Prompt Component Cost Dataset for LLM Deployment

It contains a main full-factorial experiment with 256 configurations (215,040 inferences) and four engineering validation experiments: a code-level simulation (Plan C), a physics simulation (EmbodiedJev), cross-domain validation on the public ALFRED benchmark, and a real FastAPI service deployment.

Each inference records input/output/total tokens, latency, output length metrics, and success rates where applicable. The dataset includes raw per-inference records, data dictionaries, a quality report, and experiment code for reproducibility. All dependent variables are token, latency, and output length; quality scoring is not involved.