Engineering Requirements

If AI is to become efficient, trustworthy, and suitable for high‑integrity use, it must be built on rigorous, verifiable principles. These principles define the foundations upon which a coherent AI architecture can be constructed.

Unified Memory Semantics

• a single addressing model spanning hardware, runtime, and language
• hardware‑enforced aliasing and borrowing rules
• elimination of undefined behaviour

Memory must be coherent across the entire stack. Undefined behaviour cannot be tolerated in systems that require determinism and traceability.

Deterministic Execution

• optional fully deterministic mode for training and verification
• explicit concurrency primitives with no hidden reordering
• reproducible behaviour by design

Determinism is not a luxury. It is a prerequisite for certification, debugging, and scientific reproducibility.

First‑Class AI Types

• native support for tensors, graphs, sequences, and distributions
• type‑checked operations including slicing and broadcasting
• built‑in provenance and lineage metadata

AI development should not rely on ad‑hoc conventions layered on top of generic programming languages. The types of AI must be first‑class citizens.

Compositional Model Interfaces

• strongly typed
• fully inspectable
• composably modular
• serialisable with complete semantic fidelity

Models must be treated as structured, typed artefacts—not opaque blobs.

Structured Error & Uncertainty Handling

• explicit confidence thresholds
• out‑of‑distribution detection
• defined recovery and fallback paths
• comprehensive logging with full data lineage

AI systems must acknowledge uncertainty explicitly rather than burying it in heuristics or silent failures.

Energy‑Aware Execution

• hardware and runtime expose energy cost per primitive
• compiler and scheduler optimise for total system efficiency

Energy is a first‑class constraint. Efficiency must be engineered, not hoped for.