Mistake Clinic

Module-Specific Mistake Radar

Mistake	Signal	Repair
Chunk by token count only	Procedures split apart	Preserve structural boundaries
Vector-only search	Exact identifiers missed	Add lexical and hybrid baselines
End-to-end metric only	Failure stage unknown	Evaluate retrieval and generation separately
Citation presence check	Evidence does not entail claim	Verify claim-span support
Source-only deletion	Embeddings remain	Test every derived store

Practice Mistake Checks

Verify authority, identity, lineage, benchmark quality, hard negatives, freshness, deletion, and fallback.

Required evidence

Submit runnable commands, immutable source and index versions, stage-specific measurements, representative failures, one rejected alternative, and a decision with an explicit stop condition. Screenshots and untraceable generated prose do not pass.

Oral defense

Trace one output backward to its source, explain the first failing stage in one counterexample, justify the baseline and metric, and demonstrate update or deletion behavior.

Source backbone

Use Designing Data-Intensive Applications, official engine documentation, primary retrieval research, and NIST AI RMF selectively.