1# Research log — UQO WP 25 (Agentic General Equilibrium)23Decisions, evidence and open tasks, in chronological order. Dates are absolute. Every entry states (a) the question,4(b) assumptions, (c) what was implemented/tested, (d) evidence, (e) threats, (f) next tasks.56## 2026-10-27 — Day 178### 00. Set-up9* Repository created from the UQO template (`_modele-working-paper`), number 25, package `age` (`src/age/…`).10* Local inference: `mlx_lm.server` on M3U96a (Qwen2.5-7B-Instruct-4bit :8081, Llama-3.2-3B-Instruct-4bit :8083) and11 M3U96b (Mistral-7B-Instruct-v0.3-4bit :8082), reached through SSH tunnels (`scripts/llm_serve.sh`, `scripts/llm_tunnel.sh`).12 No paid provider key is used: keys found on the laptop are marked burned by their owner and were left untouched.13* Literature audit and calibration data collected by two parallel team members (see `docs/literature_matrix.md`,14 `docs/references_verification.md`, `docs/parameter_ledger.csv`, `data/raw/README.md`).1516### 01. Core theory (gate 2)17* (a) Question: is there a *distinct* equilibrium mechanism from delegation, beyond "AI in the production function"?18* (b) Assumptions: one-step-ahead information friction; human forecast errors idiosyncratic, model errors common to all19 users of the same foundation model; price-taking; quadratic approximation (Angeletos–Pavan class).20* (c) `scripts/00_lq_derivations.py` (sympy) derives strategies, aggregate law, welfare, private vs social value.21* (d) **Evidence:** the private value of delegation equals the social marginal value *identically* (`wedge == 0`) under the22 CRS restriction `U_xX + U_XX = 0`, `U_Xθ = 0`; confirmed by direct computation of expected utility from the full23 quadratic payoff and by a property-based test over random (α, D, M, τ). With `U_XX = −υ U_xX`, υ ≠ 1 gives a wedge24 of sign (1 − υ). → Proposition 3 (efficiency benchmark) and Proposition 4(a).25* (e) Threats: the benchmark is a second-order approximation; it relies on symmetric ownership and complete markets;26 the literature audit flags Angeletos–Pavan 2007 and Colombo–Femminis–Pavan 2014 as the right comparison — we must27 present the result as "the delegation externality is zero in exactly the cases where their efficiency results apply,28 and non-zero in the frictions we add", not as a new efficiency theorem.29* (f) Next: derive the uninsurable-wage-risk wedge (4b) and the collateral wedge (Prop. 7) explicitly.3031### 02. Mapping the economy to the game32* Savings mapping (household saves in capital, news about TFP): **negative result**. Partial-equilibrium response of33 log savings to a known TFP news shock κ₀ = −0.0015 (γ = 2, ρ_z = 0.9): wealth and substitution effects offset, the34 stake of the consumption–savings decision is negligible, and the GE news response is the standard comovement35 puzzle (investment falls on impact). Kept in the paper as evidence that *which* decision is delegated matters.36* Portfolio mapping (allocation between risky capital claims and a safe store of value, mean-variance): κ₀ = 1/(γσ_R²)37 = 19.5 with σ_R = 0.16 (Shiller 1984–2025 real equity s.d. 0.157), α_s = −(1−α)R^K κ₀ = −1.33 (strong strategic38 substitutability through diminishing returns). With illustrative information parameters (predictable return39 component s.d. 6 %, human noise 10 %, model noise 4 %) the first adopter gains 3.3 % of consumption per year; the40 gain falls to 0.4 % when everyone delegates (the common signal is less valuable once it is already in the aggregate).41 All three s.d. are illustrative and are swept in the robustness section.42* Compute market, short run (fixed K_Q): interior delegation share ≈ 0.47 in the baseline; inference demand is small43 relative to production compute with q̄ = 0.002, so crowding-out is weak → the H6 experiments must use a larger q̄44 (inference-heavy delegation) and report the threshold at which crowding-out flips the labour share. Decision: make45 q̄ a scenario parameter ("inference intensity") rather than a calibrated constant.4647### 03. Calibration notes (from the ledger)48* Labour share level 0.618 (1984–2026, BLS index rescaled by PWT), capital share (Fernald) 0.354, δ = 0.055, TFP AR(1)49 ρ = 0.88–0.93 (linear detrend) with innovation s.d. 0.0126–0.0135; real T-bill 0.57 %, TIPS 10y 0.97 %;50 DFA 2026:Q2 top 1 % 32.5 %, top 10 % 68.9 %, bottom 50 % 2.3 %; equity real return 9.9 % s.d. 15.7 % (1984–2025).51* Baseline choices: β = 0.96 (annual, r ≈ 4 % with growth abstracted), γ = 2, α = 0.33, δ = 0.06, ρ_z = 0.9,52 σ_ε = 0.013. Deviations from ledger means are documented in `paper/sections/calibration.tex`.5354### Open TODO55- [ ] Prop. 4(b) wage-risk wedge, Prop. 7 collateral multiplier (static), proofs in appendix.56- [ ] Model B solver + decomposition; Model C engine + banking + three market designs; LLM lab + audit logs.57- [ ] Pre-registered experiments (`experiments/pre_registration.md`), figures, tables, paper, referee report, git push.5859### 04. Day-1 results and decisions (evening)60* Model A: portfolio mapping is the baseline; PV first adopter 3.3 % of consumption, last 0.4 %; wage-noise s.d. 7.7 %61 (M=1) / 3.2 % (M=8) at full adoption; wedge = 0 to 1.9e-7 on the grid (benchmark); hand-to-mouth Pigouvian tax 4.5× fee62 (M=1), 0.8× (M=8). Compute crowding-out confirmed in sign, tiny in size (labour share +0.2 pp at most) — reported as a63 negative result; inference-intensive delegation self-limits (hump in q̄).64* Model B (n_a=160, 5 GH nodes, a_min=−0.5): no-AI Gini 0.683 / top-10 % 0.457 / r 3.9 %; delegation only → Gini 0.701,65 top-10 % 0.466 (rich delegate first); full broad → Gini 0.575; full concentrated (owner class, capitalised rents) → 0.661,66 top-10 % 0.474. Decision: owner class as a point mass outside the household block (Prop. 6 two-class economy);67 delegation-by-decile computed over households only. Bug fixed: rents were not paid to anyone under concentration.68* Model C (20 seeds): H2 confirmed in all designs (walrasian of_corr 0.11 → 0.97 → 0.43 for D=0 / D=1,M=1 / M=8); volume69 collapse with a shared model (no-trade); leverage defaults highest with private signals. Stress: needed persistent70 dividends (ρ_d=0.95, valuation prices persistence), preloaded leverage 0.95 of capacity, auto-repay off, −10 s.d. shock,71 thin ladder (tick 0.25 %, 40 levels, depth 0.1): call market amplification 2.6× (margin 0.2), 1.5× (margin 0.5), none72 with 4 models; CDA: 10 % trough, no cascade. Disclosed in table notes and referee report.73* LLM lab (quick: 8 agents × 24 periods × 3 seeds, Qwen/Mistral/Llama 4-bit): early cells show Qwen agents under a shared74 signal LESS correlated (0.14–0.32) than the Bayesian control (0.98) — consistent with Hansen–Lee 2025; to be confirmed75 when the run completes and written into results_c (the draft paragraph must be revised accordingly).76* Paper: 69 pages compiled (pdflatex, aer.bst), 144 verified references, 104 generated macros + fallbacks.77* LLM lab provenance note: a helper process (`--policies=llm_diverse --tag=helper`, 19:44–20:12) was started to pre-warm the78 cache and then killed because it contended for the Qwen server; its audit records share the `lab_<model>` run ids (same files),79 its csv outputs were deleted, and `results/llm_lab/cells.csv` comes only from the main run. Llama-3.2-3B's schema failures are80 of one type: `"order_type": "hold"` (valid side, invalid order type) — counted as invalid (no lenient repair), reported as such.81* LLM lab findings (cells bayesian/surrogate/shared/diverse complete, 8 agents × 24 periods × 3 seeds): Qwen agents on one82 shared signal → conditional order-flow corr. 0.20 vs 0.65 (Bayesian) and 1.00 (surrogate) — LLM "interpretation noise"83 restores part of the LLN (Hansen–Lee 2025); but R² of returns on the common error 0.20 vs 0.02 (they trade the common84 signal harder); volume 11 vs 4.4 trades/period (no no-trade); 50–75 % of actions clipped; LLM agents borrow (lev 0.03–0.09);85 Llama-3.2-3B schema failures ("order_type": "hold"). The draft paragraph of results_c was rewritten accordingly; intro and86 abstract updated. The 1-vs-3-model ordering within LLM cells is unresolved at 3 seeds.87* Lab complete 21:47 (72 cells; Qwen 8 180 calls / 3.3 M tokens, Mistral 1 296, Llama; zero cost). Prompt variation: 0.12 vs 0.2088 (single prompt) vs 0.16 (three models) — prompt and model diversity reduce correlation comparably at this scale.8990### Open TODO (end of day 1)91- [ ] SSRN deposit; referee items 2, 7, 9 of docs/referee_report.md; effective-M calibration; Krusell–Smith extension.92