Mistake Clinic
Module-Specific Mistake Radar
| Mistake | Signal | Repair |
|---|---|---|
| Weights are the version | Prompt/index mismatch | Version the complete system |
| Mean latency only | Users see long tails | Measure percentiles and queues |
| Cost per token only | Retries and review hidden | Cost successful outcomes |
| Logs without correlation | Regression source unknown | Trace component versions |
| Rollback on paper | Old dependencies unavailable | Rehearse rollback |
Practice Mistake Checks
Verify bounds, idempotency, backpressure, version correlation, privacy, tail latency, unit cost, quality signals, abort criteria, fallback, rollback, and ownership.
Required evidence
Submit reproducible deployment configuration, correlated telemetry, load and failure results, user-impact and unit-cost calculations, one rejected alternative, and the executed response. A dashboard screenshot without raw query and version context does not pass.
Oral defense
Trace one request across boundaries, explain a tail-latency or quality regression, demonstrate containment and rollback, and defend the residual operational risk.
Source backbone
Use Building Secure and Reliable Systems, Google SRE books, OpenTelemetry specifications, and NIST AI RMF.