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