The Economics of Orchestration: Can Culturally-Composed Teams of Small Language Models Outperform Frontier Models at Equal Cost? A Cost-Matched Experiment on Business Solution Generation — Reproducibility Package

Abstract

Reproducibility package for the paper 'The Economics of Orchestration'. Organizations deploying large language models face a budget choice: pay for a frontier model, or orchestrate cheaper models into a multi-step pipeline. Across three pre-registered studies (a 2x2x2 factorial over five LLM families and 19 business problems with 1,900 blinded generation-judgment runs, a refinement probe, and an adaptive designer study with a confirmatory replication wave), the paper shows a three-step team structure lifts a lower-tier model to 98.8% of frontier quality at 88% of the cost, while a single lower-tier call captures 98% at 23% of the cost. This archive contains the final paper PDF plus the complete companion repository: all pipeline definitions, prompts, pinned model versions and prices, raw model outputs, token and latency logs, judge prompts and raw scores, analysis scripts, and pre-registration materials — sufficient to reproduce every reported statistic without access to the DeeWaanAI platform used to execute the runs (accessed August–September 2026).