Overview — Project 6: Model Lab

The previous projects treated models as replaceable capabilities. Model Lab opens the box far enough to make good engineering decisions without pretending you need to train a frontier model from scratch.

You will inspect token-to-logit mechanics, load selected local runtimes, measure memory and throughput, route easy and hard cases, and decide whether a failure needs retrieval, verification, more compute, or tuning.

The five days

Day	Build	Rust pressure	Model lesson
24	transparent causal block	matrices, checked shapes, numeric errors	tokens are IDs; logits are not answers
25	replaceable local runtime	tensors, device errors, native boundaries	model artifact plus runtime is the executable unit
26	calibrated cascade	traits, policies, measurements	small models win narrow verified tasks
27	generate-check-repair/search	bounded concurrency, state	extra compute must buy measured quality
28	data and tuning experiment plan	dataset types, provenance, comparison	tune behavior only after simpler layers fail

Run the opening exhibits:

cd rust/rust-ai-engineering
cargo run -q -p mosaic-models --example inspect_model_bundle
cargo run -q -p mosaic-ml --example causal_mechanics
cargo run -q -p mosaic-ml --example quantization_batch_device

Start with Day 24 — Transformer Mechanics →.