VectorIR

01 / The problem

What needs to change

AI agents can generate source code quickly, but ordinary source is difficult to validate semantically before execution and expensive to repair or hand off with precise evidence.

02 / The approach

A focused path through the complexity

VectorIR narrows the problem to deterministic JSON-centric logic and represents behavior as a canonical module that can be validated, interpreted, compiled, patched, replayed, and verified before another agent trusts it.

03 / Key capabilities

What the system is designed to do

• Schema and semantic validation with stable diagnostics
• Fuel-bounded deterministic interpretation and replay
• Readable JavaScript compilation with source annotations
• Capability-scoped host calls and deny-by-default grants
• Compact packets, semantic patches, benchmarks, and Codex skill

04 / Architecture

Technology and system shape

• TypeScript monorepo with specification, validator, runtime, and compiler packages
• Canonical versioned JSON module format
• CBOR transport and digest-enforced handoff packets
• CLI workflow for validation, tests, execution, compilation, and repair
• Proof applications for decision routing and interactive explanation

05 / Results

What the repository demonstrates

• Version 1 includes the complete module format, validator, interpreter, compiler, patching, transport, handoff, benchmarks, examples, and skill workflow.
• The project documents where structured IR improves validation and repair while preserving where TypeScript remains the better choice.

06 / Lessons

What the work teaches

• AI-native formats are most useful when their scope is explicit and their tradeoffs are measured against ordinary code.
• Verified handoff depends on evidence, stable diagnostics, and reproducible execution—not format novelty alone.

07 / Screenshots

Approved visuals