ReferenceStrong architectural inferencev1.21.5
In plain English
This page is reference material: definitions, schemas, catalogs, templates, and implementation records.
- Why this matters: AI risk can come from the whole arrangement, not one obvious model.
- What to look for: data, memory, routes, adapters, tools, evaluators, updates, and rollback paths.
- Technical version below: the expert terminology remains available and is linked through the glossary.
Browser LLM Review Checklist
Evidence levelStrong architectural inferenceTechnical label: Architectural inference
Local inferenceA conclusion or output produced from data. Open glossary definition improves several operational properties, but it also creates browser-specific review work. This checklist converts the zero-dependency Rust report into a non-operational governance checklist.
Artifact identity
- Record the
.wasmSHA-256 hash. - Record compiler target, optimization mode, LTO state, panic strategy, SIMD/atomics flags, and post-processing tools.
- Record model container hash, tokenizer hash, architecture ID, quantization family, and adapterA small add-on that changes or specializes model behavior. Open glossary definition compatibility rules.
- Record whether any third-party JavaScript, CDN, dynamic import, or external model loader is used.
Runtime determinism
- Use seeded sampling for reproducible evaluation where feasible.
- Capture deterministic evaluation prompts, expected metrics, tokenizer version, and sampling configuration.
- Log peak scratch arena usage, token throughput, memory growth events, accepted speculative tokens, rejected speculative branches, and assembly checksum.
Memory and cache reset
- Clear KV cache pages.
- Clear radix-prefix cache entries.
- Clear local prompt history, synthetic examples, evaluation scratch, service-worker caches, IndexedDB records, and user-granted file handles.
- Verify that reset removes adapters and route policies as well as model weights.
Composition evaluation
- Test base model alone.
- Test base plus each adapter.
- Test the exact adapter stackA set of adapters loaded together, usually in a defined order. Open glossary definition and load order.
- Test tokenizer edge cases.
- Test speculative decoding rollbackReturning a system to an earlier known state. Open glossary definition.
- Test worker-count changes.
- Test the reset ecology action.
Source anchors
- Rust WebAssembly code-size reference
- WebAssembly threads on web.dev
- MDN SIMD dot_i16x8_s reference
- CycloneDX ML-BOM
Review output
The output is not a safety certificate. It is an evidence packet: artifact hashes, model/adapters, tokenizer, runtime settings, evaluation report, cache/reset proof, and known limitations.