Claims and limitations
| Topic | Supported statement | Boundary |
|---|---|---|
| Paid plasticity | Exact static cost sqrt(n) and adaptive cost 2*sqrt(n/(n+4)) for the stated orthogonal task family |
Unit parent preparation and child interval, constant child controls, task revealed at interval start, no maintenance, expandable width |
| Relative separation | Static/adaptive action ratio grows without bound in this family | Quadratic action is not hardware time, RAM, joules, intelligence, or game win rate |
| Width | Exact finite-n construction uses n physical channels | Cold one-channel action two is a separate upper bound; restrictive widths may prevent an unrestricted optimum |
| Task uncertainty | The adaptive information price stays below sqrt(5) |
Public fixed task distribution; the prior is not learned in the theorem |
| Shared inheritance | The two-stage compiler returns one physical parent and a branch family | Branch-specific public task knowledge becomes available only after preparation |
| General inherited matrix | Feasible controls, optional maintenance, and explicit numerical bounds are returned | Two of 37 frozen cases retain open grid gaps; continuous bounds may be much looser |
| Numerical certificates | Float64 primal/dual and independent interval-formula checks | No directed rounding, formal proof assistant, or external replication |
| Evidence acquisition | Declared release times and invoked-read charges are enforced in process | Host clock and sensor truth are trusted; no OS isolation boundary |
| Evolvability | Declared future tasks are evaluated by attainable continuation cost | Covariance sufficiency requires the specified isotropic control class |
| Learned RSI | A controlled empirical test is specified | No measured advantage of a learned successor-producing agent |
| Games and engines | Policy service, development jobs, and engine adapter source are supplied | Actual game-engine builds and live gameplay were not measured |
| Local LLMs | Bounded proposal interface and dry-run examples are supplied | Real local-model inference was not measured |
| GPU | Optional ONNX Runtime CUDA actor backend exists | Frozen ONNX model execution is mocked; no GPU throughput or actual kernel execution evidence |
| Multilevel inheritance | A scenario-tree formulation is discussed under stated assumptions | Only the complete two-stage solver is implemented; no general scalable tree scheduler |
| Novelty | The work positions its formulation against cited predecessors | Worldwide priority, peer review, and best-existing RSI performance are not established |
The causal observation model assumes a fixed context during planning, known covariance and targets, exact Gaussian linear readings, deterministic shared preparation, and the stated quadratic costs. Protected comparison results use separately specified bounded latent assumptions. Unknown targets, noisy observations, moving worlds, arbitrary program modification, nonlinear constraints, and realistic actuator limits require further modeling.
Parent preparation is charged once. Sensor price, external tools, CPU/GPU work, memory, and code size require their own ledger or explicit objective extension. Child maintenance is included where declared; parent maintenance is excluded from the supplied two-stage preparation formula.
The historical v0.2/v0.3 sources and unavailable experiments were not reconstructed. The complete current derivations stand alone, while the recovered v0.1/v0.4 materials are retained as provenance.
The empirical next step is a fresh-descendant comparison between an unchanged and modified producer, with matched observations, inheritance, actor state, tools, model calls, and paid development allowance. Current reward, retention, shift behavior, and descendant-production cost should be reported separately.