The labs now pair their interactive models with engineering notes on the same page. Each experiment starts with a small enough system to inspect and follows it far enough to expose a correctness condition or a practical tradeoff.
Consensus under a partition #
The Raft lab keeps separate server logs and terms. Create a minority write, elect a newer leader on the majority, then heal the partition and inspect which entries survive. The notes connect that trace to election restrictions, current-term commitment, linearizable reads and client retries.
Indexes as page operations #
The B+tree workbench inserts and deletes through local splits, redistribution and merges. Persistent page IDs and read/write counts make the storage work visible. Range scans follow linked leaves; a separate byte budget explains why real indexes can have much larger fan-out than the drawing.
Retrieval with an exact baseline #
The vector search lab compares cosine, inner product and Euclidean distance, then measures recall against exact search as the number of probed inverted lists changes. Normalization and approximation are separate interventions, so a ranking change has an identifiable cause.
Learning beyond the training curve #
The decision-boundary trainer compares a linear classifier with a trainable hidden layer. Fixed train/validation data, controlled label noise, L2 regularization and both loss curves help separate representation limits from optimization and generalization.
Decoding as a specified pipeline #
The sampling desk exposes penalties, temperature, truncation order and renormalization. Entropy, retained mass and seeded empirical frequencies show why a short list of parameter values is not a complete description of a sampler.