A probabilistic model predicts likely next tokens rather than objective facts, so its outputs are non-deterministic and have no built-in model of truth. section 1.3
Model settings and added retrieval shift tradeoffs among accuracy, creativity, safety, cost, and latency; no configuration maximizes all five. Treat generated code as uncertain: fluency and a plausible first result do not guarantee correctness. section 1.3 section 1.4
Repeated ISBN validation prompts
Repeated prompts produced different Python ISBN-13 validators. One implementation rejected hyphenated input; another referenced an undefined variable and crashed. Tests for hyphens, checksums, length, and non-numeric characters expose failures that a plausible-looking answer can hide. section 1.4




