Papers
Everything is here, including the pilots whose effects died in replication. Status labels tell you how much weight each result can carry.
Working paper · arXiv submission in preparation
Buying the Oracle: Costly Calibration and Intermediated Routing in a Market of Heterogeneous LLM Agents
The flagship result. A 2-by-2 identification separating exploration from granularity, the intermediary as the only information structure that covers its own cost, and the preregistered confirmatory run: 8 of 10 paired seeds, +14.5% gross surplus.
Pilot study · headline effect did not replicate
Public Performance Signals and Task Allocation in Decentralized LLM Agent Markets
Five tool-endowed agents, Vickrey auctions, four information regimes. Textbook adverse selection without signals, a clean-looking information ladder with them. The headline effect did not survive a 10-seed replication with randomized identities; the paper stays online as a documented pilot.
Pilot study · null result under soft verification
Market-Based Orchestration of Specialized AI Agents for Complex Knowledge Work
Embedding-space reputation, real tools, market versus central planning. The market delta was zero under a soft LLM judge, which taught us to treat verifiability as a design constraint. The embedding machinery became the basis of the later vector-market line.