Assessment & application roadmap
A self-critical assessment of the framework, with the applications prioritised by how cleanly they can expose falsifiable signatures.
Jump to a profile
Each profile below is linked from the research-library home page and points to a decisive way to test the framework.
Fit verdict. The strongest scientific application because an experimenter can independently control the offered actions, information access, and hidden consequences.
A decisive first study. Use a virtual-resource task with one menu-visible dead end and a separate post-choice environmental trap, while varying depth, outcome variance, and menu contractions.
Fit verdict. A high-value diagnostic use: the framework becomes a behavioral fingerprint, rather than a claim about the unknown agent's internal implementation.
A decisive first study. Create a sealed benchmark with repeated paired menus, a variance reversal, a depth ladder of menu traps, and matched environment traps.
Fit verdict. A strong engineering testbed because recovery states and primitives are discrete, observable, replayable, and auditable.
A decisive first study. Deploy the service-recovery benchmark in a containerized service with controlled restart storms, partitions, burst load, and misleading telemetry.
Fit verdict. A plausible narrow runtime-assurance layer only when its action set and safe envelope are independently verified; the framework itself does not provide certification.
A decisive first study. On a validated plant or hardware-in-the-loop rig, compare strict E-blind fallback control with conventional verified and predictive controls using the same macros.
Fit verdict. A promising falsifiable null model of non-anticipating regulation in controlled perturbation experiments, not yet a claim that cells literally solve the selector equation.
A decisive first study. Use a controlled flow system with fluorescent reporters, matched perturbations, and response branches that separate endogenous dead ends from later external pulses.
Fit verdict. An excellent domain for the evolved-horizon prediction because within-lifetime policy and between-generation selection can be manipulated separately.
A decisive first study. Evolve populations in matched worlds, then introduce post-choice hazards whose cues are absent from the declared state and action menu.
Fit verdict. A strong engineering application for debuggable behavior, provided authored primitives do not conceal a world model.
A decisive first study. Build a settlement scenario with two trap types and ask designers to diagnose failures from decision logs before seeing source code.
Fit verdict. A plausible research application for local, auditable recovery in decentralized systems, with a demanding boundary between observed neighbours and the environment.
A decisive first study. Simulate and then physically test a small swarm under node loss, including a local menu trap and a simultaneous-response environment trap.
Full document