The Failure Library

A public corpus of real agent failures, with forensic analyses attached.

Each case follows one fixed structure, so that failures can be compared instead of retold. Production failures are discussed today as one-off incidents inside individual teams, which makes them impossible to name, compare, or accumulate knowledge about.

The case structure

Every entry answers the same fifteen questions.

01Case
02System / domain
03Trace provenance
04User goal
05Outcome
06Trajectory
07Observed failure
08Failure primitive
09Existing evaluator
10What the evaluator missed
11Acceptance criterion
12Regression case
13Proposed grader
14Intervention
15Counterfactual / generalization question

Cases

Case 001 — forthcoming.

From Traces to Evals

A second line of work: how production trajectories should — and should not — become repeatable evaluations. Which parts of a trace are signal, what a grader can actually measure, and when one incident generalises.

When production failures become evals, how do we prevent the failures we’ve already seen from becoming our definition of correctness?

Contribute traces