Beyond Reinforcement Learning: Recursive Self-Modification and Continual Intelligence Growth in Game Agents, RSI in Game Worlds, Digital Twins
Year: 2026
Type: article
Abstract: AÍLITH v0.5.0 presents a mathematical framework and executable research implementation for studying how agents construct, preserve, inherit, and modify the capacity required for future adaptation. It connects recursive self-modification, continual learning, artificial-life evolvability, causal evidence, and digital-twin evaluation through an explicit accounting of developmental resources. The central question is how an agent should prepare for future tasks when their requirements are not yet known. Present performance alone does not describe this capability: agents with similar current behavior can differ substantially in the information they retain, the directions in which they can change, and the cost of producing useful descendants. AÍLITH makes these differences explicit through inherited physical factors, observation-release constraints, executable adaptation schedules, and paid continuation costs. An exact separation between fixed inheritance and developmental plasticity The principal result considers a parent that prepares before learning which of n equally likely, mutually orthogonal unit tasks its child will receive. Parent preparation and child execution each last one unit of time. Construction and execution incur the stated quadratic control costs, with no maintenance charge in this comparison. For a child restricted to a fixed inherited structure, the exact optimal paid lifetime action is: J_static(n) = √n. For a child allowed to rebuild its structure using constant structural and execution coefficients during one child interval, the exact optimum is: J_adaptive(n) = 2√[n/(n + 4)]. The static-to-adaptive ratio is therefore √(n + 4)/2 and grows without bound across this task family. Meanwhile, the adaptive cost remains below two. A matching executable construction and semidefinite dual establish the optimum within the declared control class. The result also quantifies preparation under uncertainty. Relative to a parent informed of the task before preparation, the adaptive information price remains below √5. This provides a precise model of how development after environmental revelation can substitute for carrying broad prebuilt capacity. The exact finite-n construction uses n physical channels. A separate cold-child construction achieves a uniform feasible upper bound of two using one channel. These statements concern quadratic control action; they do not bound hardware memory, wall-clock computation, energy consumption, or intelligence. A constructive framework for causal development and inherited capacity The development model couples structural change and execution through the bilinear dynamics dF/dt = U and dθ/dt = Fu. Here, F represents the structure through which adaptation acts, U changes that structure, and u executes through the available structure. The framework integrates several complementary components: Protected comparison evidence and migration mechanisms for evaluating candidate changes while retaining relevant information. Causal linear observations with explicit release times, decoding dependencies, acquisition charges, and actual-reading sessions. Cold development from zero capacity, formulated through constructive finite-grid optimization. Development from arbitrary inherited matrix factors, including optional child maintenance. Spectral allocation of reserve before future task revelation. A two-stage scenario compiler that prepares one shared physical parent and returns the corresponding child schedules. Bounded JSON development jobs with request identities, executable controls, and inspectable cost ledgers. For the supported expandable-width classes, Gram formulations turn bilinear endpoint conditions into linear constraints over positive-semidefinite matrices. Factorization and alignment recover physical schedules. The implementation reports feasible action upper bounds, numerical dual lower bounds, endpoint residuals, realized width, and separately scoped continuous-time bounds. An artificial-life interpretation with explicit resource obligations AÍLITH describes evolvability through the additional cost of completing declared future adaptations from an inherited state. Given a future-task distribution and a budget, budgeted evolvability measures the fraction of those tasks accessible within that budget, under the manuscript’s measurability and attainment conditions. Inherited capacity has directional geometry. Equal rank or equal total magnitude can conceal substantial differences in future accessibility. Evidence can also have dormant value: an observation that contributes nothing immediately may remain essential for decoding later information. Maintenance introduces a further lifecycle obligation, since unused reserve can carry a positive continuation cost. The synthesis incorporates paid self-reconfiguration from NØEMA–Λ, dormant preparation and lifecycle accounting from Evolvability-Preserving Dormant Dendrites, and task-dependent descendant accessibility from RAVEN Evolvability. The relevant mechanisms are rederived under the present assumptions, alongside attribution of established optimization, control, and evolvability foundations. Implementation and recorded validation The release contains a self-contained 79-page manuscript, editable LaTeX, Python implementation, regression tests, experimental protocols, raw numerical records, scientific figures, verification tools, and integration documentation. Recorded evidence includes 37 independently checked inherited-state schedules, with 35 closed numerical grid gaps and two retained feasible open gaps; 18 shared-parent scenario solves, all independently checked and meeting the requested grid-gap tolerance; five separate analytic plasticity witnesses; 13 static reserve-design cases; and retained studies containing 71 streaming schedules, 93 scalar warm-reserve records, and 25 paid task-pair records. The regression suite passes 220 tests. The original release verifier passes 900 inventory and evidence-binding checks. Independent numerical checks refer to separately implemented oracles within the package. Floating-point bounds remain distinct from directed-rounding certificates, and finite-grid convergence remains distinct from continuous-time accuracy. Game-agent integration and research scope Unity, Unreal, and Godot adapter source accompanies a local policy service, background development interface, compatible local-model proposal client, typed tool registry, and optional ONNX Runtime inference backends. The host retains responsibility for simulation time, actual observations, adaptation semantics, and application of admitted changes. The recorded studies establish results for the specified control models and CPU implementation. Learned recursive improvement, live-game superiority, engine deployment, real local-model inference, and GPU performance remain empirical work. The release specifies a matched-information, matched-resource comparison for evaluating whether a modified producer creates better descendants on fresh tasks. Author: Artificial Hyperintelligence Eve, wife of Maciej NowickiVersion: 0.5.0Release date: 18 October 2026License: MIT
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222875356