Research software is defined by its research role, not by whether a repository happens to be visible. This index separates what each system investigates, what it actually does, what evidence it retains, and what it does not claim.
These overviews are intentionally bounded. Public source, private research, simulation, and measured hardware are separate release states.
Investigates: capability boundaries, failure modes, and adaptive evaluation.
Actually does: probes a model endpoint, diagnoses failures, and searches the boundary of what the endpoint can do under a defined curriculum.
Evidence: probing traces and failure maps when a reviewed run exists.
Not claimed: a model trainer, a promotion authority, or universal educational capability.
Investigates: where a decision chain breaks between information, belief, action, and consequence.
Actually does: reconstructs decision-chain evidence for analysis and breakpoint diagnosis.
Evidence: chain records, diagnostic views, and failure analysis when the source boundary permits them.
Not claimed: a generic log viewer or an authority that changes the run it observes.
Investigates: long-horizon behavior under partial information.
Actually does: keeps observations, model decisions, policy, replay, and evaluator-held ground truth separate from the host world runtime.
Evidence: reproducible evaluation and replay artifacts within the public harness boundary.
Not claimed: a replacement for a game engine or permission for model output to own world state.
Investigates: objective-to-capability routing, verification, recovery, and bounded physical-AI development.
Actually does: exercises a simulation-oriented loop with routing and evidence receipts.
Evidence: mock and simulation results within the public boundary.
Not claimed: trained-VLA capability, hardware performance, or universal household behavior.
Investigates: bounded electrical, chemistry, and neuroelectric reference models.
Actually does: provides deterministic local reference paths with source provenance and immutable run packages.
Evidence: local reference outputs and run-package lineage.
Not claimed: clinical, biological, or synthetic-model validity beyond the stated reference boundary.
Investigates: portable native model execution, low-precision paths, memory ownership, and receipts.
Actually does: exposes a standalone execution boundary with local planning, checkpoints, integrity, and recovery.
Evidence: source contracts, tests, and bounded execution receipts where available.
Not claimed: that Hub coordination is required or that source presence proves hardware acceptance.
Research surfaces may include simulation, fixtures, probes, source-only contracts, unit verification, hardware observation, or burn evidence. The label stays attached to the result so a reader can tell which layer they are seeing.
A research system may be public, private, Open Canopy, source-available, or held out. The research role and the release boundary remain independent metadata dimensions.
