Skip to content

bijux-vex — read this first

bijux-vex is a vector execution engine. It MUST run vector computations under explicit execution contracts, record the consequences, and make determinism vs. non-determinism auditable.

What vector execution means

  • Inputs are execution requests (text or vectors) routed through explicit plans.
  • Backends are execution substrates, not storage layers.
  • Artifacts are replayable records of how execution happened, not generic indexes.

Deterministic vs non-deterministic contracts

  • Deterministic: bit-stable, replay MUST match exactly, no hidden randomness MAY exist.
  • Non-deterministic (experimental): randomness MUST be declared, bounded, and audited; replay MUST check distributional consistency within declared bounds.

Why ANN is treated as execution (not indexing)

  • ANN choices (graph walks, sampling, pruning) MUST be treated as execution-time decisions.
  • Backends MUST refuse deterministic contracts when they cannot honor exactness.
  • Approximation metadata (MUST include randomness sources and reproducibility bounds) is part of the execution artifact.

Replay as a first-class invariant

  • Replay MUST mean “re-run the same execution plan under the same contract.”
  • Deterministic replay ⇒ equality; non-deterministic replay ⇒ declared envelope.
  • Provenance chains MUST include contract, randomness profile, and artifact signature.

Typing philosophy

  • Runtime contracts and invariants are primary; static typing is advisory.
  • Type checkers MUST NOT be treated as soundness gates. Invariants, provenance, and conformance tests enforce correctness.
  • ExecutionIntent is represented canonically as intent strings (exact_validation, reproducible_research, exploratory_search, production_retrieval); the enum is a convenience veneer only.

All other spec documents link back here as the canonical mental model.