[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"glossary-non-determinism::en":3,"gloss-cluster-non-determinism::en":23,"gloss-next-non-determinism::en":9},{"slug":4,"category":5,"name":6,"definition":7,"meta_desc":8,"faq":9,"schema_markup":9,"related":10},"non-determinism","core-ai","Non-Determinism","Non-determinism is the property of producing different output from the same input. Language models have it by design and by accident: sampling introduces deliberate variation, and even with sampling turned down, differences in batching, hardware and serving infrastructure mean that identical requests are not guaranteed to return identical text. For anything built on top, this is not a detail — it is the property that invalidates most ordinary engineering habits.\n\nTesting is the first casualty. An assertion that the output equals a fixed string fails intermittently and gets deleted, which is how AI features end up with no tests at all. The workable replacement asserts properties rather than text — that the answer parses, that it names only entities present in the source, that it stays within a length, that it refuses when it should — and runs over a set of inputs rather than one, accepting a rate rather than a single pass.\n\nDebugging changes shape too. A report that cannot be reproduced is not evidence that nothing happened; it is the normal case. That is why the recorded run matters more here than in deterministic systems: without the exact inputs, context and tool results captured at the time, a one-off failure is unexaminable, and the temptation is to declare it a fluke and move on.\n\nAgents amplify all of this, because non-determinism compounds along a chain of decisions. The same request can take a different path, call different tools, cost a different amount and take a different length of time on two consecutive runs, which is why variance is a first-class operational property of an agent rather than noise around an average.\n\nThe practical stance is not to fight it everywhere. Reduce it where it buys nothing — pin what can be pinned, keep prompts under version control, isolate the generative step so the surrounding pipeline stays deterministic — and accept it where the variation is the point. Then measure the distribution instead of the sample, and treat any conclusion drawn from a single run as an anecdote.","Non-determinism means the same input can produce different output. It breaks exact-match tests and reproduction, and it compounds along an agent's chain.",null,[11,14,17,20],{"slug":12,"name":13},"agent-trajectory","Agent Trajectory",{"slug":15,"name":16},"eval-harness","Eval Harness",{"slug":18,"name":19},"hallucination","Hallucination",{"slug":21,"name":22},"regression-testing","Regression Testing",[24,28,32,36,39,42,45,48,51,54,57,60],{"slug":25,"category":5,"name":26,"updated_at":27},"agentic","Agentic AI","2026-09-08T21:32:34+00:00",{"slug":29,"category":5,"name":30,"updated_at":31},"alignment-tax","Alignment Tax","2026-09-08T21:32:35+00:00",{"slug":33,"category":5,"name":34,"updated_at":35},"artificial-intelligence","Artificial Intelligence (AI)","2026-09-08T21:32:36+00:00",{"slug":37,"category":5,"name":38,"updated_at":27},"attention","Attention",{"slug":40,"category":5,"name":41,"updated_at":35},"beam-search","Beam Search",{"slug":43,"category":5,"name":44,"updated_at":31},"benchmark-contamination","Benchmark Contamination",{"slug":46,"category":5,"name":47,"updated_at":31},"catastrophic-forgetting","Catastrophic Forgetting",{"slug":49,"category":5,"name":50,"updated_at":35},"computer-vision","Computer Vision",{"slug":52,"category":5,"name":53,"updated_at":31},"constitutional-ai","Constitutional AI",{"slug":55,"category":5,"name":56,"updated_at":27},"context-window","Context Window",{"slug":58,"category":5,"name":59,"updated_at":35},"deep-learning","Deep Learning",{"slug":61,"category":5,"name":62,"updated_at":27},"diffusion-model","Diffusion Model"]