audit/figures.md

Figure Provenance

For each paper figure: which generation script, which input data sources, which experiment.

Figure 1 — fig:multi_market_equity_curves (Q1, main text)

• Source script: scripts/generate_cross_market_case_figures.py → plot_cross_market_equity()
• PDF: figures/case_cross_market_equity_curves.pdf
• Inputs: equity_curve.csv from each of 10 runs in Experiment 1 (US 6M + CN 6M, 5 mandates each)
• What it shows: Daily normalized wealth path per mandate, two stacked panels (CN top, US bottom).
• Last regenerated: commit 579ed3d (cash/exposure merge, 1×3 layout)

Figure 2 — fig:regime_robustness (Q2, main text)

• Source script: scripts/generate_cross_market_case_figures.py → plot_cross_market_shift()
• PDF: figures/case_cross_market_shift.pdf
• Inputs: Aggregated total_return and max_drawdown from each of 10 runs in Experiment 1, via the multi-personality comparison bundles
• What it shows: Two-panel slopegraph (return / MDD) of CN-vs-US shift per mandate.
• Last regenerated: commit 579ed3d

Figure 3 — fig:behavior_profile (Q3, main text)

• Source script: scripts/generate_cross_market_case_figures.py → plot_cross_market_behavior()
• PDF: figures/case_cross_market_behavior_shift.pdf
• Inputs: avg_turnover_ratio, avg_cash_ratio, avg_position_days from each of 10 Experiment 1 runs, via comparison bundles
• What it shows: Three-panel slopegraph (turnover / cash ratio / holding days) of CN-vs-US shift per mandate.
• Last regenerated: commit 579ed3d

Note: gross exposure was an earlier fourth panel; it is the near-complement of cash ratio and was therefore removed to avoid redundancy.

Figure 4 — fig:sector_style_matrix (Q3, main text)

• Source script: scripts/generate_sector_style_matrix_figure.py → render_market(US_PROFILE_RUNS, ...)
• PDF: figures/case_us_sector_style_matrix.pdf
• Inputs: trades.csv from each of 5 US runs in Experiment 1 + universe CSV universe/sector_style_universe.csv
• What it shows: 5-panel heatmap of per-mandate trading concentration over the 5×4 US sector/style grid.
• Critical fix: Earlier this script used Experiment-4 (GPT-5.4 3M) run IDs, which made FV appear to "collapse" to a single Mid-Growth Financial. The current commit a4bf9e6 rewires it to the Experiment-1 (6M case study) runs, where FV in fact spreads across 16 tickers.

Figure 5 — fig:sector_style_matrix_cn (Q3, appendix)

• Source script: scripts/generate_sector_style_matrix_figure.py → render_market(CN_PROFILE_RUNS, ...)
• PDF: figures/case_cn_sector_style_matrix.pdf
• Inputs: trades.csv from each of 5 CN runs in Experiment 1 + universe CSV
• What it shows: Mirror of Figure 4 for the CN A-share universe.

Figure 6 — fig:backend_robustness (Q4, main text)

• Source script: scripts/generate_backend_robustness_figure.py
• PDF: figures/case_backend_robustness.pdf
• Inputs: derived/gpt54_robustness/backend_comparison.csv, which aggregates the 5 GPT-5.4 runs (Experiment 4) and their default-backend counterparts (Experiment 3)
• What it shows: Per-mandate return and average cash ratio, default vs GPT-5.4 backend, on the matched US 3M window.

Figure 7 — fig:paradigm_flow (Section 2)

• Source: Hand-drawn schematic (not generated from data)
• PDF: figures/persona_benchmark_flow.pdf
• What it shows: System flow diagram: fixed components on the left, doctrine module as the manipulated variable in the middle, mandate set and evaluation outputs on the right.