Evals cheat sheet: a plain-language guide to knowing whether an AI system actually works

blaizedsouza · x · 2026-11-05

A cheat sheet and companion article by sermakarevich lays out how to evaluate LLM-based systems for engineers, PMs, and CEOs. Key points:

Inputs: start with 20–50 real failure cases; hand-write what "right" means per case; tune on dev set, report only on test set
Graders in tiers: free code rules (exact match, rule checks; similarity metrics are drift alarms, not quality scores), small-model hallucination detectors (HHEM, MiniCheck, LettuceDetect), LLM judges (one yes/no question per failure type with quoted evidence, panel of 3 combined by median, known biases toward first/longer/own-family answers), human labels as ground truth
Check the grader: Cohen's kappa vs humans (0.6 to gate a release, 0.8 for unsupervised), report catch rate and false-fail rate
Honest numbers: error bars on every rate; 60 cases can't detect changes under 12 points; confirm prompt winners on fresh cases
Composite systems: grade RAG retrieval (hit@k) and faithfulness separately; agents on final state, pass^k over runs, cost per success
Public benchmarks: use 5 benchmarks to shortlist 2–3 models; your own eval decides
Habits: code graders every commit, judges + human-read sample on every change, A/B tests after shipping

It ends with 4 numbers to demand: pass rate with error bar, grader-human agreement, detection floor, share of cases that got worse.

More from coding & agent