Proof & Testing
Support claims on Runtime.MiRust.com should map to evidence. The source code already treats tests as part of the architecture.
Test categories
| Test lane | What it protects |
|---|---|
| Rust unit tests | Math ops, parser rules, tokenizer, quantization, adapters, packer operators, receipts, manifests. |
| WASM ABI smoke | Export boundary, allocation, model load, generation, next token, adapter validation/apply, reset, free. |
| Browser harness | Static app IDs, local-only markers, route declarations, manifest presence, content types. |
| Browser smoke | Headless browser behavior, model load, generation, Step Token, Adapter Sidecar, selector registries, module-plan drift rejection, assembly evidence checks. |
| Route soaks | Repeated q4_0/q8_0/f32 generated assembly routes, reload behavior, no-store fetches, model release before reassembly. |
| Visual Studio scripts | Build, test, copy WASM, check model artifacts, run browser harness, reuse local server. |
Drift tests
Generated routes should reject checksum-consistent but semantically wrong metadata. The source references drift gates for:
- runtime manifest drift;
- selector byte-budget drift;
- text-sidecar budget drift;
- module-plan profile drift;
- execution profile drift;
- fetch policy drift;
- origin policy drift;
- fetch-count drift;
- resource drift;
- graph drift;
- stack drift;
- role drift;
- slot drift;
- phase drift;
- dependency drift;
- byte-count drift;
- checksum drift;
- assembly eval-case drift;
- assembly evidence-file drift;
- adapter manifest drift;
- adapter artifact drift;
- adapter Rust validation drift.
Quality gates
A runtime smoke model proves runtime behavior. It does not prove assistant usefulness.
Assistant-quality claims need converted-trained provenance and a matching passed evaluation sidecar with declared cases, prompts, expected text, actual text, totals, and checksums.
Public evidence wording
Use phrases such as:
- “source-backed at bundle time”;
- “runtime-smoke evidence”;
- “deterministic fixture”;
- “quality not claimed yet”;
- “accepted for runtime smoke”;
- “requires trained/evaluated model before product-quality claim.”
Avoid implying that deterministic smoke models are production assistants.