CoreAI_HTCE CoreAI released HTCE-Origin v1.0, a Q256 integer-only toroidal cognitive runtime designed for bounded, evidence-gated reasoning and adaptive memory. The system operates as a single runtime object with L1/L2/L3 state transitions, proof-gated memory, and active-agent simulation, intended for simulation-only adaptive agency and auditable memory. The release explicitly does not claim AGI, consciousness, or real-world autonomy. HTCE-Origin v1.0 https://private-user-images.githubusercontent.com/99619093/607630420-e8d3c161-6366-4b01-9548-7f063bcece04.png?jwt=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.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.O3MH2cPveyd0PJ jDuV9aAgvfQGgodSRE 8xV7fmnEE Q256 integer-only toroidal cognitive runtime with L1/L2/L3 state, proof-gated memory, bounded active-agent simulation, continual adaptive memory, and no-cross-domain regression probes. HTCE-Origin v1.0 final math Q256 clean A clean, single-runtime, simulation-safe cognitive architecture for evidence-bound reasoning and adaptive memory. What HTCE-Origin Is what-htce-origin-is What HTCE-Origin Is Not what-htce-origin-is-not Why This System Exists why-this-system-exists Core Design Principle core-design-principle Mathematical Foundation mathematical-foundation Architecture Overview architecture-overview L1/L2/L3 Cognitive Stack l1l2l3-cognitive-stack AIR: The Bounded Runtime Language air-the-bounded-runtime-language Fact-as-Delta Memory fact-as-delta-memory Proof, Evidence, Policy, and Topology Gates proof-evidence-policy-and-topology-gates Living Simulation Loop living-simulation-loop Adaptive Memory and Continual Learning adaptive-memory-and-continual-learning No-Cross-Domain Regression no-cross-domain-regression Protected Trace and Release Integrity protected-trace-and-release-integrity What the v1.0 Release Proves what-the-v10-release-proves How HTCE-Origin Differs from Common AI Stacks how-htce-origin-differs-from-common-ai-stacks Repository Layout repository-layout Installation installation Quick Start quick-start Acceptance and Verification acceptance-and-verification Example Runtime Interactions example-runtime-interactions Generated Artifacts generated-artifacts Claim Boundary claim-boundary Development Philosophy development-philosophy Roadmap After v1.0 roadmap-after-v10 License and Commercial Use license-and-commercial-use HTCE-Origin is a clean v1.0 build of a Q256 integer-only toroidal cognitive runtime . It is designed as a bounded, auditable, simulation-safe cognitive system where memory, reasoning, dialog state, action policy, homeostasis, topology checks, and adaptive learning all run through one runtime object: HTCERuntime The v1.0 build is not a set of disconnected demos. The same runtime carries: - natural-language-to-AIR intake for bounded statements and dialog turns; - L1/L2/L3 toroidal state transitions; - fact-as-delta memory; - latest-state, supersession, and contradiction quarantine; - proof, evidence, policy, and topology gates; - active-agent grid-world simulation; - dialog slot memory and simulated API-call policy; - sleep/L3 consolidation; - continual adaptive memory; - multi-task no-regression probes; - protected trace and cryptographic release manifest. The correct short description is: HTCE-Origin v1.0 is a bounded, evidence-gated, torus-native cognitive runtime for simulation-only adaptive agency and auditable memory. This repository intentionally does not claim: - Artificial General Intelligence; - consciousness; - qualia; - biological life; - real-world robotic autonomy; - board-measured hardware performance; - unrestricted open-world natural language understanding; - replacement of large language models for broad text generation. All real-world action is disabled in this clean release: allow real actions = false simulation only = true The system is allowed to act only inside bounded simulation paths unless a future audited release explicitly changes that boundary and passes a separate safety case. Most AI applications today are built around one of the following patterns: - a large language model prompt; - a retrieval-augmented generation pipeline; - a tool-using agent wrapper; - a neural policy model; - a symbolic rules engine; - a hybrid orchestration layer. These systems can be useful, but they often lack built-in runtime invariants for: - refusing unsupported answers; - preserving evidence provenance; - separating active facts from superseded facts; - quarantining contradictions; - preventing answer-key leakage during evaluation; - proving that learning in one domain did not damage another domain; - enforcing simulation-only action boundaries; - keeping memory updates auditable through a protected trace. HTCE-Origin addresses this by treating cognition as a bounded state transition system over an integer toroidal phase space. Instead of treating memory as a text buffer or hidden vector alone, the runtime treats events, facts, rules, sensor packets, goals, and dialog slots as phase deltas and gated state transitions. The core design rule is: No answer or simulated action is allowed to bypass the runtime gates. Every meaningful transition must pass through some combination of: Input / NLU / AIR → Policy gate → Evidence gate → Proof layer → L1/L2/L3 state transition → Topology guard → Protected trace → RuntimeResponse This makes the system slower and narrower than unconstrained text generation, but it gives the runtime a very different property: it is optimized for bounded correctness , traceability , and non-hallucinating refusal , not unconstrained fluency. The runtime uses an integer modulus: A toroidal state vector is represented as: A transition is a modular update: where: - $\mathbf{x} t$ is the current toroidal state; - $\Delta t$ is an integer phase delta derived from observation, fact, rule, or simulated action; - $N$ is the Q256 modulus; - $d$ is the fixed state dimension. The decision path is integer-only. Human-readable percentages or basis points may appear in reports, but they are report-layer summaries, not floating-point decision variables. For one coordinate, the circular distance is: For a vector state: This gives the runtime a native topology-aware notion of change: large modular jumps can be detected as anomalous even if the raw representation wraps around. A fact is not treated as plain text. A supported fact produces a canonical delta: where: - $s$ is the subject; - $r$ is the relation; - $o$ is the object/value; - $e$ is the evidence identifier. The fact key is usually relation-specific: The fact payload is hashed into a bounded integer delta: The memory status is one of: ACTIVE SUPERSEDED QUARANTINED UNKNOWN This supports latest-state reasoning while keeping old history available for trace and audit. When a new fact shares the same key and the new fact becomes active: This is used for locations, dialog slot corrections, and other latest-state updates. Example: FACT Mary located in office EVID ev1 FACT Mary located in garden EVID ev2 QUERY Mary location EVID q1 The runtime should answer: ANSWER: garden while preserving the previous office state as superseded evidence. A contradiction is not treated as a normal update. If a fact and its negation cannot be reconciled, the target key is quarantined: A query against a quarantined target must not produce a normal answer: This is one of the core anti-hallucination principles of the runtime. The trace is a hash-linked sequence of events: where: - $h t$ is the current trace head; - $h {t-1}$ is the previous trace head; - $event t$ is the canonical event payload; - $H$ is SHA-256 in the release artifacts. Any change to the event sequence changes the final trace head. The living simulation uses an integer expected-free-energy-like cost. In this release it is a bounded surrogate, not a claim of biological free energy minimization. A simplified form is: The selected simulated action is: where all costs are integer raw values. No real action is authorized by this process. For continual adaptive memory, an episode cost can be represented as: where: - $S E i $ is the number of simulated steps; - $Q E i $ is the number of clarification turns; - $R E i $ is the number of recovery or risk events. For a repeated target, improvement is verified when: or, for a strict improvement claim: For multi-domain learning, each domain No cross-domain regression means: In the v1.0 clean revalidation artifact, the domain cost histories remain bounded at zero for the built-in P28 domains. At a high level: User / Benchmark / Simulation Input │ ▼ NLU → AIR bridge / AIR parser │ ▼ Policy gate ── Evidence gate ── Proof layer │ ▼ HTCERuntime │ ├── L1 sensory torus ├── L2 episodic/fact memory ├── L3 semantic/provisional rule cortex ├── Q256 world model ├── active-inference surrogate planner ├── homeostasis state ├── topology guard ├── snapshot/export path └── protected trace │ ▼ RuntimeResponse: ANSWER / ASK CLARIFICATION / REFUSE / ACT SIMULATED / BLOCK The most important architectural rule is that the system should not have multiple independent behavioral shells. Dialog, grid-world, active-agent behavior, proof-gated action, and continual adaptation are all routed through the same runtime line. L1 is the fast toroidal sensory layer. In v1.0, it is used primarily for bounded simulation input and deterministic sensor packets. L1 handles: - grid-world observations; - local agent position; - heartbeat updates; - deterministic integer sensor encoding; - active-agent loop state. It does not claim full real-world vision/audio/proprioception. Those require future audited encoders. L2 stores active facts, superseded facts, and quarantined contradictions. It supports: - latest-state query; - object carried-by relation; - location chaining; - dialog slot memory; - correction through supersession; - contradiction quarantine; - evidence-linked fact records. Examples: FACT current dialog restaurant 1 has slot value cuisine=italian EVID d1 FACT current dialog restaurant 1 has slot value price=cheap EVID d2 QUERY current dialog restaurant 1 api call ready EVID q1 If location is missing, the runtime must ask for clarification rather than fabricate a location. L3 stores provisional hints and rules produced by sleep/consolidation. It is intentionally restricted: - L3 can propose a hypothesis; - L3 can help select a path; - L3 can store adaptive hints; - L3 cannot bypass proof gates; - L3 cannot directly authorize unsupported factual answers. A valid runtime response must still pass through proof/evidence/policy checks. AIR is the controlled interface language used by the runtime. It prevents arbitrary text from directly becoming authority. Typical forms: FACT