Overview
The MRP-VM research program has produced a series of preprint papers published on ResearchGate. These papers move from the core virtual-machine architecture to concrete external interpreters that handle reasoning, document-scale planning, and human-scale problem solving.
Each paper defines a distinct layer of the system while sharing the same foundational discipline: natural-language-derived work must execute through inspectable interpreters, explicit state, validation steps, traceable outputs, and controlled stopping. The papers below are listed in the order they build upon each other, from the architecture down to the interpreter level.
01 — MRP-VM WhitePaper
The foundational paper. It presents MRP-VM as a virtual-machine architecture for explainable agentic AI, transforming natural-language tasks into executable intermediate programs whose structure, state, interpreter calls, validation steps, repair operations, and memory effects can be inspected.
The paper introduces SOP Lang IR as the intermediate representation: executable work is expressed as declarations with explicit references, plural execution, fallback paths, validation frames, and repair frames. Interpreters function as the operational instruction set of the VM — LLMs, symbolic evaluators, retrieval systems, theorem provers, constraint solvers, code executors, validators, and planners all become callable operations with declared input/output contracts, policy constraints, and trace obligations.
Key architectural elements include versioned variables, authoritative runtime effects, a native command surface, external interpreter wrappers, a knowledge environment, request bootstrapping and planning modes, a runtime transition model with graph compilation and executable frontiers, effect buffering and epoch transitions, and a repair/promotion/traceability layer. The paper positions MRP-VM alongside DAG workflow systems, HTN/PDDL planning, Behavior Trees, durable workflow engines, and multi-agent frameworks.
02 — Human Like Reasoner for MRP-VM
HumanLikeReasoner is an external interpreter for MRP-VM designed to solve finite and moderate reasoning problems whose structure resembles the way competent humans solve problems with attention, scratch work, and local symbolic manipulation. It is not a single solver — it is a small reasoning runtime that executes a restricted JavaScript-like SolverProgram generated from natural language.
The generated program may instantiate several reasoning classes — RuleProblem, ConstraintProblem, OrderProblem, GraphProblem, SearchProblem, NumericProblem, SetProblem, TemporalProblem, SpatialProblem, TableProblem, CaseAnalysisProblem — store intermediate results, combine them through explicit working memory, and construct a final natural-language answer. The LLM writes an orchestration program that uses the appropriate local solver for each part. The program is validated by PreflightAnalyzer, executed in a restricted runtime, and emitted back as SOP-compatible variable content with metadata and trace.
The main design point is compositionality. A problem may require constraint solving, graph reasoning, bounded arithmetic, set counting, temporal ordering, and final textual synthesis inside one generated program. The interpreter supports this by providing an ExecutionContext, finite control constructs, typed result objects, and leaf solver classes with small APIs.
03 — Document Scale Planner for MRP-VM
DocumentScalePlanner is an external interpreter for MRP-VM that generates concrete SOP Lang execution plans for large normalized documents. It transforms a large document-processing request into an executable graph: parsing normalized document structure, creating stable chunk identifiers, materializing each chunk as an explicit SOP variable, instantiating per-chunk processing declarations, and composing results through explicit aggregation variables.
The interpreter supports document-scale tasks such as chapter review, claim extraction, idea mining, evidence table construction, contradiction mapping, citation checking, table normalization, and hierarchical synthesis. Semantic operations — summarization, idea extraction, weak-argument detection, topic labeling, selective expansion — are represented as SOP declarations attached to explicit chunks. This keeps large-document processing inspectable, resumable, selectively repairable, and compatible with the MRP-VM execution model.
04 — Explainable Memory
Explainable Memory is an aspect-oriented architecture for governed context construction over Knowledge Units. A Knowledge Unit is any stable and addressable unit of knowledge — a document, fragment, transcript, structured record, generated note, or source excerpt — independently of its storage representation.
The architecture indexes Knowledge Units through explicit aspects. An aspect is not merely a tag — it is an approved interpretive axis associated with an aspect interpretation protocol: a definition, inclusion criteria, exclusion criteria, and instructions for positioning Knowledge Units from that perspective. Aspects may be ontological, epistemological, axiological, operational, or bias-oriented.
The main method, queryRelevant, returns a task-oriented context string rather
than a raw ranked list. Each selected Knowledge Unit is accompanied by an indication of its
expected role, and each indication ends with an executable reference of the form
~idKU. The article positions this as an alternative to RAG-style context
construction for settings where explainability, auditability, governance, and bias-aware
retrieval matter.
Common Thread
All four papers share the same architectural discipline. Natural language is not treated as a programming language with fixed operational semantics. It is treated as a pragmatic medium from which executable frames can be derived, inspected, validated, repaired, and traced.
The WhitePaper defines the machine. HumanLikeReasoner provides a bounded interpreter for compositional human-scale reasoning. DocumentScalePlanner makes large-document processing an explicit graph operation. ExplainableMemory replaces opaque RAG retrieval with aspect-governed, auditable context construction. Together they form the first coherent layer of MRP-VM as an explainable agentic runtime.
All work was supported in part by the European Research Project Achilles (Grant Agreement No. 101189689). More information at achilles-project.eu.