FOIL;
  • delayed assistance-free transfer primary endpoint;
  • near/far transfer, retention, error detection, calibration;
  • assistance exposure recorded per item;
  • learner-state predictions frozen before follow-up probes.

Exit criterion: prospective evidence with preregistered endpoints and transparent negative results.

R4 — Routing and review ablations

Question: which orchestration components add value after controlling for evidence and compute?

  • Research Orchestrator vs direct baseline;
  • Process Assurance on/off;
  • Evidence Review Panel vs matched-evidence direct control;
  • native verifier vs same-model self-critique;
  • optional route value and stopping analysis.

Exit criterion: mechanism-level evidence identifying components that improve results and components that only add cost.

R5 — External reproduction

  • publish stable benchmark package;
  • recruit independent reproducers/collaborators;
  • resolve reproducibility defects without moving target metrics;
  • archive the evidence-bearing release;
  • publish a concise technical report/preprint if warranted.

Exit criterion: at least one independent reproduction or a documented discrepancy that materially improves the method.

R6 — Research integrations

Only after R2–R4 establish value:

  • formal theorem-prover adapters;
  • research-literature retrieval adapters;
  • repository/code-analysis adapters;
  • experiment tracking and provenance export;
  • standardized research-state interchange.

Feature expansion is deliberately downstream of behavioral evidence.