Jens Fischer

Evals

How to know an AI feature actually works.

Without evals you're tuning prompts by vibe. With evals you can change models, refactor prompts, and ship with confidence. They are the single highest-leverage piece of infrastructure on an AI team.

The minimum viable eval

  1. Collect 20-50 real examples. From production traces, user requests, or hand-written.
  2. Label them. A pass/fail or a 1-5 rubric. Do it yourself the first time.
  3. Run them on every prompt change. A script, a CI step, whatever. The cadence matters more than the tool.
  4. Watch the score over time. When it drops, find out which examples broke.

That's it. You can spend the next year improving on this. You can't skip it.

import anthropic

client = anthropic.Anthropic()

EXAMPLES = [
    {"input": "Refund my order #1234", "expected": "refund"},
    {"input": "Where is my package?",  "expected": "track"},
    # ...
]

def classify(text: str) -> str:
    msg = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=20,
        system="Return one of: refund, track, account, other. Nothing else.",
        messages=[{"role": "user", "content": text}],
    )
    return msg.content[0].text.strip().lower()

passed = sum(classify(ex["input"]) == ex["expected"] for ex in EXAMPLES)
print(f"{passed}/{len(EXAMPLES)} passed")

What to grade

  • Exact match. When the answer is structured (JSON, classification, extraction).
  • LLM-as-judge. When the answer is open-ended. Use a stronger model than the one being evaluated. Calibrate the judge against human labels.
  • Code execution. When the answer is code. Run it.
  • Trajectory. For agents. Did it call the right tools, in roughly the right order, without obvious mistakes?

Things that go wrong

  • Evals drift. Real usage changes. Re-sample from production every few weeks.
  • LLM judges are biased. They favor longer answers, their own family of models, and answers that look confident. Calibrate.
  • You optimize the eval, not the product. Watch out for prompts that game the metric. Spot-check real usage.
  • No one looks at the data. Half the value is reading 20 traces and noticing the same failure three times.

Reading