Evals / 03
Failure taxonomies
A failure taxonomy is a short, named list of the ways your system goes wrong. Without one, every bug report is a snowflake, every fix is a guess, and "it's dumb sometimes" passes for analysis.
What it is
A stable set of five to eight failure categories for one task, each with a name, a one-line definition, and a canonical example. For a support bot on RAG, a real taxonomy looks like: retrieval miss (the right document was never fetched), grounding failure (the document was fetched and the answer ignored it), fabricated fact, policy violation, refusal when answerable, format break. It lives next to the golden set, and every failing case gets a tag.
Why it exists now
LLM failures feel infinite and vague, so teams argue in anecdotes. Naming converts vibes into counts, and counts change decisions: "62 percent of failures are retrieval misses" means the next sprint is retrieval work, not another round of prompt polish. The taxonomy is also what makes a judge auditable, because a judge that outputs a category tag can be checked, and it is how support, product, and engineering end up describing the same failure with the same word.
The 20-minute kata
- Pull 20 recent failures from logs, tickets, or your golden set's failing cases.
- Write a category name on each. Invent freely on the first pass.
- Merge until five to eight categories remain. Give each a one-line definition and one canonical example.
- Tag all 20 cases with the final categories and count.
- For the biggest bucket, write a one-sentence fix hypothesis. That sentence is your roadmap.
If two buckets keep absorbing the same case, the definitions are too close. Sharpen or merge them.
What good looks like
- Five to eight categories, not thirty. A taxonomy nobody can hold in their head does not get used.
- Categories are distinct enough that two engineers tag the same ten cases the same way. This is checkable: try it.
- Every eval failure and every escalated production case gets a tag at triage, automatically where possible.
- Counts get reviewed on a cadence, and the biggest bucket visibly drives what gets built next.
- The taxonomy is versioned. New categories get added when reality invents new failures, and empty ones get retired.
How it's tested
"Your system gets things wrong sometimes. Walk me through the kinds of wrong, ranked by how often they happen."
The red flag: "hallucination" as the only category. One word carrying six different failure modes means nobody has looked closely enough to fix any of them.