Overview: the question, the three model families, and how the pieces connect

The question

How likely is an AI investment bust over the next three years, and how much damage would it do? Two different events are tracked, because they are not the same thing:

Market crash. Chip stocks (the PHLX Semiconductor index, SOX) fall 40% or more from a peak.
Economic bust. End-customer AI spending falls 15% or more below the path the buildout was planned for (the plan path: revenue keeps pace with the AI capital stock).

The answer comes in three parts: Part I the odds, Part II the damage if it happens, Part III the two combined into expected damage, with an odds tracker.

The three model families

Family	Question	Technique	File
Model G and the pooled odds	How likely is a bust?	Four independent estimates (a revenue simulation, history, market prices, warning indicators) pooled by weighted log-odds	`ai_bust_probability.py` → doc 03
Model D (with Models A, B, C, E)	If a bust happens, how does it spread?	12-sector contagion network with funding, credit and legal-delay channels; Monte Carlo	`ai_bust_models.py` → doc 02
Model F	Same, at firm level, with behaviour and policy	Weekly agent-based simulation with nine switchable mechanisms	`ai_bust_abm.py` → doc 04

How the files connect

flowchart LR
  subgraph PartI[Part I: odds]
    G[Model G: revenue vs plan path]
    H[History: 8 past booms + run-up study]
    M[Market prices: options + credit spreads]
    I[Indicators: credit growth R-zone]
    G --> P[Log-odds pooling]
    H --> P
    M --> P
    I --> P
  end
  P -->|p_bust| S[Reweighted shortfall sampler]
  G -->|shortfall distribution| S
  S --> D[Model D: 12-sector network]
  S --> F[Model F: weekly agent-based model]
  F -->|neocloud failure curve| M
  F -->|40% crash line| G
  D --> E[Expected damage: Part III]
  F --> E
  L[ai_bust_live.py: fetchers, calibration, SQLite, FastAPI] -->|LIVE inputs, Model F overrides| P
  L -->|Model F overrides| F
  E --> W[paper.html, dashboard.html]
  L --> W

Two links are circular by design and are flagged in the paper as sources of non-independence: the credit-spread estimate borrows Model F's neocloud failure curve and Model G's shortfall distribution, and Model G's market-crash line borrows Model F's 10–12.5% line.

Files

File	Role
`ai_bust_models.py`	Models A (revenue gap), B (depreciation), C (neocloud solvency), D (contagion), E (macro) and the 10,000-draw Monte Carlo
`ai_bust_abm.py`	Model F, the weekly agent-based model, its scenario runner, ablations, cliff sweep, knockout tests and Monte Carlo
`ai_bust_probability.py`	Part I (Model G, history, market, indicators, pooling, tracker, sensitivities) and Part III (expected damage); `--calibrate-mapping`
`ai_bust_live.py`, `dashboard.html`, `manual_inputs.json`	The live-data pipeline, API and dashboard
`run_abm_final.py`	Model F ablation, cliff and knockout runs that produce the `abm_final_*.json` files
`build_site.py`	Builds the paper as a web page from `working_paper.md` and the result files
`test_ai_bust_live.py`	Ten offline unit tests for the live pipeline
`probability_results.json`	The paper's published results (as of 2026-10-05)
`abm_final_*.json`, `abm_shock_to_realised.json`	Model F sweep results
`probability_results_rerun_20261006.json`, `backup_20261006/`	The 6 October 2026 reproduction run and the shipped originals it is compared with
`fixtures/`	The recorded snapshot of the paper's inputs
`docs/`	This specification, the paper (`paper.md`, `paper.html`)

Simulation register

Counts are from the result files and code constants. A Model F "scenario run" is a scenario plus its no-shock twin (plus an unconstrained run when the power ceiling is on).

ID	Simulation	Kind	Size	Seeds	Output
S1	Model G	Monte Carlo over uncertain inputs	40,000 paths	20261005	`probability_results.json`
S2	Pooling of the four methods	joint draws with random log-odds weights	40,000	20261005	same
S3	Model G sensitivity	one-at-a-time tornado + 4 stress cases	20,000 paths × 27 settings	7	`tornado_g`
S4	Model G upgrade ladder	cumulative ladder and leave-one-out	40,000 paths × 11 specifications	7	`g_upgrade_ablation`
S5	Model D	sector network, with and without legal friction	10,000 draws × 2	11 / 20260930	`model_d`, `model_d_legal`
S6	Model F, Part III	agent-based, shortfalls from Model G	400 draws × 3 (v2, v1, alternative mapping)	5	`model_f`, `model_f_v1`, `model_f_by_shock`
S7	Model F ablation	each of 9 mechanisms added/removed, 22 cases, shocks 10/20/30%	1,584 scenario runs	1000–1023	`abm_final_ablation.json`
S8	Model F cliff curve	P(neocloud failure) along 17 shock sizes, 3 configurations	1,632 scenario runs	2000–2031	`abm_final_cliff.json`
S9	Model F knockout	wipe out one lender's capital at t=1 in a 15% shock; 8 nodes + control; 2 network designs	432 scenario runs	3000–3023	`abm_final_knockout.json`
S10	Shock-to-outcome mapping	calibrates Part I's shortfall to Model F's demand shock	156 scenario runs	1–12	`abm_shock_to_realised.json`
S11	Live recalibration	smaller version of the chain on current inputs, each refresh	10,000 paths, 4,000 D draws, 40 F draws	20261006	`live_history.db`

The model equations and algorithms are in docs 02–04. build_site.py regenerates the same register, with runtimes, in the paper's Section 18.

Reading order

The paper, docs/paper.md (or paper.html): what is claimed and why.
This overview, then docs 03 → 02 → 04 for the exact mathematics.
docs/05-live-pipeline.md for the dashboard, docs/06-data-and-assumptions.md for where the inputs come from.