Limitations (living document; mirrors paper/sections/limitations.tex)
- Efficiency benchmark is second-order and price-taking. Proposition 3 holds in the LQ approximation with complete markets and symmetric ownership; market power of model providers, non-pecuniary spillovers and higher-order terms break it (Prop. 4a gives the direction).
- Which decision is delegated is assumed. Savings mapping has no stake (kappa0 = -0.0015); the portfolio mapping is the baseline. Firm-level input decisions (David–Hopenhayn–Venkateswaran channel) are not modelled.
- Information parameters are illustrative (sigma_theta_R, sig_H_R, sig_A_R, M, kappa, tau_s, qbar; Model B return noise; Model C dividend/noise/market-maker parameters). Levels of welfare effects are orders of magnitude, not estimates.
- Model B has no aggregate risk; it isolates cross-sectional channels; top-1% share too low (standard Aiyagari limitation).
- Model C is a laboratory: temporary trading equilibrium per period, market maker ladder as liquidity, 8–20 agents. The margin-spiral channel (H4) could not be reproduced in the lab: transient model errors mean-revert once dividends fail to confirm them, and the mechanical margin rule deleverages agents at the first breach; see results_c.tex for the numbers.
- LLM agents are not humans: three small 4-bit open-weight models, prompt/temperature dependence, logged; results are about the architecture, not about deployed commercial agents.
- No external causal identification; validation roadmap in calibration.tex.
- Independent-AI-principal counterfactual implemented but not reported (no institutional benchmark).
- Engineering caveats: the Walrasian call only clears among submitted limits (no price when one side is empty); CDA prices are last-trade prices (stale when no trade); the owner class in Model B is outside the household block.