From Switched-Off to Trusted — FNCE 6067 Group Project

Research paper + reproducible code on AI × Blockchain (DeFAI): why finance needs both decentralised AI (so a model cannot be switched off) and explainable, governed AI (so each decision can be justified and monitored).

• report.md — the ~1,850-word paper (≤2000-word limit). Print this for submission.
• figures/ — the five original infographics, all generated by the code.
• src/ — the Python that produces every figure and every number cited in the report.

Reproduce everything

pip install -r requirements.txt
python src/generate_all.py      # writes the 5 PNGs into figures/ and prints key metrics

Every figure is regenerated from scratch; Figures 4–5 use a synthetic, seed-fixed dataset, so the cited numbers (AUC 0.731, applicant P(default) 97%, CSI 0.41, Months-Employed PSI 0.46, kill-switch trigger at month 5) reproduce exactly.

The five figures

File	Layer	What it shows	Course concept
fig1_recall_timeline.png	—	The 12 Jun 2026 model recall and the decentralised-AI token rally	DeFAI lecture
fig2_ai_stack.png	framework	4-layer AI stack with two opposite risks; where each answer applies	AI stack + autonomy ladder
fig3_centralised_vs_decentralised.png	infrastructure	"One switch vs no switch" — the hedge thesis	DeFAI / Bittensor "Bitcoin for AI"
fig4_shap_credit.png	application	SHAP: explaining one declined loan, global + local	Explainability artifacts (SHAP/LIME)
fig5_psi_drift.png	application	PSI/CSI drift dashboard + kill-switch trigger	Model Risk Management lifecycle

Source layout

src/
  theme.py           shared visual style + mandatory source-note footer
  data.py            datasets (DeFAI event facts + synthetic credit DGP + PSI)
  figures_defai.py   fig1, fig2, fig3   (infrastructure-layer story)
  figures_credit.py  model + fig4 (SHAP) + fig5 (PSI)   (application-layer story)
  generate_all.py    single entry point

Provenance / honesty notes

• The DeFAI event magnitudes (TAO +30%, etc.) are reconstructed from the course lecture and the reporting it cites (Grayscale, The Block, CNN/NBC/Fortune); they illustrate the mechanism and are not a price feed.
• The credit dataset is synthetic with a known data-generating process — it exists to demonstrate the SHAP and PSI techniques reproducibly, not to model a real lender.
• All figures are the authors' own work; sources are printed on every figure and listed in report.md.