Overall progress12%
1

Mathematical Foundations

Dec 2025 — Jun 2026 · In progress

Developing the mathematics of computation — formalizing ternary vector algebra, proving its properties, and building simulations before writing production code.

Ternary algebra specificationCompleted

Formal definition of the ternary vector space — presence (+1), absence (0), negation (−1) — and the fundamental operations: BIND, SUPERPOSE, GATE, PERMUTE.

AlgebraSpecification
Consistency proofs & property testsCompleted

Proof of associativity, commutativity, and distributivity for every operation. Property-based testing across millions of random vectors.

ProofsVerification
Noise tolerance & capacity analysisCompleted

Formal bounds on dimensionality vs. capacity vs. noise tolerance — the numbers that set hardware requirements for every subsequent phase.

CapacityNoiseBounds
Encoding schema specificationIn progress

Rules for mapping structured data (records, sequences, graphs) into ternary vectors. Positional encodings and hierarchical composition.

EncodingSchemasMapping
Simulation & reference implementationIn progress

Python implementation of the full algebra for experimental validation of mathematical claims. An executable specification that proves the theory correct.

PythonSimulationValidation
2

Core Engine

Jun 2026 — Aug 2026 · Planned

Coding the core engine — translating the validated mathematics into a high-performance Rust runtime. Ternary vectors map directly onto machine words.

Rust core library (mindprism-core)Planned

Zero-copy, SIMD-optimized implementation of all four algebraic operations. Ternary vectors packed into a compact bit-level representation.

RustSIMDZero-copy
Memory layout & cache optimizationPlanned

Structure-of-Arrays layout for batch operations. Cache-line alignment. Benchmarks with linear scaling from 1K to 100K dimensions.

MemoryCacheSoA
Python bindings (py-mindprism)Planned

Native Python package via PyO3. NumPy-compatible. Researchers can call bind(a, b) from a Jupyter notebook.

PyO3PythonNumPy
Benchmark & regression suitePlanned

Continuous benchmarks for throughput, memory consumption, and algebraic correctness. Every PR runs 10M tests.

CriterionCIRegression
3

Text Modality & HAL Layer

Aug 2026 — Oct 2026 · Planned

Text modality, the Hardware Abstraction Layer (HAL), and a web interface for the text modality. The first perceptual channel: natural language.

Text encoder (token → ternary)Planned

Character-level encoding that maps token sequences into ternary vectors through positional BIND and hierarchical SUPERPOSE. Preserves word order, syntax, and semantic similarity.

NLPEncoderTokens
HAL — Hardware Abstraction LayerPlanned

Universal interface between raw modalities and the algebraic core. The HAL normalizes inputs from any modality into a single ternary vector space.

HALAbstractionArchitecture
Reasoning chains for textPlanned

Multi-step reasoning over text: premise → GATE → conclusion. Each reasoning step produces a human-readable trace — the system explains why it reached a given answer.

ReasoningXAITraces
Web interface for text modalityPlanned

Interactive web app: the user enters text queries and watches the reasoning engine work in real time. Visualization of vector similarity and reasoning traces.

Web UIVisualizationREST API
4

Audio Modality

Oct 2026 — Dec 2026 · Planned

Audio modality — sound as a first-class perceptual input. From raw waveforms to structured ternary representations.

Audio encoder (waveform → ternary)Planned

Spectral decomposition into frequency bands, each mapped to positions in the ternary vector. Temporal structure through sequential BIND operations.

FFTSpectralEncoder
Speech & sound classificationPlanned

Real-time classification of audio streams: voice commands, environmental sounds, speaker identification. With full interpretability.

SpeechClassificationReal-time
Cross-modal text ↔ audioPlanned

The first cross-modal demo: describe a sound in text — the system finds matching audio clips. Both modalities share a single vector space through the HAL.

Cross-modalHALRetrieval
Audio reasoning & anomaly detectionPlanned

Anomalous-sound detection in industrial and security contexts. The system explains which spectral features triggered the anomaly flag.

AnomalyReasoningXAI
5

Visual Modality

Jan 2027 — May 2027 · Planned

Visual modality — the most complex perceptual channel: images, video, spatial relations, and object compositions in ternary vector space.

Image encoder (pixels → ternary)Planned

Hierarchical encoding: local patches → regional features → global composition. Each level uses BIND for position and SUPERPOSE for merging.

VisionHierarchicalSpatial
Object detection & scene graphsPlanned

Detection of objects and their spatial relations as structured ternary relations. "Red cup ON the table NEAR the window" — a single algebraic expression.

DetectionScene graphComposition
Video temporal encodingPlanned

Extension to video by binding frames with temporal positional vectors. Actions and motion patterns become trajectories in ternary space.

VideoTemporalMotion
Full cross-modal reasoningPlanned

Text, audio, and visual data in a single vector space. "Find the video where someone says this phrase near this object" — one similarity query.

MultimodalCross-modalReasoning
6

Physical Environment

May 2027 — Sep 2027 · Planned

Moving into implementation in the physical world — robots, smart homes, sensor networks. The same algebraic engine drives perception and decision-making in the real world.

Sensor fusion encoderPlanned

Fusing data from cameras, microphones, lidars, temperature, and motion sensors into unified ternary representations. The HAL normalizes all physical modalities.

SensorsFusionIoT
Robotics perception & planningPlanned

Object recognition, spatial mapping, and task planning on robotic platforms in real time. Every decision is auditable and explainable.

RoboticsPlanningReal-time
Smart home integrationPlanned

Voice commands, presence detection, sound analysis, and visual scene understanding — all on edge hardware. The home understands context, not just triggers.

Smart homeEdgeVoice
Safety & constraint reasoningPlanned

The physical world demands safety guarantees. The reasoning engine evaluates actions against encoded constraints before execution — algebraic verification, not heuristics.

SafetyConstraintsVerification
7

Knowledge Domain Marketplace

May 2027 — Sep 2027 · Planned · Parallel with Phase 6

A marketplace for knowledge domains — a platform where experts publish, share, and monetize ready-made domain encodings, and users assemble custom reasoning systems from off-the-shelf components.

Domain package format & registryPlanned

A standard "knowledge domain" format — an encapsulated set of encoders, reasoning schemas, and reference vectors for a specific subject area. A versioned registry.

PackagingRegistryVersioning
Quality & validation pipelinePlanned

Every published domain goes through automated validation: encoding consistency, reasoning correctness, capacity bounds, and conflict detection.

ValidationCI/CDQuality
Marketplace web platformPlanned

A catalog of knowledge domains with search, ratings, usage statistics, and dependency graphs. Preview a domain's capabilities right in the browser.

Web platformSearchDiscovery
Monetization & licensingPlanned

Flexible pricing models: free open-source domains, one-time purchases, and subscriptions to regularly updated knowledge bases. Transparent analytics for authors.

MonetizationLicensingAnalytics
8

Specialized AI Chip

Oct 2027 — Aug 2028 · Planned

Developing a specialized AI chip — custom silicon that executes ternary vector operations natively. The algebra is the hardware. One tape-out, no second chances.

Instruction set architecture (ISA)Planned

Native machine instructions for BIND, SUPERPOSE, GATE, PERMUTE, and similarity search. Each instruction operates on ternary words in a single cycle.

ISAArchitectureDesign
RTL design & simulationPlanned

RTL implementation in Verilog/SystemVerilog. Full functional simulation confirming correctness against the software reference. Timing analysis.

RTLVerilogSimulation
FPGA prototype & validationPlanned

A complete FPGA prototype before committing to ASIC. Real workloads from all previous phases run on the prototype to validate performance.

FPGAPrototypeValidation
ASIC tape-outPlanned

Final physical design, layout, and tape-out at a leading foundry. Physical verification (DRC/LVS), signal-integrity analysis, and power estimation.

ASICTape-outManufacturing
SDK for chip integrationPlanned

SDK for integrating the MindPrism chip into hardware projects. Drivers, reference boards, documentation, and performance guides.

SDKDriversIntegration

Want to shape the roadmap?

We're looking for research partners, early adopters, and domain experts. Tell us what interpretable AI should do for your field.

Get in touch Read the blog