# Quantization Contract

> How converted numeric representation remains explicit, bounded, and testable.

- Site: TinyRustLM
- Canonical: https://TinyRustLM.mirust.com/quantization/
- Version: 1.8.2
- Updated UTC: 2026-08-30T01:36:19Z

## Representation before speed

Quantized tensors bind signed range, scale placement, grouping, packing order, tail behavior, accumulator width, and deterministic dequantization. Unsupported features are rejected rather than approximated silently.

- Explicit group profile
- Finite scales only
- Checked tails and packing
- Deterministic reconstruction

## Noninferiority gate

Converted scalar, optimized native, and browser-WASM execution use the same prompts, token IDs, positions, seeds, sampling profile, and stop policy. Quantization and kernel differences are localized before behavioral qualification.

- Layerwise comparison
- Preregistered thresholds
- No tolerance tuning after failure
- No lucky-seed selection