Explains parameter scale, artifact precision, runtime workspace, context state, cache storage, and why planning factors must be replaced by measurements.
Architecture guide: this topic defines a modular tiny-model planning contract. It does not claim that model artifacts exist, are compatible, or execute on this WordPress site.
Keep measures separate
Parameter count is not artifact size, artifact size is not peak runtime memory, and peak runtime memory is not device compatibility. Precision, tensor layout, tokenizer, working buffers, context, backend, and browser implementation all change the result.
Planning equation
The composer uses visible factors for weights, adapters, workspace, context state, and orchestration. These values support comparison only. Replace each factor with measurements from the exact artifact and implementation before publication.
Sensitivity analysis
- Vary K-active independently from N-total.
- Vary context and output length.
- Compare cold cache, warm cache, CPU, and GPU paths.
- Measure fallback behavior and failed allocations.
- Report median, tail latency, peak memory, and initialization separately.
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?