Pre-registration — UQO WP 25 experiments
Registered 2026-10-27 before any factorial experiment was run (theory derivations and the baseline steady state had
been computed; no Model B/C simulation had been executed). Any deviation is recorded in docs/research_log.md.
Primary outcomes (per model)
- Model A (
scripts/10_model_a.py): delegation shareD*, aggregate model-noise variance of capital/wages/returns, private vs social marginal value (wedge), welfare relative to full information, compute rental, automated share and labour share as functions of (τ_A, M, q̄, K_Q, ownership split, hand-to-mouth share). - Model B (
scripts/20_model_b.py): r, w, labour/capital/compute income shares, Gini, top 1 % / top 10 % wealth shares, delegation share by wealth decile, consumption-equivalent welfare by wealth percentile; decomposition by switching on one channel at a time (productivity, displacement, information/decision quality, ownership, borrowing limit). - Model C (
scripts/30_model_c.py): return volatility, bid–ask spread and Kyle-λ (price impact regression), order-flow correlation across agents, default rate, bank losses, price amplification of a common forecast shock, by market design (Walrasian, CDA, bilateral) × signal structure (private, 1 model, M models) × leverage (margin m) × agent policy (Bayesian mean-variance, rule-based, LLM single, LLM diverse, LLM shared-with-varied-prompts).
Hypotheses and falsification conditions
| H | Prediction (within the structural model) | Falsified if |
|---|---|---|
| H1 | D* is interior and increasing in τ_A, decreasing in the compute price and in supervision time; adoption is concentrated among high-stake (wealthy) households in Model B |
D* is monotone in the wrong direction, or adoption in Model B is not increasing in wealth |
| H2 | Order-flow correlation across agents and the price variance attributable to the common shock rise with D and fall with M (≈ 1/M); Kyle-λ under stress is higher with a shared model | correlation/variance do not increase with D or do not fall with M in the Bayesian-agent control; if LLM agents show the opposite, the control decides |
| H3 | Concentrated compute ownership raises the top-1 % wealth share and the Gini relative to broad ownership at the same aggregate output; the effect is larger with scarcity rents | top shares are unchanged or lower under concentration |
| H4 | With margin constraints the price response to a common model shock is amplified (multiplier > 1) and increases with leverage; without leverage the multiplier is 1 | amplification ≤ 1 or non-increasing in leverage |
| H5 | Welfare in Model A is independent of M in the CRS benchmark (Prop. 3) and increasing in M once hand-to-mouth workers or collateral constraints are present; a leverage cap reduces default rates and amplification in Model C | welfare depends on M in the benchmark (would reject Prop. 3's numerics) or leverage caps increase amplification |
| H6 | With fixed compute, raising delegation intensity q̄ raises p_Q, lowers ι and raises the labour share; with elastic compute supply the labour share falls | opposite signs |
Design
- Common random numbers across treatment cells; 20 independent replications per cell for Model C (seeds 1–20), bootstrap 95 % intervals; sensitivity heat-maps for Model A over (τ_A, M) and (q̄, ε_Q).
- Negative controls: zero-AI limit (must reproduce the RBC / Aiyagari benchmarks), M → ∞ (model noise must vanish), τ_A = τ_H (no private gain ⇒ D = 0).
- LLM runs: temperature 0 and 0.7, three local models, prompt-variation control, seed stability; actions validated against JSON schemas and clipped by constraints; every call logged. LLM results are reported next to the Bayesian control and never interpreted as evidence about human preferences.
- Out-of-distribution shock: a 3-s.d. common forecast shock and a 50 % margin increase in Model C.
What we will not claim
External causal identification; realism of LLM agents as humans; that Model C outcomes are general equilibria; that the illustrative information parameters are estimates.