The Deterministic Memory Problem

The Deterministic Memory Problem

Valori originated from a simple experiment.

A Python-based vector engine was run on:

• macOS ARM (Apple Silicon)
• Linux x86

The same:

• embeddings
• vectors
• code
• model
• seed

And yet — the output diverged.

The divergence did not appear during retrieval or indexing.

It appeared at ingestion.

At the moment embeddings touched the substrate, the memory world-line split.

---

Key Realization

Memory state is not only a data structure —
it is an artifact of the substrate it runs on.

Floating-point behavior varies with:

• instruction pipelines
• fused-multiply-add rules
• rounding modes
• register widths
• reduction order

Two executions that appear “identical” do not evolve the same state.

This breaks:

• deterministic replay
• compliance verification
• safety analysis
• cross-hardware reproducibility
• forensic traceability

In regulated computing environments, that is unacceptable.

Valori was built to constrain the substrate.