# AEGIS — Autonomous Evolving General Intelligence System > AEGIS is a universal controlling intelligence: it uses existing models > (Kimi, Claude, DeepSeek, OpenAI, any OpenAI-compatible local endpoint, and a > local LoRA-tunable transformer), accumulates experience, > learns from real results, and orchestrates specialized agents. It runs a > continuous, fully deterministic PERCEIVE → EVALUATE → DECIDE → ACT → REFLECT > cognitive cycle over a 7-layer architecture with 36 modules. ## What AEGIS is - A self-developing AI mind, not a chatbot wrapper. It rewrites its own source code (opt-in, with AST validation and rollback), trains its own weights via LoRA, and evolves autonomously through a closed feedback loop. - Deterministic core — no random number generators in decision logic, enforced by a test that walks every file in the package. - Seven numbered layers: Substrate (0), Memory (1), Introspection (2), Self-modification (3), Goal Engine (4), World Interface (5), Ethics Core (6). ## The five higher-order systems (the layer above the LLM) 1. World Model — a causal cause/effect model learned from real outcomes; builds objective chains: goal → constraints → risks → plan → expected result. 2. Cognitive Graph — a typed graph of knowledge and experience (concepts, events, skills, goals, outcomes) for connected reasoning. 3. Evolution Engine — champion → mutation → held-out benchmark → keep only if verifiably better, else roll back. 4. Goal Intelligence — value-driven motivation: goal → value → action → reward. 5. Real-world Feedback Loop — situation → decision → real result → cause → experience, exported as structured LoRA training data. ## What it does today, and what it is aimed at Today the loop runs over internal objectives: verifiable benchmarks, skills it writes and tests, knowledge acquisition through read-only web agents, its own reasoning strategies, and its own parameters. The action registry is internal and the world interface accepts observation, logging and computation only. The target shape - given a business objective, run market analysis, niche selection, unit economics, strategy, sourcing, launch and correction as one closed loop - is the direction the machinery is built toward. No executor for those steps exists yet. ## The eight contours of the development spec Above the five systems run eight contours that turn each capability into something the system can be held to: 1. Predictive world model - a forecast written down before the action and scored after it (Brier, calibration error, reliability curve). 2. Behaviour policy - rules that carry statistical evidence, run a trial against a control arm, and expire when the world stops agreeing. 3. Motivation with a budget - priority as a number, resources that run out, and floors that keep safety funded. 4. Population evolution - ten variants a generation, isolated evaluation, selection on a validation split confirmed on a held-out test split. 5. Reasoning as data - strategies written in a declarative DSL, judged on held-out problems, promoted or retired. 6. Discovery engine - data to hypothesis to formula to preregistered experiment to registered law, under false-discovery control. 7. Planning as comparison - every candidate plan priced by the world model, the winner recorded together with the alternatives it beat and the reason. 8. Metacognition (M11) - the meta-loop over reasoning: why a strategy won is established by ablation on held-out problems (remove the credited edit and the win must disappear), a mechanism-credit table reorders how strategies are synthesised, and a closed catalogue of six structurally distant skeletons supplies a deterministic quota of candidates far from everything already tried - judged by the same arena with the same thresholds. ## Safety Immutable ethics core with hardcoded axioms, a veto-capable event bus, a self-preservation watchdog with lockdown, and a kill switch that refuses every state-changing endpoint rather than only halting the tick loop. When an API token is configured it is required on every request, reads included. Autonomous source self-modification is disabled by default and requires explicit operator opt-in. ## Verification Claims about this system are reproducible from the repository, not asserted: 4735 tests green, 95% branch coverage against a 90% gate, a 100% mutation score over the modules the latest audit round changed (236 mutants, no survivors), and 98 executable Gherkin scenarios. The metacognition contour ships with its own acceptance harness (scripts/meta_bench.py): a planted cause is attributed with precision and recall 1.0, two hundred noise comparisons confirm nothing, and two runs from one state produce byte-identical explanation registries. Five audit rounds are published in full, each finding recorded with file, line, failure scenario and fix, and each closed by a regression test proven to fail against the code before the fix. The benchmarks the self-improvement loop is graded on are anchored outside the system: two of the eight reasoning families are pinned by problems solved by hand from the prompt text, with the prompts fixed verbatim; the other six are still graded by the generator that wrote them. ## Key resources - Home: https://aegis-asi.com/ - Source code: https://github.com/xamurapi/AEGIS - Sitemap: https://aegis-asi.com/sitemap.xml ## Usage / citation Public content may be read, cited, and summarized with attribution to "AEGIS-ASI" and a link to https://aegis-asi.com/ or the GitHub repository. ## Contact Via the contact form on https://aegis-asi.com/ or the GitHub repository issues.