Notes on getting machine learning into production and keeping it there.
Usage counts and average accuracy are vanity in production. Override rate, correction rate, failed validations and refusal accuracy are what move before an incident.
Read article · 7 minNumeric models drift in ways a chart catches. When the output is code, prose or a derivation, the drift is qualitative and silent. Here is how to surface it.
Read article · 8 minGrading one model with another feels independent. If they share training and assumptions, they share blind spots, and the check passes on the same mistake.
Read article · 7 minLLM-as-judge is the default and often the wrong one. Match the validator to the output: code where you can check it, a human for judgment, a model only for the open-ended.
Read article · 7 minAveraged accuracy is meaningless when a single rare error carries the whole consequence. Set the bar on the tail, and build the eval set to find it.
Read article · 8 minSuccess criteria, a representative task set, tool-use correctness and cost per resolved task — and where to set a go/no-go bar before you ship.
Read article · 8 minGrounding rate, refusal accuracy, recall at k and tool-call accuracy each predict a kind of incident. Overall accuracy predicts none of them.
Read article · 7 minAggregates rise while things break. A diff between two runs is an afternoon of work and it changes what a team argues about.
Read article · 6 minTokens per call and average latency both flatter the system. The number a business can act on is what one completed piece of work costs, at the tail.
Read article · 6 minPolicy pass rate, over refusal and sensitive data in output. A system can score well on every other line of a benchmark and still be the one that leaks.
Read article · 7 minA system that answers differently on the third attempt is a support problem. A suite run once per case can never show it.
Read article · 6 minReal users paraphrase, misspell, pad with irrelevant detail and paste in text carrying instructions. Robustness is the measurement of what that does to your answers.
Read article · 7 minSchema validity, required fields and refusal precision. The measurements that decide whether the system downstream survives contact with your agent.
Read article · 6 minTool selection accuracy is the easy measurement and the least informative. What decides production behaviour is what happens after a call times out.
Read article · 7 minIf the passage holding the answer never reaches the prompt, no model recovers it. Measure retrieval separately or you will spend months tuning the wrong component.
Read article · 7 minHallucination is not a vibe. It is the share of claims no retrieved passage supports, plus what the system does when the answer is not there at all.
Read article · 8 minAn agent retrieves, decides, calls tools and writes. Each part fails on its own, and a single accuracy figure is the average that hides which one did.
Read article · 7 minA contradiction needs both halves in context at once. Retrieval ranks passages by similarity to a query, which is not the same thing, and no model detects a disagreement it has only seen one side of.
Read article · 9 minIf the passage holding the answer never reaches the prompt, no model can recover it. Most disappointing document question answering systems are failing at chunking and search, not at generation.
Read article · 9 minWithout a fixed set of cases and a definition of correct, every change is an opinion and the team argues from anecdotes. The harness is what turns it into a measurement.
Read article · 8 minWe will say plainly whether machine learning is the right tool for it.
Start a conversation