Eval Drift — When Your Test Set Lies

~12 min · evaluation, drift

Level 0Apprentice

0 XP0/100 lessons0/14 achievements

0/120 XP to next level120 XP to go0% complete

Eval sets get stale

A golden set built in 2025 may not represent 2026 traffic. The categories shifted, the language changed, the user base grew, the policies were updated. A prompt that scores 95% on the old set can be measurably worse on the actual current traffic.

Drift signals

Pass rate on the golden set drifts upward over time without prompt changes (the model got better at answering yesterday's questions).
Production complaint rate rises while eval pass rate stays constant.
New intents appear in production that aren't represented in the golden set.
Reference outputs become stale (policy changes; the "correct" answer is now different).

Maintenance routine

Quarterly: sample 100 production cases, classify them, compare distribution to the eval set.
Add cases from new intents.
Retire cases that are no longer representative.
Re-review reference outputs for currency.

Code

Drift detection sketch·python

from collections import Counter

def intent_distribution(cases):
    return Counter(c["category"] for c in cases)

eval_dist = intent_distribution(load_golden_set())
prod_dist = intent_distribution(sample_production(n=200))

# Symmetric KL or just diff per category
for cat in set(eval_dist) | set(prod_dist):
    e = eval_dist.get(cat, 0) / sum(eval_dist.values())
    p = prod_dist.get(cat, 0) / sum(prod_dist.values())
    if abs(e - p) > 0.05:
        print(f"drift: {cat} eval={e:.2%} prod={p:.2%}")

External links

Anthropic — Maintain test sets

Exercise

Compare your golden set's category distribution against the last 30 days of production traffic. Identify and add three categories that are under-represented in the golden set.