Maps untrusted model bytes, raw pointers, tokenizer and tensor parsers, source conversion, manifests, browser origin, unsafe memory code, and model-output authority.
Implementation evidence: this topic is grounded in the reviewed GGUF.MiRust.com source snapshot. It documents observed code and artifacts without claiming broad deployment, model quality, or production readiness.
Primary boundaries
- Model and tokenizer bytes are untrusted parser input.
- Raw WASM pointer/length pairs are untrusted host input.
- Raw trained-source directories and manifests are untrusted conversion input.
- Browser origin controls which app, model, and sidecar bytes are served.
- Generated text is data, never execution authority.
Unsafe code
The workspace warns on unsafe code. The WASM allocation and raw-slice boundary explicitly uses unsafe operations. Safety relies on pointer/length validation and correct host deallocation discipline.
Positive controls
Bounds checks, checked arithmetic, explicit error codes, no remote APIs, no arbitrary plugin execution, path traversal checks, and quality non-claims reduce risk.
Missing controls
Cryptographic artifact admission, signed provenance, CSP/COOP/COEP policy, fuzz evidence, memory quotas beyond transfer length, cancellation, worker isolation, capability manifests, license enforcement, and model-output policy are not complete.
Authority rule
A future skill module must state what it can read, emit, and request. The runtime—not model output—must enforce side effects and human approval.
Scope
This starter page defines the questions, boundaries, evidence, and failure modes that should be recorded before a capability is presented as supported.
Engineering considerations
- Identify the source, version, target environment, and owner.
- Separate observed values from estimates and externally reported values.
- Record trade-offs, unsupported cases, and fallback behavior.
- Link performance statements to a compatible benchmark methodology.
Verification questions
- What exact artifact, revision, backend, and environment were reviewed?
- Which assumptions could change the result?
- Which data should be retained so another engineer can reproduce the conclusion?