docs: sync scripts/render_academic_pdfs.py with Round-7 audit

scripts/render_academic_pdfs.py

@@ -335,7 +335,7 @@ for b_key, b_label in [
 
 # Table P1-5: Capacity Baselines vs Conformer 1.19B across Horizons & Years (Strict Non-Zero ROC-AUC)
 TBL_BASELINES = [
-    ["Baseline Architecture", "Params", "1m NZ", "3m NZ", "5m NZ", "15m NZ", "30m NZ", "60m NZ", "1d NZ", "'24 Mean", "'25 Mean", "'26 Mean"]
+    ["Baseline Architecture", "Params", "1m NZ ('24)", "3m NZ ('24)", "5m NZ ('24)", "15m NZ ('24)", "30m NZ ('24)", "60m NZ ('24)", "1d NZ ('24)", "'24 Mean", "'25 Mean", "'26 Mean"]
 ]
 # Feature & small-model baselines
 cap = AUDIT.get("capacity_and_feature_baselines", {})
@@ -368,7 +368,7 @@ TBL_BASELINES.append([
     "0.5000", "0.5000", "0.5000"
 ])
 # 2. Linear-8F (Raw 8 Candlestick Features)
-TBL_BASELINES.append(format_baseline_row("Linear Logistic / Ridge (8 Features)", "63 (9/h)", l8))
+TBL_BASELINES.append(format_baseline_row("Ridge Regression (8 Features)", "63 (9/h)", l8))
 # 3. Linear-56F (8 Features x 7 Lags)
 TBL_BASELINES.append(format_baseline_row("Linear Logistic (56 Lagged Features)", "399 (57/h)", l56))
 # 4. HistGBM-56F (Gradient Boosted Trees)
@@ -445,18 +445,43 @@ TBL_FROB_EXACT.append([
     f"{g24['global_rms_update_magnitude']:.3e}"
 ])
 shuf = AUDIT.get("shuffled_label_and_gradient_audit", {})
-s_fp32 = shuf.get("shuffled_label_960steps_fp32_master", {})
-s_cos = shuf.get("cosine_true_wf24_vs_shuffled_960", 0.019184)
+s_mod = shuf.get("per_module_ep3_and_shuffled_control", {})
+s_glob = s_mod.get("Global_Model_283_Tensors", {})
+s_tok = s_mod.get("Inception_Tokenizer", {})
+s_bb = s_mod.get("Conformer_Backbone", {})
+s_cos = s_glob.get("cosine_true_wf24_vs_shuffled_960", 0.019184)
 TBL_FROB_EXACT.append([
-    "Shuffled-Label 960-Step Control",
+    "Shuffled-Label Control (Global)",
     "1,189.14M",
     "—",
-    f"{s_fp32.get('global_rel_frobenius_pct', 0.5401):.4f}%",
+    f"{s_glob.get('shuffled_rel_frobenius_pct', 1.802):.4f}%",
     "—",
     "—",
     f"{s_cos:.6f} (vs True)",
-    "0.04838",
-    f"{s_fp32.get('rms_update_magnitude', 2.592e-4):.3e}"
+    f"{s_glob.get('base_w_rms', 0.01438):.5f}",
+    f"{s_glob.get('shuffled_update_rms_mean', 2.592e-4):.3e}"
+])
+TBL_FROB_EXACT.append([
+    "  └─ Shuffled Tokenizer Control",
+    "4.87M",
+    "—",
+    f"{s_tok.get('shuffled_rel_frobenius_pct', 1.178):.4f}%",
+    "—",
+    "—",
+    "97.9% of Noise Ceiling",
+    f"{s_tok.get('base_w_rms', 0.02574):.5f}",
+    f"{s_tok.get('shuffled_update_rms_mean', 3.032e-4):.3e}"
+])
+TBL_FROB_EXACT.append([
+    "  └─ Shuffled Backbone Control",
+    "1,180.06M",
+    "—",
+    f"{s_bb.get('shuffled_rel_frobenius_pct', 1.807):.4f}%",
+    "—",
+    "—",
+    "83.5% vs 24.4% True",
+    f"{s_bb.get('base_w_rms', 0.01432):.5f}",
+    f"{s_bb.get('shuffled_update_rms_mean', 2.588e-4):.3e}"
 ])
 
 # Table P2-2: Pre-2024 Stress Forgetting Audit (2020-03 COVID Crash, 2022-06 Rate Shock, 2023-12 Anchor)
@@ -551,16 +576,32 @@ for k_base, lbl in [
         f"{ph['1m']['abs_logit_rms_shift']:.4f}/{ph['15m']['abs_logit_rms_shift']:.4f}/{ph['1d']['abs_logit_rms_shift']:.4f}"
     ])
 
-# Table P3-2: Non-Overlapping Contiguous Backtest — Frozen 2024 Policy vs Benchmarks & Demeaned Rules
+# Table P3-2: Non-Overlapping Contiguous Backtest — Synchronous 258-Step Grid
 TBL_BACKTEST = [
-    ["Policy & Execution Spec", "OOS Yr", "Trades / Steps", "Turn / Trd", "Gross/Trd", "BE State (Iso)", "SR 0bp (DSR)", "SR 1.5bp (Iso)", "SR 3.0bp", "CumRet 1.5bp", "MaxDD 1.5bp"]
+    ["Policy & Execution Spec", "OOS Yr", "Trades / Steps", "Turn", "Gross μ ± σ (bp)", "BE State (Iso)", "SR 0bp (DSR)", "SR 1.5bp (Iso)", "SR 3.0bp", "CumRet 1.5bp", "MaxDD 1.5bp"]
 ]
-# Benchmark 1: Buy and Hold
-for yr, sr, ret in [("2024", "+0.42", "+14.2%"), ("2025", "+0.38", "+11.8%"), ("2026", "-0.31", "-4.1%")]:
-    TBL_BACKTEST.append(["Equal-Weight Buy & Hold", yr, "15 / 10,304", "0.000", "+0.14bp", "N/A", f"{sr} (N/A, M=1)", f"{sr} ({sr})", f"{sr}", ret, "18.4%"])
-# Benchmark 2: Market Neutral
-for yr, sr, ret in [("2024", "+0.21", "+3.1%"), ("2025", "+0.14", "+1.9%"), ("2026", "+0.08", "+0.4%")]:
-    TBL_BACKTEST.append(["Market-Neutral (Beta=0)", yr, "15 / 10,304", "0.450", "+0.08bp", "0.18bp", f"{sr} (0.120)", f"{float(sr)-0.2:+.2f}", "-0.45", ret, "6.2%"])
+# Benchmark 1: Buy and Hold (synchronous 258-step grid)
+bh_sync = AUDIT["benchmarks_1bar_delay_and_dsr"]["benchmarks_synchronous_258_steps"]["Equal_Weight_Buy_and_Hold"]
+for yr in ["2024", "2025", "2026"]:
+    v = bh_sync[yr]
+    TBL_BACKTEST.append([
+        "Equal-Weight Buy & Hold", yr, f"{v['n_trades']} / {v['n_steps']}",
+        f"{v['turnover_per_trade']:.3f}", f"{v['gross_mean_ret_bps']:+.2f} ± {v['per_step_vol_bps']:.1f}",
+        v["breakeven_one_way_bps"], f"{v['annualized_sharpe_0bp']:+.2f} (N/A, M=1)",
+        f"{v['annualized_sharpe_1.5bp']:+.2f} ({v['annualized_sharpe_1.5bp']:+.2f})",
+        f"{v['annualized_sharpe_3.0bp']:+.2f}", f"{v['cum_return_1.5bp_pct']:+.2f}%", f"{v['max_drawdown_1.5bp_pct']:.2f}%"
+    ])
+# Benchmark 2: Market Neutral (synchronous 258-step grid)
+mn_sync = AUDIT["benchmarks_1bar_delay_and_dsr"]["benchmarks_synchronous_258_steps"]["Market_Neutral_Beta0"]
+for yr in ["2024", "2025", "2026"]:
+    v = mn_sync[yr]
+    TBL_BACKTEST.append([
+        "Market-Neutral (Beta=0)", yr, f"{v['n_trades']} / {v['n_steps']}",
+        f"{v['turnover_per_trade']:.3f}", f"{v['gross_mean_ret_bps']:+.2f} ± {v['per_step_vol_bps']:.1f}",
+        f"{v['breakeven_one_way_bps']:+.2f}bp", f"{v['annualized_sharpe_0bp']:+.2f} (N/A, M=1)",
+        f"{v['annualized_sharpe_1.5bp']:+.2f} ({v['annualized_sharpe_1.5bp']:+.2f})", f"{v['annualized_sharpe_3.0bp']:+.2f}",
+        f"{v['cum_return_1.5bp_pct']:+.2f}%", f"{v['max_drawdown_1.5bp_pct']:.2f}%"
+    ])
 # 1-Bar Delayed Execution Policy (t+1 entry, y - y_1m)
 b1 = AUDIT["benchmarks_1bar_delay_and_dsr"]["best_2024_1bar_delay_policy"]
 for yr, k in [("2024", "2024_1bar_delay"), ("2025", "2025_untouched_1bar_delay"), ("2026", "2026_untouched_1bar_delay")]:
@@ -570,7 +611,7 @@ for yr, k in [("2024", "2024_1bar_delay"), ("2025", "2025_untouched_1bar_delay")
         tag, yr,
         f"{yv['n_trades']} / {yv['n_steps']}",
         f"{yv['turnover_per_trade']:.3f}",
-        f"{yv['gross_mean_ret_bps']:+.2f}bp",
+        f"{yv['gross_mean_ret_bps']:+.2f} ± {yv['per_step_vol_bps']:.1f}",
         f"{yv['breakeven_one_way_bps']:+.2f}bp",
         f"{yv['annualized_sharpe_0bp']:+.2f} ({yv['dsr_0bp']:.3f})",
         f"{yv['annualized_sharpe_1.5bp']:+.2f} ({yv['annualized_sharpe_1.5bp_isolated']:+.2f})",
@@ -602,34 +643,22 @@ for h in ["3m", "5m"]:
 # COMPILE PAPER 1: Scaling 1.189B Financial Conformers & 1m Base-Rate Analysis
 # =============================================================================
 p1_abstract = (
-    "Applying billion-parameter neural sequence modeling to high-frequency financial market microstructure requires confronting "
-    "severe non-stationarity, heavy-tailed innovations, and discrete tick-grid artefacts such as zero-return flat bars. In this "
-    "paper, we introduce and empirically audit the <b>Financial Conformer Foundation Model</b>, a <b>1,189,135,740-parameter "
-    "(1.189B)</b> hybrid convolution-attention architecture pre-trained across <b>14 calendar years</b> (January 2010 – December "
-    "2023; 168 monthly Parquet shards, 15,030 steps) of 1-minute US equity bars across 15 liquid instruments. Our architecture "
-    "combines an 8-feature stationary candlestick transform, a 6-branch <b>Multi-Scale Inception 1D Convolutional Tokenizer</b> "
-    "(K in {3, 7, 15, 31, 63, 127} minutes, stride s=4) compressing L=2048-minute windows into T=512 tokens, <b>20 Deep Conformer "
-    "Blocks</b> (d_model=2048) with <b>Grouped Query Attention (GQA)</b>, <b>Rotary Position Embeddings (RoPE)</b>, depthwise "
-    "separable convolutions (K_dw=7), and <b>SwiGLU</b> feed-forward layers (d_ff=5504), and dual regression/classification heads "
-    "across 7 horizons (1m, 3m, 5m, 15m, 30m, 60m, 1d=390m). Across 855 logged pre-training telemetry checkpoints, batch loss "
-    "standard deviation contracts by <b>-15.2%</b> across epochs (0.0079 to 0.0071 to 0.0067). Evaluating on a unified, stratified "
-    "15-ticker contiguous cohort across 2024 (8,044 windows), 2025 (8,060 windows), and 2026 (8,100 windows), we perform a rigorous "
-    "<b>Base-Rate and Flat-Bar Decomposition</b> of the I(y_{t,h} &gt; 0) directional objective. We prove that because 6.46%–7.86% "
-    "of 1-minute bars are exact flat bars (y=0), strictly positive returns occur on only 46.15%–46.51% of 1m bars, shifting the "
-    "mean 1m logit to -0.1627 so that uncalibrated z &gt; 0 thresholding collapses to predicting down/flat (53.49%–53.85% raw "
-    "accuracy, matching a FlatBar-Down heuristic). Crucially, removing this static threshold artefact via <b>Strict Non-Zero "
-    "ROC-AUC (y != 0)</b>, <b>Causal Prior-Calibrated Accuracy</b>, and <b>|y|-Quartile Stratification</b> characterizes "
-    "threshold-independent ranking: Epoch 3 achieves 1m All-Bar ROC-AUC of <b>0.5375 / 0.5262 / 0.5274</b> (inflated by flat-bar detection), "
-    "Strict Non-Zero ROC-AUC of <b>0.5177 / 0.5161 / 0.5198</b>, Prior-Calibrated Accuracy of <b>52.04% / 51.23% / 51.25%</b>, and uncalibrated "
-    "two-sided 5-minute Non-Zero Accuracy of <b>51.74%</b> (2024), <b>49.90%</b> (2025), and <b>51.18%</b> (2026). "
-    "Crucially, capacity controls across all seven horizons reveal that a 16-parameter linear ridge model achieves <b>0.5289 1m NZAUC</b> in 2024, "
-    "outperforming the 1.189B Conformer (0.5177) and establishing a definitive null result on scale in intraday return prediction."
+    "We investigate whether scaling deep neural architectures to <b>1,189,135,740 parameters</b> confers predictive advantages in financial market microstructure. "
+    "Our model combines an 8-feature stationary candlestick transform, a 6-branch Multi-Scale Inception 1D Convolutional Tokenizer, 20 Deep Conformer Blocks "
+    "(d_model=2048, GQA, RoPE, SwiGLU), and dual regression/classification heads across 7 horizons (1m–1d). Evaluating across 24,204 stratified contiguous windows "
+    "(2024–2026), we perform an empirical <b>Base-Rate and Flat-Bar Decomposition</b> of the directional objective. Because 6.46%–7.86% of 1-minute bars are "
+    "exact flat bars (y=0), uncalibrated thresholding collapses to static base-rate capture (53.49%–53.85% raw accuracy). On strict non-zero returns (y != 0), "
+    "Epoch 3 achieves small but positive 1m ranking (<b>Strict Non-Zero AUC = 0.5177 / 0.5161 / 0.5198</b>, +0.018 above chance). "
+    "However, capacity benchmarks establish that a <b>63-parameter linear ridge model</b> achieves <b>0.5289 1m NZAUC</b> in 2024, outperforming the 1.189B Conformer (0.5177; a +0.0112 advantage slightly exceeding the empirical noise floor). "
+    "Across 3m to 1d horizons, performance differences between the 1.189B Conformer and ridge remain under 0.0065, sitting within the measured empirical noise floor "
+    "(±0.010 AUC across 5 seeds). Furthermore, a 50M Conformer control dips below chance at 60m (0.4925) and 1d (0.4853), establishing that scaling to 1.189B parameters "
+    "provides no capacity advantage in intraday return prediction."
 )
 
 p1_elements = [
     ("h1", "1. Introduction & Microstructure Capacity Challenges"),
     ("p", "Applying neural architectures beyond 10^9 parameters to financial time series presents unique statistical traps: (1) <b>Non-Stationary Drift</b> across multi-year macro cycles [11, 12]; (2) <b>Discrete Tick-Grid Quantization and Zero-Return Flat Bars</b>, where 6.5%–8.6% of 1-minute bars exhibit zero price change (y_{t,1m}=0), biasing asymmetric binary labels d_{t,h} = I(y_{t,h} &gt; 0); and (3) <b>Severe Capacity Saturation</b>, where low signal-to-noise ratios prevent deep overparameterized models from outperforming shallow feature baselines."),
-    ("p", "In this work, we introduce and audit a <b>1,189,135,740-parameter Financial Conformer Foundation Model</b> pre-trained on 14 years (2010–2023) of 1-minute OHLCV data across 15 liquid US equities and ETFs (AAPL, AMD, AMZN, BAC, GOOGL, INTC, IWM, JPM, META, MSFT, NFLX, NVDA, QQQ, SPY, TSLA). Beyond reporting raw metrics, we provide a complete mathematical decomposition separating static base-rate/flat-bar effects from true threshold-independent ranking alpha, and benchmark against linear and gradient-boosted capacity controls across all 7 forecast horizons."),
+    ("p", "In this work, we introduce and audit a <b>1,189,135,740-parameter Financial Conformer Foundation Model</b> pre-trained on 14 years (2010–2023) of 1-minute OHLCV data across 15 liquid US equities and ETFs (AAPL, AMD, AMZN, BAC, GOOGL, INTC, IWM, JPM, META, MSFT, NFLX, NVDA, QQQ, SPY, TSLA). Beyond reporting raw metrics, we provide an empirical decomposition separating static base-rate/flat-bar effects from true ranking alpha, and benchmark against linear and gradient-boosted capacity controls across all 7 forecast horizons."),
     ("h1", "2. Exact Feature Engineering & Architecture Specification"),
     ("p", "Matching our production implementation in <i>features.py</i> and <i>config_1b.py</i> verbatim, each 1-minute bar (O_t, H_t, L_t, C_t, V_t) is transformed into an F=8-dimensional stationary feature vector x_t in R^8 clipped to [-10.0, +10.0], consisting of log return, close-to-open gap, high-low log spread, signed candle body, upper wick, lower wick, 20-bar Parkinson volatility, and 20-bar log-volume z-score:"),
     ("eq", "x_t = clip([ ln(C_t/C_{t-1}), ln(O_t/C_{t-1}), ln(H_t/L_t), ln(C_t/O_t), ln(H_t/max(O_t,C_t)), ln(min(O_t,C_t)/L_t), sigma_Park(20), Z_vol(20) ], -10, +10)", "1"),
@@ -647,13 +676,13 @@ p1_elements = [
     ("p", "A critical methodological insight concerns the binary label d_{t,h} = I(y_{t,h} &gt; 0). At the 1-minute horizon, exact flat bars (y_{t,1m} = 0) account for f_flat = 7.86% (2024), 6.85% (2025), and 6.46% (2026) of all observations, and input bars with ln(H_t/L_t)=0 occur on 13.33%–16.07% of windows (after which P(y_{t,1m} &lt;= 0 | ln(H_t/L_t)=0) = 57.09%–60.63%). Because d_{t,1m} assigns flat bars to class 0, the strictly positive base rate P(y_{t,1m} &gt; 0) is only 46.51% (2024), 46.36% (2025), and 46.15% (2026). Under BCE training, the directional head centers its 1m logit distribution around the negative base-rate log-odds (mean z_{1m} = -0.1627, std = 0.1232). Consequently, uncalibrated thresholding at z_{t,1m} &gt; 0 yields a 0.0% positive prediction rate (predicting 0 on all bars) and achieves a raw accuracy of 1 - P(y&gt;0) = 53.49%, 53.64%, and 53.85%. Symmetrically, at 60m and 1d, positive equity drift shifts mean logits to +0.0603 and +0.1275 (std 0.0135 and 0.0160), so uncalibrated z &gt; 0 predicts +1 on 100.0% of bars."),
     ("eq", "z_cal_{t,h} = z_{t,h} - (1/W) * sum_{j=1}^W z_{t-j,h};   Acc_cal(h) = P( I(z_cal_{t,h} > 0) == I(y_{t,h} > 0) ),   NZAUC(h) = AUC(z_{t,h}, y_{t,h} | y_{t,h} != 0)", "5"),
     ("p", "To eliminate this static threshold bias completely, we evaluate: (1) <b>Strict Non-Zero Accuracy (y != 0)</b>; (2) <b>Causal Prior-Calibrated Accuracy</b> (thresholding z_{t,h} at its trailing 256-step causal mean); (3) <b>All-Bar and Strict Non-Zero ROC-AUC (NZAUC)</b>; (4) <b>Non-Overlapping Horizon Rank IC</b> stepped by S_h = ceil(tau_h / 32) windows (giving N_eff = 627–630 non-overlapping daily observations per year across the 15 tickers); and (5) <b>1-Minute |y|-Quartile Stratification</b> (Table 4 and Figure 2)."),
-    ("fig", ("papers/figures/fig2_oos_directional_accuracy.png", "Figure 2: (a) All-Bar vs. Strict Non-Zero (y != 0) ROC-AUC across all 7 horizons. (b) 1-Minute Prior-Calibrated Accuracy and ROC-AUC across strictly non-zero |return| quartiles (Q1..Q4), proving genuine discrimination beyond flat bars.")),
+    ("fig", ("papers/figures/fig2_oos_directional_accuracy.png", "Figure 2: (a) All-Bar vs. Strict Non-Zero (y != 0) ROC-AUC across all 7 horizons. (b) 1-Minute Prior-Calibrated Accuracy and ROC-AUC across strictly non-zero |return| quartiles (Q1..Q4), showing empirical discrimination marginally above chance (+0.018 AUC).")),
     ("tbl", (TBL_OOS_UNIFIED, "Table 3: Unified 15-ticker stratified OOS evaluation of Frozen Epoch 3 across 2024 (8,044 win), 2025 (8,060 win), and 2026 (8,100 win).", [52, 24, 28, 32, 34, 34, 38, 38, 38, 38, 38, 44])),
     ("tbl", (TBL_1M_BUCKETS, "Table 4: 1-Minute return-magnitude |y| quartile decomposition (Flat y=0 vs. strictly non-zero Q1–Q4 quartiles; note Q1 AUC matches overall 1m AUC as signal concentrates in smallest moves).", [88, 56, 56, 38, 56, 38, 56, 56, 38])),
-    ("tbl", (TBL_BASELINES, "Table 5: Capacity baselines vs. Conformer 1.19B across all 7 horizons and multi-year out-of-sample cohorts (Strict Non-Zero ROC-AUC).", [96, 36, 28, 28, 28, 28, 28, 28, 28, 38, 38, 38])),
-    ("p", "Tables 3–5 establish three rigorous findings: (1) While uncalibrated 1m raw accuracy (53.49%–53.85%) reflects the base rate captured by the <i>FlatBar_Down_Heuristic</i> (which has 0.5000 AUC and 0.5000 NZAUC), Epoch 3 achieves genuine threshold-independent 1m ranking: <b>All-Bar AUC = 0.5375 / 0.5262 / 0.5274</b> and <b>Strict Non-Zero NZAUC = 0.5177 / 0.5161 / 0.5198</b>, with Prior-Calibrated Accuracy of <b>52.04% / 51.23% / 51.25%</b>; (2) Across non-zero 1m return quartiles (Table 4), calibrated accuracy remains above 50% across Q1–Q3 in 2024 (51.43%–53.53% CalAcc, 0.5273–0.5375 AUC) and peaks on the largest Q4 moves (mean 13.83 bps) in 2026 (<b>51.37% CalAcc, 51.80% RawAcc, 0.5271 AUC</b>); and (3) At 3m and 5m—where uncalibrated logits actively fire on both sides (27.0%–39.4% positive rate)—Epoch 3 achieves <b>51.74% (2024), 49.90% (2025), and 51.18% (2026) Strict Non-Zero Accuracy at 5m</b> (AUC 0.5256, 0.5016, 0.5134). Crucially, benchmarking against shallow linear models (Table 5) reveals that Ridge on 8 features achieves <b>0.5289 NZAUC at 1m</b> in 2024, outperforming the 1.189B Conformer (0.5177), demonstrating that 1.19B parameters provides no capacity benefit over shallow models on low-SNR financial returns."),
+    ("tbl", (TBL_BASELINES, "Table 5: Capacity baselines vs. Conformer 1.19B across 7 horizons on the 2024 out-of-sample cohort (Strict Non-Zero ROC-AUC) alongside multi-year non-zero means (2024–2026).", [96, 36, 28, 28, 28, 28, 28, 28, 28, 38, 38, 38])),
+    ("p", "Tables 3–5 establish three empirical findings: (1) While uncalibrated 1m raw accuracy (53.49%–53.85%) reflects the base rate captured by the <i>FlatBar_Down_Heuristic</i> (which has 0.5000 AUC and 0.5000 NZAUC), Epoch 3 achieves small but positive 1m ranking: <b>All-Bar AUC = 0.5375 / 0.5262 / 0.5274</b> and <b>Strict Non-Zero NZAUC = 0.5177 / 0.5161 / 0.5198</b> (+0.018 above chance), with Prior-Calibrated Accuracy of <b>52.04% / 51.23% / 51.25%</b>; (2) Across non-zero 1m return quartiles (Table 4), calibrated accuracy remains above 50% across Q1–Q3 in 2024 (51.43%–53.53% CalAcc, 0.5273–0.5375 AUC) and peaks on the largest Q4 moves (mean 13.83 bps) in 2026 (<b>51.37% CalAcc, 51.80% RawAcc, 0.5271 AUC</b>); and (3) At 3m and 5m—where uncalibrated logits actively fire on both sides (27.0%–39.4% positive rate)—Epoch 3 achieves <b>51.74% (2024), 49.90% (2025), and 51.18% (2026) Strict Non-Zero Accuracy at 5m</b> (AUC 0.5256, 0.5016, 0.5134). Benchmarking against shallow models (Table 5) shows that while Ridge on 8 features achieves <b>0.5289 NZAUC at 1m</b> in 2024 (outperforming Conformer's 0.5177; a +0.0112 advantage slightly exceeding the ±0.010 spread), performance differences across all other horizons (3m to 1d) remain under 0.0065. Multi-seed measurement (5 seeds) confirms that the empirical noise floor is ±0.010 AUC; furthermore, the 50M Conformer control dips below chance at 60m (0.4925) and 1d (0.4853), establishing that differences below 0.010 sit within the noise floor and indicating no capacity advantage from scaling to 1.189B parameters."),
     ("h1", "5. Conclusion"),
-    ("p", "By separating static base-rate threshold collapse from threshold-independent ranking and benchmarking against shallow feature-only models across all seven horizons, we establish a definitive empirical null result on scale: while the 1.189B Conformer captures short-horizon bid-ask microstructure, its predictive capacity provides zero advantage over a 16-parameter linear ridge model on intraday financial returns.")
+    ("p", "By separating static base-rate threshold collapse from threshold-independent ranking and benchmarking against capacity controls across all seven horizons, we document an empirical null result on scale: while the 1.189B Conformer captures short-horizon bid-ask microstructure, its predictive capacity provides no demonstrated advantage over a 63-parameter (9 per horizon) linear ridge model on low-SNR intraday equity returns.")
 ]
 
 compile_paper(
@@ -663,29 +692,19 @@ compile_paper(
     keywords="Financial Foundation Models, Conformer Architecture, Market Microstructure, Flat-Bar Base-Rate Calibration, Non-Zero ROC-AUC.",
     abstract_text=p1_abstract,
     elements=p1_elements,
-    references=p1_refs if 'p1_refs' in globals() else [
+    references=[
         "Vaswani, A., et al. (2017). Attention is all you need. NeurIPS, 30, 5998–6008.",
-        "Brown, T., et al. (2020). Language models are few-shot learners. NeurIPS, 33, 1877–1901.",
-        "Touvron, H., et al. (2023). LLaMA: Open and efficient foundation language models. arXiv:2302.13971.",
-        "Dosovitskiy, A., et al. (2021). An image is worth 16x16 words: Transformers for image recognition at scale. ICLR.",
-        "Zhou, H., et al. (2021). Informer: Beyond efficient transformer for long sequence time-series forecasting. AAAI.",
-        "Nie, Y., et al. (2023). A time series is worth 64 words: Long-term forecasting with transformers (PatchTST). ICLR.",
-        "Das, A., et al. (2024). A decoder-only foundation model for time-series forecasting (TimesFM). ICML.",
-        "Ansari, A. F., et al. (2024). Chronos: Learning the language of time series. TMLR.",
-        "Kaplan, J., et al. (2020). Scaling laws for neural language models. arXiv:2001.08361.",
-        "Hoffmann, J., et al. (2022). Training compute-optimal large language models (Chinchilla). NeurIPS, 35.",
-        "Kim, T., et al. (2022). Reversible instance normalization for accurate time-series forecasting (RevIN). ICLR.",
+        "Gulati, A., et al. (2020). Conformer: Convolution-augmented transformer for speech recognition. Interspeech, 5036–5040.",
+        "Su, J., et al. (2024). RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568, 127063.",
+        "Zhang, B., & Sennrich, R. (2019). Root mean square layer normalization. NeurIPS, 32.",
+        "Shazeer, N. (2020). GLU variants improve transformer. arXiv:2002.05202.",
         "Cont, R. (2001). Empirical properties of asset returns: stylized facts and statistical issues. Quantitative Finance, 1(2), 223–236.",
         "Sirignano, J., & Cont, R. (2019). Universal features of price formation in financial markets. Quantitative Finance, 19(9), 1449–1459.",
         "Zhang, Z., Zohren, S., & Roberts, S. (2019). DeepLOB: Deep convolutional neural networks for limit order books. IEEE TSP, 67(11), 3001–3012.",
         "Bouchaud, J. P., et al. (2018). Trades, quotes and prices: financial markets under the microscope. Cambridge University Press.",
         "Gu, S., Kelly, B., & Xiu, D. (2020). Empirical asset pricing via machine learning. The Review of Financial Studies, 33(5), 2223–2273.",
         "Kelly, B., Malamud, S., & Zhou, K. (2024). The virtue of complexity in return prediction. The Journal of Finance, 79(1), 459–503.",
-        "Gulati, A., et al. (2020). Conformer: Convolution-augmented transformer for speech recognition. Interspeech, 5036–5040.",
-        "Ainslie, J., et al. (2023). GQA: Training generalized multi-query transformer models from multi-head checkpoints. EMNLP.",
-        "Su, J., et al. (2024). RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568, 127063.",
-        "Zhang, B., & Sennrich, R. (2019). Root mean square layer normalization. NeurIPS, 32.",
-        "Shazeer, N. (2020). GLU variants improve transformer. arXiv:2002.05202."
+        "Kim, T., et al. (2022). Reversible instance normalization for accurate time-series forecasting (RevIN). ICLR."
     ]
 )
 
@@ -694,34 +713,19 @@ compile_paper(
 # COMPILE PAPER 2: Continual Walk-Forward Adaptation, Exact Cosine Geometry & Pre-2024 Stress Forgetting
 # =============================================================================
 p2_abstract = (
-    "Deploying billion-parameter financial foundation models under non-stationary regime drift requires balancing forward plasticity "
-    "against backward forgetting of historical crisis manifolds. In this paper, we audit an <b>Online Walk-Forward Continual "
-    "Adaptation Protocol</b> for our <b>1,189,135,740-parameter Financial Conformer</b> across 2024 (WF-2024) and 2025 (WF-2025), "
-    "resolving three foundational mathematical and empirical questions. First, we replace unweighted per-tensor cosine averages "
-    "(which overweight 7-element bias vectors relative to 2048x5504 weight matrices) with the <b>Exact Concatenated Global "
-    "Inner-Product Cosine Similarity</b> derived via the Hilbert-space Law of Cosines in float64, proving that every module and "
-    "checkpoint pair strictly satisfies the Cauchy-Schwarz lower bound cos &gt;= sqrt(1 - δ_F^2): global Frobenius drift is "
-    "<b>0.5439%</b> (cos_exact = <b>0.99998521</b> &gt;= 0.99998521) for Ep3->WF24, <b>0.5725%</b> (cos_exact = <b>0.99998361</b>) "
-    "for WF24->WF25, and <b>0.7860%</b> (cos_exact = <b>0.99996911</b>) cumulatively for Ep3->WF25. Second, we test whether the "
-    "observed plasticity hierarchy is merely an Adam step-size denominator artefact (small ||W||_RMS inflating relative drift) by "
-    "decomposing relative drift into per-parameter RMS weight norm ||W||_RMS and absolute RMS update ||ΔW||_RMS. We prove that the "
-    "Inception Tokenizer actually possesses the <b>largest</b> weight norm (||W||_RMS = 0.02574 vs. 0.01432 for the backbone and "
-    "0.01106 for the return head), yet still exhibits the highest relative drift (1.1773%) because its absolute RMS update "
-    "(<b>3.030e-4</b>) is <b>4.00x larger</b> than the backbone (7.575e-5) and <b>11.3x larger</b> than the return head (2.676e-5). "
-    "Similarly, GQA Attention updates <b>3.32x more</b> in absolute RMS terms than SwiGLU FFN (1.452e-4 vs. 4.374e-5) despite near-identical "
-    "weight norms (0.01241 vs. 0.01209). Third, we directly test backward catastrophic forgetting on <b>Pre-2024 Historical Stress "
-    "Shards</b> (March 2020 COVID-19 crash, June 2022 Fed rate shock, December 2023 anchor; 7,896 windows) and compute <b>Paired "
-    "Block-Bootstrap 95% Confidence Intervals</b> (B=500) on strictly forward 2026 OOS data (8,100 windows). On Pre-2024 stress "
-    "regimes, WF-2024 (0.5139 AUC, 50.97% CalAcc) and WF-2025 (0.5135 AUC, 51.10% CalAcc) show zero overall degradation vs. Frozen "
-    "Epoch 3 (0.5130 AUC, 50.90% CalAcc; ΔAUC = +0.0005), while on strictly forward 2026 OOS data, testing on strict non-zero "
-    "AUC yields 1m ΔNZAUC = <b>+0.0013 [-0.0074, +0.0100]</b> (raw p = 0.77, Holm p = 1.0 across 35 hypothesis tests), which does "
-    "not achieve statistical significance under multiplicity control. Running a 960-step control on shuffled labels (AdamW, lr=1e-5, seed 42) demonstrates "
-    "that Tokenizer RMS update displacement (3.03e-4) sits directly on the theoretical Adam noise ceiling (3.10e-4) with an update cosine "
-    "of +0.019184, providing no evidence of structured continual drift beyond optimizer Brownian motion."
+    "We audit an <b>Online Walk-Forward Continual Adaptation Protocol</b> for the 1.189-billion parameter Financial Conformer across 2024 and 2025. "
+    "Using the exact float64 Hilbert-space Law of Cosines, we show that global weights shift by <b>0.5439%</b> in 2024 (cos = 0.99998521) and <b>0.5725%</b> in 2025 "
+    "(cos = 0.99998361). Decomposing updates into weight norm ||W||_RMS and update displacement ||ΔW||_RMS reveals a sharp structural divergence: input tokenizer "
+    "updates sit at <b>97.9% of the theoretical Adam Brownian ceiling</b> (3.030e-4 vs. 3.098e-4), matching a shuffled-label noise control (3.032e-4, 97.9%). "
+    "In contrast, deep Conformer backbone updates under real labels are <b>3.4x smaller</b> than under shuffled noise (7.575e-5 vs. 2.588e-4, 24.4% vs. 83.5% of ceiling), "
+    "while global drift is suppressed threefold (0.5439% vs. 1.8020%), consistent with weak but non-random gradient signals suppressed by Adam second-moment damping. "
+    "The cross-update cosine (+0.0192 ± 0.0018) is small but ~10 SE from zero, rejecting exact orthogonality while confirming that walk-forward trajectory alignment "
+    "with true data remains marginal. Evaluating on Pre-2024 historical crisis manifolds (7,896 windows) and forward 2026 data (8,100 windows) with paired block bootstraps "
+    "across 35 hypothesis tests confirms that walk-forward deltas span zero under Holm-Bonferroni correction (Holm p >= 0.50), confirming no statistically significant forward adaptation edge."
 )
 
 p2_elements = [
-    ("h1", "1. Introduction: Testing Stability and Plasticity Rigorously"),
+    ("h1", "1. Introduction: Testing Stability and Plasticity Systematically"),
     ("p", "When a billion-parameter model pre-trained on 2010–2023 market microstructure is updated online across 2024 and 2025, three methodological pitfalls must be avoided: (1) <b>In-Sample Evaluation Confusion:</b> Evaluating WF-2024 on 2024 or WF-2025 on 2025 measures in-sample adaptation rather than forward out-of-sample transfer; true backward forgetting requires testing on <b>pre-2024 historical stress periods</b> (such as the March 2020 COVID-19 liquidity crisis and June 2022 Federal Reserve tightening shock), while true forward transfer requires evaluating on strictly held-out <b>2026</b> data with paired bootstrap confidence intervals; (2) <b>Tensor-Averaged Cosine Distortion:</b> Computing an unweighted mean of per-tensor cosines across 283 tensors weights 7-parameter bias vectors equally with 11.27M-parameter SwiGLU matrices, violating the global norm inequality cos &gt;= sqrt(1 - δ_F^2); and (3) <b>Adam Step-Size Confounding:</b> Because AdamW normalizes updates by second moments, one must verify whether higher relative drift ||ΔW||_F / ||W||_F in boundary layers reflects genuinely larger absolute parameter updates ||ΔW||_RMS or merely smaller denominator norms ||W||_RMS."),
     ("h1", "2. Exact Global Cosine Geometry & Adam Plasticity Arithmetic"),
     ("p", "For any module M_m containing P_m scalar parameters across tensor index set J_m, let ||W_m^(A)||_F and ||W_m^(B)||_F denote the exact Frobenius norms of checkpoints A and B, and let ||ΔW_m||_F = ||W_m^(B) - W_m^(A)||_F. Defining the relative Frobenius drift δ_F = ||ΔW_m||_F / ||W_m^(A)||_F and norm ratio r_m = ||W_m^(B)||_F / ||W_m^(A)||_F, the exact global concatenated inner-product cosine similarity in R^{P_m} is given by the Hilbert-space Law of Cosines:"),
@@ -729,14 +733,14 @@ p2_elements = [
     ("p", "To separate absolute update magnitude from denominator norm scaling, we decompose δ_F into the ratio of the per-parameter root-mean-square update ||ΔW_m||_RMS and the per-parameter RMS weight norm ||W_m^(A)||_RMS:"),
     ("eq", "||Delta W_m||_RMS = sqrt( ||Delta W_m||_F^2 / P_m ),   ||W_m^(A)||_RMS = sqrt( ||W_m^(A)||_F^2 / P_m )   ==>   delta_F(M_m) = ||Delta W_m||_RMS / ||W_m^(A)||_RMS", "2"),
     ("fig", ("papers/figures/fig3_walk_forward_comparison.png", "Figure 3: (a) Decomposition of relative Frobenius drift into absolute per-parameter RMS update ||ΔW||_RMS vs. base RMS weight norm ||W||_RMS. (b) 7-Horizon mean ROC-AUC across Pre-2024 historical stress cohorts (2020-03 COVID, 2022-06 Rates, 2023-12) and 2024–2026 OOS cohorts.")),
-    ("tbl", (TBL_FROB_EXACT, "Table 1: Exact float64 Law-of-Cosines global cosine similarity, Cauchy-Schwarz lower bound, per-parameter RMS weight/update decomposition, and shuffled-label control.", [95, 42, 46, 46, 48, 46, 102, 42, 48])),
-    ("p", "Table 1 proves four exact mathematical results: (1) <b>Strict Cauchy-Schwarz Satisfaction:</b> Every single subsystem and global checkpoint pair strictly satisfies cos_exact &gt;= sqrt(1 - δ_F^2), and the larger cumulative drift Ep3->WF25 (0.7860%) corresponds to a strictly lower exact cosine (0.99996911) than Ep3->WF24 (0.5439%, cos_exact = 0.99998521); (2) <b>Genuine Gradient Concentration vs. Adam Arithmetic:</b> The Inception Tokenizer has the <b>largest</b> denominator weight norm in the entire network (||W||_RMS = 0.02574, 1.80x larger than the backbone's 0.01432), yet still exhibits the largest relative drift (1.1773%) because its absolute RMS parameter update (<b>3.030e-4</b>) is <b>4.00x larger</b> than the backbone (7.575e-5). Furthermore, within the 20 Conformer blocks, GQA Attention (||W||_RMS = 0.01241) and SwiGLU FFN (||W||_RMS = 0.01209) have nearly identical weight norms, yet GQA Attention updates <b>3.32x more</b> in absolute RMS terms (1.452e-4 vs. 4.374e-5); (3) <b>Update Vector Orthogonality in R^{1.19x10^9}:</b> The exact cosine between ΔW_{Ep1->Ep2} and ΔW_{Ep2->Ep3} is <b>+0.180473 (79.60°)</b>, reflecting coherent pre-training descent, whereas between annual walk-forward updates ΔW_{Ep3->WF24} and ΔW_{WF24->WF25} it is <b>-0.009270 (90.53°)</b>, consistent with nearly independent regime adjustments in 10^9 dimensions; and (4) <b>Shuffled-Label Random Walk Proof:</b> Optimizing 960 steps on shuffled labels (AdamW, lr=1e-5, batch 32, seed 42) produces a Tokenizer RMS displacement of <b>3.031e-4</b>, sitting directly on the theoretical Adam noise ceiling (η*sqrt(K) = 1e-5 * sqrt(960) = 3.098e-4). The cross-update cosine between true-label and shuffled-label drift is <b>+0.019184</b>, providing no evidence of structured continual drift beyond optimizer Brownian motion."),
+    ("tbl", (TBL_FROB_EXACT, "Table 1: Exact float64 Law-of-Cosines global cosine similarity, Cauchy-Schwarz lower bound, per-parameter RMS weight/update decomposition, and like-for-like shuffled-label control starting from Ep3.", [95, 42, 46, 46, 48, 46, 102, 42, 48])),
+    ("p", "Table 1 establishes four quantitative empirical results: (1) <b>Strict Cauchy-Schwarz Satisfaction:</b> Every single subsystem and global checkpoint pair strictly satisfies cos_exact &gt;= sqrt(1 - δ_F^2), and the larger cumulative drift Ep3->WF25 (0.7860%) corresponds to a strictly lower exact cosine (0.99996911) than Ep3->WF24 (0.5439%, cos_exact = 0.99998521); (2) <b>Gradient Concentration vs. Adam Arithmetic:</b> The Inception Tokenizer has the <b>largest</b> denominator weight norm in the entire network (||W||_RMS = 0.02574, 1.80x larger than the backbone's 0.01432), yet still exhibits the largest relative drift (1.1773%) because its absolute RMS parameter update (<b>3.030e-4</b>) is <b>4.00x larger</b> than the backbone (7.575e-5). Furthermore, within the 20 Conformer blocks, GQA Attention (||W||_RMS = 0.01241) and SwiGLU FFN (||W||_RMS = 0.01209) have nearly identical weight norms, yet GQA Attention updates <b>3.32x more</b> in absolute RMS terms (1.452e-4 vs. 4.374e-5); (3) <b>Update Vector Trajectory in R^{1.19x10^9}:</b> The exact cosine between ΔW_{Ep1->Ep2} and ΔW_{Ep2->Ep3} is <b>+0.180473 (79.60°)</b>, reflecting coherent pre-training descent, whereas between annual walk-forward updates ΔW_{Ep3->WF24} and ΔW_{WF24->WF25} it is <b>-0.009270 (90.53°)</b>, consistent with nearly independent regime adjustments in 10^9 dimensions; and (4) <b>Shuffled-Label Like-for-Like Control:</b> Optimizing 960 steps on shuffled labels starting from exact Ep3 master weights (||W||_RMS = 0.014384, across 3 seeds) confirms that Tokenizer displacement (3.032e-4) sits at 97.9% of the theoretical Adam noise ceiling (η*sqrt(K) = 1e-5 * sqrt(960) = 3.098e-4) under both true and shuffled labels (noise-like boundary diffusion). In contrast, in the 20 Conformer backbone blocks, true gradient updates (7.575e-5, 24.4% of ceiling) are 3.4x smaller than shuffled-label noise diffusion (2.588e-4, 83.5% of ceiling), while global weight drift is suppressed threefold (0.5439% vs. 1.8020%), demonstrating that true labels produce weak but non-random gradient signals that Adam second-moment damping and weight decay suppress. The cross-update cosine between true-label and shuffled-label drift is +0.0192 ± 0.0018, small but ~10 SE from zero, rejecting exact orthogonality while confirming that walk-forward trajectory alignment with true data remains marginal."),
     ("h1", "3. Pre-2024 Stress Forgetting Audit & 2026 Paired Bootstrap CIs"),
     ("tbl", (TBL_FORGETTING, "Table 2: Direct Catastrophic Forgetting Audit on Pre-2024 Historical Stress Regimes (2020-03 COVID, 2022-06 Rates, 2023-12) and Forward 2024–2026 OOS.", [98, 36, 62, 62, 62, 58, 58, 40, 40])),
-    ("tbl", (TBL_BOOTSTRAP, "Table 3: Paired Block-Bootstrap 95% Confidence Intervals (B=500, block=64) on Strict Non-Zero AUC with Holm Multiplicity Control.", [60, 88, 88, 44, 44, 94, 94])),
-    ("p", "Tables 2 and 3 resolve the catastrophic forgetting and forward-transfer questions with statistical precision: (1) <b>Pre-2024 Crisis Retention:</b> Evaluated on 7,896 windows from March 2020 (COVID-19 crash), June 2022 (Fed rate shock), and December 2023, WF-2024 (0.5139 AUC, 50.97% CalAcc) and WF-2025 (0.5135 AUC, 51.10% CalAcc) preserve or slightly exceed Frozen Epoch 3 (0.5130 AUC, 50.90% CalAcc). On the March 2020 COVID-19 crash, 1-minute non-zero discrimination shows <b>ΔNZAUC = +0.0044 (95% CI [-0.0112, +0.0200], raw p = 0.58)</b> and 1-day <b>ΔNZAUC = -0.0164 [-0.2643, +0.2315]</b>, both statistically indistinguishable from zero (Holm p = 1.000); and (2) <b>Multiplicity-Controlled 2026 Forward Transfer:</b> On strict non-zero AUC (Table 3), the 2026 forward adaptation step (Epoch 3 -> WF-2025) is <b>ΔNZAUC = +0.0013 (95% CI [-0.0074, +0.0100], raw p = 0.77; 0.5198 -> 0.5211)</b>. Across all 35 hypothesis tests spanning 5 comparisons and 7 horizons, zero cells achieve statistical significance under Holm-Bonferroni or Benjamini-Hochberg FDR control at α = 0.05 (1m Holm p = 1.0, BH p = 0.77), confirming that online adaptation does not produce statistically defensible alpha beyond noise."),
+    ("tbl", (TBL_BOOTSTRAP, "Table 3: Paired Block-Bootstrap 95% Confidence Intervals (B=2500, block=64) on Strict Non-Zero AUC with Holm Multiplicity Control.", [60, 88, 88, 44, 44, 94, 94])),
+    ("p", "Tables 2 and 3 resolve the catastrophic forgetting and forward-transfer questions with statistical precision: (1) <b>Pre-2024 Crisis Retention:</b> Evaluated on 7,896 windows from March 2020 (COVID-19 crash), June 2022 (Fed rate shock), and December 2023, WF-2024 (0.5139 AUC, 50.97% CalAcc) and WF-2025 (0.5135 AUC, 51.10% CalAcc) preserve or slightly exceed Frozen Epoch 3 (0.5130 AUC, 50.90% CalAcc). On the March 2020 COVID-19 crash, 1-minute non-zero discrimination shows <b>ΔNZAUC = +0.0044 (95% CI [+0.0009, +0.0079], raw p = 0.0145; a single post-hoc cell out of 35 that yields Holm p = 0.5075, confirming no statistically significant departure from zero after multiplicity correction)</b> and 1-day <b>ΔNZAUC = -0.0164 [-0.0348, +0.0020]</b> (raw p = 0.0811, Holm p = 1.0000); and (2) <b>Multiplicity-Controlled 2026 Forward Transfer:</b> On strict non-zero AUC (Table 3), the 2026 forward adaptation step (Epoch 3 -> WF-2025) is <b>ΔNZAUC = +0.0014 (95% CI [-0.0006, +0.0034], raw p = 0.1615; 0.5198 -> 0.5211)</b>. Across all 35 hypothesis tests spanning 5 comparisons and 7 horizons, zero cells achieve statistical significance under Holm-Bonferroni or Benjamini-Hochberg FDR control at α = 0.05 (1m Holm p = 1.0000, BH p = 0.65), confirming that online adaptation does not produce statistically defensible alpha beyond noise."),
     ("h1", "4. Conclusion"),
-    ("p", "By combining exact float64 Law-of-Cosines geometry, per-parameter RMS update decomposition, seed-verified shuffled-label noise controls, pre-2024 crisis stress testing, and multiplicity-corrected paired bootstrap intervals on 2026, we demonstrate that online fine-tuning displacement is noise-driven optimizer diffusion that fails to produce statistically significant out-of-sample gains.")
+    ("p", "By combining exact float64 Law-of-Cosines geometry, per-parameter RMS update decomposition, seed-verified shuffled-label noise controls, pre-2024 crisis stress testing, and multiplicity-corrected paired bootstrap intervals on 2026, we demonstrate that online fine-tuning displacement is heavily dominated by optimizer diffusion that fails to produce statistically significant out-of-sample gains.")
 ]
 
 compile_paper(
@@ -753,13 +757,11 @@ compile_paper(
         "Lu, J., et al. (2018). Learning under concept drift: A review. IEEE TKDE, 31(12), 2346–2363.",
         "Kim, T., et al. (2022). Reversible instance normalization for accurate time-series forecasting (RevIN). ICLR.",
         "López de Prado, M. (2018). Advances in financial machine learning. John Wiley & Sons.",
-        "McCloskey, M., & Cohen, N. J. (1989). Catastrophic interference in connectionist networks. Psych. Learning & Motiv., 24, 109–165.",
         "Kirkpatrick, J., et al. (2017). Overcoming catastrophic forgetting in neural networks (EWC). PNAS, 114(13), 3521–3526.",
-        "Zenke, F., Poole, B., & Ganguli, S. (2017). Continual learning through synaptic intelligence. ICML, 3987–3995.",
-        "Hu, E. J., et al. (2022). LoRA: Low-rank adaptation of large language models. ICLR.",
         "Lee, Y., et al. (2023). Surgical fine-tuning improves adaptation to distribution shifts. ICLR.",
         "Kunsch, H. R. (1989). The jackknife and the bootstrap for general stationary observations. The Annals of Statistics, 17(3), 1217–1241.",
         "Politis, D. N., & Romano, J. P. (1994). The stationary bootstrap. Journal of the American Statistical Association, 89(428), 1303–1313.",
+        "Kingma, D. P., & Ba, J. (2015). Adam: A method for stochastic optimization. ICLR.",
         "Loshchilov, I., & Hutter, F. (2019). Decoupled weight decay regularization (AdamW). ICLR."
     ]
 )
@@ -769,45 +771,45 @@ compile_paper(
 # COMPILE PAPER 3: Controlled Mechanistic Ablation & Non-Overlapping Frozen-Policy Backtest
 # =============================================================================
 p3_abstract = (
-    "Evaluating mechanistic interpretability and economic executability in billion-parameter financial models requires rigorous "
+    "Evaluating mechanistic interpretability and economic executability in billion-parameter financial models requires systematic "
     "experimental controls against off-distribution clamping artefacts, denominator norm inflation, overlapping multi-period returns, "
     "and ex-post configuration selection bias. In this paper, we conduct a peer-review-hardened mechanistic and execution audit of "
     "our <b>1,189,135,740-parameter Financial Conformer Foundation Model</b> across 24,204 stratified contiguous windows (2024–2026). "
     "First, we upgrade our 7-horizon Inception tokenizer ablation (K in {3, 7, 15, 31, 63, 127}) with <b>In-Distribution "
     "Mean-Ablation</b> (clamping each branch to its empirical post-SiLU channel mean rather than zero), <b>Random 341-Channel and "
     "682-Channel Controls</b>, <b>Variance-Normalized Centered Logit Perturbation</b> (||Δ(z - z_bar)||_2 / ||z - z_bar||_2), and "
-    "<b>Denominator-Free RMS Logit Shift</b> (RMS(Δz)). We prove that the apparent 15m peak in uncentered relative L2 perturbation "
+    "<b>Denominator-Free RMS Logit Shift</b> (RMS(Δz)). We show that the apparent 15m peak in uncentered relative L2 perturbation "
     "(||Δz||_2 / ||z||_2 = 33.61% at 15m vs. 4.49% at 1d for K=7) is a <b>denominator effect</b> caused by the baseline mean logit "
     "crossing zero at 15m (z_bar_{15m} = +0.0118, ||z||_RMS = 0.0405) vs. large static base-rate offsets at 1m (z_bar_{1m} = -0.1627, "
     "||z||_RMS = 0.2041) and 1d (z_bar_{1d} = +0.1275, ||z||_RMS = 0.1285). In denominator-free RMS units, logit sensitivity decays "
     "monotonically with horizon (0.0403 at 1m -> 0.0213 at 5m -> 0.0136 at 15m -> 0.0058 at 1d for K=7), while in centered L2 units "
-    "structured branches (21.36%–34.29%) and scale groups (30.54%–33.42%) generally exceed matched 341-channel (20.61% ± 0.90%) "
+    "structured branches (21.36%–34.29%) and scale groups (Micro 40.20%, Meso 49.90%, Macro 37.07%) generally exceed matched 341-channel (20.61% ± 0.90%) "
     "and 682-channel (30.97% ± 1.07%) random controls, though K=3 sits within 1 SD of the control mean and accuracy deltas are "
     "indistinguishable from random channels. Second, we construct a <b>Non-Overlapping Contiguous Per-Ticker Backtest</b> (stepping by S_h = ceil(tau_h / 32) "
     "windows so 60m and 1d holds never overlap) with true <b>State-Dependent Turnover</b> |w_{i,k} - w_{i,k-1}|, compounded percentage drawdowns, Bailey &amp; "
     "López de Prado (2014) <b>Deflated Sharpe Ratios (DSR)</b> across all M=35 trials, and a <b>Single Frozen 2024 Policy</b> applied untouched to 2025 and 2026. "
-    "Under 1-bar execution delay (t+1 entry, y - y_1m), 2024 net Sharpe at 1.5 bps drops from +2.04 to +0.59, while frozen transfer to 2025 and 2026 yields "
-    "negative net returns (-2.38 and -5.10). Fixed demeaned rules at 3m and 5m yield sub-0.5 bp breakeven thresholds (0.35–0.44 bps) and Deflated Sharpe "
-    "Ratios of DSR <= 0.030, showing that pre-cost edge cannot survive taker fees or adverse selection in passive maker execution."
+    "Under 1-bar execution delay (t+1 entry, y - y_1m), removing prompt 1-minute order-flow continuation cuts 2024 net Sharpe from +2.04 to +0.48, and "
+    "flips 2025 gross Sharpe from +1.01 to -0.15, while frozen transfer to 2025 and 2026 yields negative net returns (-0.99 and -1.67). Fixed demeaned rules "
+    "at 3m and 5m yield sub-0.5 bp breakeven thresholds (0.35–0.44 bps) and Deflated Sharpe Ratios of DSR <= 0.030, showing that pre-cost edge cannot survive taker fees or adverse selection in passive maker execution."
 )
 
 p3_elements = [
     ("h1", "1. Introduction: Eliminating Denominator & Selection Biases"),
-    ("p", "Rigorous evaluation of neural trading signals requires addressing two subtle methodological traps: (1) <b>Ablation Denominator & Off-Distribution Bias:</b> Zero-clamping post-SiLU activations shifts the mean input to the linear projection W_proj, and normalizing perturbation norms ||Δz_h||_2 by uncentered ||z_h||_2 artificially inflates sensitivity at horizons where the mean logit z_bar_h is close to zero; and (2) <b>Backtest Overlap, Turnover & Selection Bias:</b> Evaluating 60m or 1d (390m) returns on 32-minute sliding windows double-counts overlapping returns, charging 2c per bar ignores position persistence |w_{i,k} - w_{i,k-1}|, and cherry-picking different (horizon, hurdle) pairs per year inflates out-of-sample Sharpe ratios."),
+    ("p", "Evaluating neural trading signals requires addressing two subtle methodological traps: (1) <b>Ablation Denominator & Off-Distribution Bias:</b> Zero-clamping post-SiLU activations shifts the mean input to the linear projection W_proj, and normalizing perturbation norms ||Δz_h||_2 by uncentered ||z_h||_2 artificially inflates sensitivity at horizons where the mean logit z_bar_h is close to zero; and (2) <b>Backtest Overlap, Turnover & Selection Bias:</b> Evaluating 60m or 1d (390m) returns on 32-minute sliding windows double-counts overlapping returns, charging 2c per bar ignores position persistence |w_{i,k} - w_{i,k-1}|, and cherry-picking different (horizon, hurdle) pairs per year inflates out-of-sample Sharpe ratios."),
     ("h1", "2. Controlled Mechanistic Ablation & Denominator Decomposition"),
     ("p", "Let f_{t,c} denote channel c in {1..2046} of the concatenated post-SiLU tokenizer output prior to W_proj in R^{2046 x 2048}, with empirical mean μ_c = E_{t}[f_{t,c}]. For any channel subset C (|C|=341 for single branches, |C|=682 for 2-branch groups), we evaluate both <b>In-Distribution Mean-Ablation</b> (f_{t,c} <- μ_c for c in C) and <b>Zero-Ablation</b> (f_{t,c} <- 0), benchmarked against <b>Random 341-Channel and 682-Channel Controls</b> drawn uniformly across all 2,046 channels. For each horizon h, we compute Uncentered Relative L2, Variance-Normalized Centered L2, and Denominator-Free RMS Logit Shift:"),
     ("eq", "L2_raw(h) = ||Delta z_h||_2 / ||z_h^(0)||_2,   L2_cent(h) = ||(Delta z_h) - mean(Delta z_h)||_2 / ||z_h^(0) - mean(z_h^(0))||_2,   RMS(Delta z_h) = ||Delta z_h||_2 / sqrt(N)", "1"),
     ("fig", ("papers/figures/fig4_kernel_ablation.png", "Figure 4: (a) Variance-Normalized Centered Logit L2 Perturbation under In-Distribution Mean-Ablation and Zero-Ablation vs. Random 341ch/682ch Controls. (b) Decomposition of the 15m denominator effect: ||z_h||_RMS reaches its minimum at 15m (0.0405), inflating uncentered L2_raw, whereas denominator-free RMS(Δz_h) decays monotonically.")),
     ("tbl", (TBL_ABLATION, "Table 1: Controlled 7-horizon mechanistic ablation on the 2024 OOS cohort (N=8,044 windows across 15 tickers; 25-draw random controls reported with mean ± SD).", [86, 24, 62, 46, 46, 46, 120, 80])),
-    ("p", "Table 1 and Figure 4 establish three rigorous mechanistic results: (1) <b>Proof of the 15m Denominator Effect:</b> Baseline logit RMS is ||z||_RMS = 0.2041 at 1m (mean -0.1627), 0.1017 at 3m, 0.0799 at 5m, <b>0.0405 at 15m (mean +0.0118)</b>, 0.0401 at 30m, 0.0618 at 60m, and 0.1285 at 1d (mean +0.1275). Because ||z||_RMS is 5x smaller at 15m than at 1m, uncentered L2_raw peaks at 15m (33.61% for K=7), whereas denominator-free <b>RMS(Δz_h) decays monotonically with horizon</b> (0.0403 at 1m -> 0.0136 at 15m -> 0.0058 at 1d); (2) <b>Structured Branches vs. 25-Draw Random Controls:</b> Under Mean-Ablation, structured kernel centered L2 shifts (21.36%–34.29%) and scale groups (Micro 30.54%, Meso 31.82%, Macro 33.42%) generally exceed matched Random 341-Ch (20.61% ± 0.90%) and Random 682-Ch (30.97% ± 1.07%) controls, though shortest kernel K=3 (21.36%) sits within 1 SD of the control mean; and (3) <b>Absence of Statistically Significant Branch Specialization:</b> Across all ablated branches, overall ΔNZAUC ranges from -0.0004 to -0.0009 across single kernels and -0.0004 to -0.0008 across groups, with all 95% paired bootstrap confidence intervals spanning zero (e.g., K=3: [-0.0016, +0.0008]; Macro: [-0.0020, +0.0004]). Crucially, group deltas sit entirely within the random 682-channel control spread (-0.0005 ± 0.0006), proving that ranking degradation is governed by channel count rather than receptive-field semantics."),
+    ("p", "Table 1 and Figure 4 establish three empirical mechanistic results: (1) <b>Quantification of the 15m Denominator Effect:</b> Baseline logit RMS is ||z||_RMS = 0.2041 at 1m (mean -0.1627), 0.1017 at 3m, 0.0799 at 5m, <b>0.0405 at 15m (mean +0.0118)</b>, 0.0401 at 30m, 0.0618 at 60m, and 0.1285 at 1d (mean +0.1275). Because ||z||_RMS is 5x smaller at 15m than at 1m, uncentered L2_raw peaks at 15m (33.61% for K=7), whereas denominator-free <b>RMS(Δz_h) decays monotonically with horizon</b> (0.0403 at 1m -> 0.0136 at 15m -> 0.0058 at 1d); (2) <b>Structured Branches vs. 25-Draw Random Controls:</b> Under Mean-Ablation, structured kernel centered L2 shifts (21.36%–34.29%) and scale groups (Micro 40.20%, Meso 49.90%, Macro 37.07%) generally exceed matched Random 341-Ch (20.61% ± 0.90%) and Random 682-Ch (30.97% ± 1.07%) controls, though shortest kernel K=3 (21.36%) sits within 1 SD of the control mean; and (3) <b>Absence of Statistically Significant Branch Specialization:</b> Across all ablated branches, overall ΔNZAUC ranges from -0.0004 to -0.0009 across single kernels and -0.0004 to -0.0008 across groups, with all 95% paired bootstrap confidence intervals spanning zero (e.g., K=3: [-0.0014, +0.0006]; Macro: [-0.0021, +0.0005]). Group deltas sit entirely within the random 682-channel control spread (-0.0005 ± 0.0006), consistent with ranking degradation being governed by channel count rather than receptive-field semantics."),
     ("h1", "3. Non-Overlapping Contiguous Backtest, State Turnover & Frozen 2024 Policy Transfer"),
-    ("p", "Within each contiguous (shard, ticker) trajectory (spaced by s_win = 32 minutes), we step by S_h = max(1, ceil(tau_h / 32)) windows (S_h = 1 for 1m–30m, S_{60m} = 2 = 64m, S_{1d} = 13 = 416m) so holding periods <b>never overlap</b>. At each step k, we track state-dependent turnover Turn_{i,k} = |w_{i,k} - w_{i,k-1}| (including initial entry from 0 and terminal exit to 0 at shard boundaries) alongside isolated round-trip cost 2|w_{i,k}|, compound portfolio equity E_t = E_{t-1}(1 + R_{p,t}^{net}) across the 15 tickers to compute bounded percentage MaxDD in [0%, 100%], and compute the Bailey &amp; López de Prado (2014) <b>Deflated Sharpe Ratio (DSR)</b> adjusting for skewness, kurtosis, and M=35 trials:"),
-    ("eq", "R_{p,t}^net(c) = (1 / max(5, N_act,t)) * sum_{i=1}^{15} [ w_{i,t} (exp(y_{i,t,h}) - 1) - c * |w_{i,t} - w_{i,t-1}| ],   BE_state = Gross_mean / mean(|Delta w|)", "2"),
+    ("p", "Within each contiguous (shard, ticker) trajectory (spaced by s_win = 32 minutes), we step by S_h = max(1, ceil(tau_h / 32)) windows (S_{60m} = 2 = 64 minutes, yielding 258 non-overlapping portfolio steps per calendar year across 15 tickers, spanning 275.2 active trading hours, or a fraction of 0.168 of the 1,638 annual US trading hours). Holding periods therefore strictly never overlap. Step net portfolio return is R_{p,t}^{net}(c) = R_{p,t}^{gross} - c * Turnover_t, and sample cumulative return over the 258 steps is CumRet_{sample} = prod_{t=1}^{258}(1 + R_{p,t}^{net}) - 1 approx 258 * mean(R_{p,t}^{net}). Annualized Sharpe ratio is strictly computed from step mean and volatility via SR = sqrt(258) * mean(R_{step}) / std(R_{step}) = 16.062 * mean(R_{step}) / std(R_{step}), and state break-even cost satisfies BE_state = mean(R_gross) / mean(Turnover). Compounding across the 258-step sample grid provides an empirically verifiable basis for percentage MaxDD in [0%, 100%], alongside the Bailey &amp; López de Prado (2014) <b>Deflated Sharpe Ratio (DSR)</b> adjusting for skewness, kurtosis, and M=35 trials (with passive benchmarks holding M=1 unsearched):"),
+    ("eq", "R_{p,t}^net(c) = R_{p,t}^gross - c * Turnover_t,   BE_state = mean(R_gross) / mean(Turnover)", "2"),
     ("fig", ("papers/figures/fig5_slippage_friction_curve.png", "Figure 5: (a) Non-overlapping state-hold breakeven cost (bps) across horizons under a fixed τ=1.0σ demeaned hurdle in 2024, 2025, and 2026. (b) Untouched out-of-sample transfer of the Frozen 2024 Policy (60m, τ=0.0σ) vs. fixed 3m (τ=1.0σ) across execution cost tiers.")),
-    ("tbl", (TBL_BACKTEST, "Table 2: Non-overlapping contiguous backtest of the Frozen 2024 Policy (1-bar delayed execution) vs. Buy-and-Hold and Market-Neutral baselines across 2024–2026.", [82, 42, 44, 34, 38, 56, 48, 52, 36, 42, 40])),
-    ("p", "Table 2 and Figure 5 yield three transparent, execution-grounded conclusions: (1) <b>State-Hold vs. Isolated Breakeven:</b> Because contiguous positions persist across steps when conviction remains in the same direction, average one-way turnover per active step is 0.680–0.873 (for τ=0.0σ) and ~1.60–1.75 (for τ=1.0σ), so true state-hold breakeven cost BE_state = Gross / Turn is 1.15x–2.41x higher than isolated 1-bar round-trip breakeven Gross / 2; (2) <b>Untouched OOS Transfer and 1-Bar Execution Delay:</b> Under a 1-bar execution delay model (approximated by y_{delay,t,h} = y_{t,h} - y_{t,1m}, removing immediate continuation drift), net Sharpe of the Frozen 2024 Policy at 1.5 bps fee drops from +2.04 to <b>+0.59 in 2024</b>. Untouched frozen transfer to subsequent years reveals complete out-of-sample decay, collapsing to net Sharpes of <b>-2.38 in 2025</b> and <b>-5.10 in 2026</b>. Benchmarking against Equal-Weighted Buy-and-Hold (+0.42 / +0.38 / -0.31, M=1 unsearched) and Market-Neutral (+0.21 / +0.14 / +0.08) confirms that long-horizon returns are market-beta driven; and (3) <b>Friction Bounds &amp; Adverse Selection:</b> Under the fixed demeaned policy at 3m and 5m, gross alpha is positive across all three years (+0.26 to +0.75 bps/trade), but state-hold breakeven is <b>0.35–0.44 bps</b> and Deflated Sharpe Ratio across M=35 trials is <b>DSR &lt;= 0.030</b>. Passive maker deployment is subject to adverse selection (queue decay on adverse moves), meaning standalone OHLCV signals cannot generate deployable trading returns without full order-book queue modeling."),
+    ("tbl", (TBL_BACKTEST, "Table 2: Non-overlapping contiguous backtest of the Frozen 2024 Policy (1-bar delayed execution) vs. Buy-and-Hold and Market-Neutral baselines across 2024–2026 (synchronous 258-step grid).", [80, 36, 42, 30, 52, 50, 52, 52, 38, 42, 40])),
+    ("p", "Table 2 and Figure 5 yield three execution-grounded conclusions: (1) <b>State-Hold vs. Isolated Breakeven:</b> Because contiguous positions persist across steps when conviction remains in the same direction, average one-way turnover per active step is 0.680–0.829 (for τ=0.0σ) and ~1.60–1.75 (for τ=1.0σ), so true state-hold breakeven cost BE_state = Gross / Turn is strictly satisfied in every row (e.g., 2024: +1.88 / 0.829 = +2.27 bps). For Market-Neutral (turnover 0.450/step), cost drag at 1.5 bps is 0.675 bps/step, exceeding gross return (+0.32 bps) and flipping net return to -0.355 bps/step (std 24.48 bps), net Sharpe to -0.23, and sample CumRet to -0.92% (-3.2% ann.); (2) <b>Untouched OOS Transfer and 1-Bar Execution Delay:</b> Under a 1-bar execution delay model (approximated by y_{delay,t,h} = y_{t,h} - y_{t,1m}, removing immediate continuation drift), net Sharpe of the Frozen 2024 Policy at 1.5 bps fee drops from +2.04 to <b>+0.48 in 2024</b> (gross +1.88 bp, net +0.64 bp, std 21.23 bp, sample CumRet +1.64%, +5.82% ann. across 3,843 active ticker positions on 258 steps). Crucially, removing prompt 1-minute continuation flips 2025 gross Sharpe from +1.01 to -0.15 (gross mean -0.20 bp, std 21.42 bp), demonstrating that apparent long-horizon directional predictability was predominantly immediate order-flow continuation. Untouched frozen transfer to subsequent years reveals complete out-of-sample decay, collapsing to net Sharpes of <b>-0.99 in 2025</b> (-3.40% sample CumRet, -12.2% ann.) and <b>-1.67 in 2026</b> (-6.01% sample CumRet, -21.4% ann.). Benchmarking against Equal-Weighted Buy-and-Hold (+0.42 / +0.38 / -0.31; CumRet +1.42% / +1.19% / -0.41%, M=1 unsearched) and Market-Neutral (+0.21 / +0.14 / +0.08) on the identical 258-step grid confirms that long-horizon returns are market-beta driven; and (3) <b>Friction Bounds &amp; Adverse Selection:</b> Under the fixed demeaned policy at 3m and 5m, gross alpha is positive across all three years (+0.26 to +0.75 bps/trade), but state-hold breakeven is <b>0.35–0.44 bps</b> and Deflated Sharpe Ratio across M=35 trials is <b>DSR &lt;= 0.030</b>. Passive maker deployment is subject to adverse selection (queue decay on adverse moves), meaning standalone OHLCV signals cannot generate deployable trading returns without full order-book queue modeling."),
     ("h1", "4. Conclusion"),
-    ("p", "By introducing in-distribution mean-ablation, 25-draw random-channel controls, denominator-free RMS logit metrics, 1-bar delayed execution, non-overlapping state-turnover accounting, Deflated Sharpe Ratios, and untouched frozen-policy OOS evaluation, we provide a mathematically grounded audit of billion-parameter models under real-world execution frictions (DSR <= 0.030), establishing that raw directional signals cannot overcome bid-ask spreads without order-flow execution mechanisms.")
+    ("p", "By introducing in-distribution mean-ablation, 25-draw random-channel controls, denominator-free RMS logit metrics, 1-bar delayed execution, non-overlapping state-turnover accounting, Deflated Sharpe Ratios, and untouched frozen-policy OOS evaluation, we provide an empirical audit of billion-parameter models under real-world execution frictions (DSR <= 0.030), establishing that raw directional signals cannot overcome bid-ask spreads without order-flow execution mechanisms.")
 ]
 
 compile_paper(
@@ -826,7 +828,6 @@ compile_paper(
         "Constantinides, G. M. (1986). Capital market equilibrium with transaction costs. Journal of Political Economy, 94(4), 842–862.",
         "Gu, S., Kelly, B., & Xiu, D. (2020). Empirical asset pricing via machine learning. The Review of Financial Studies, 33(5), 2223–2273.",
         "Almgren, R., & Chriss, N. (2001). Optimal execution of portfolio transactions. Journal of Risk, 3, 5–40.",
-        "Olah, C., et al. (2020). Zoom in: An introduction to circuits. Distill, 5(3), e00024-001.",
         "Zhang, Z., Zohren, S., & Roberts, S. (2019). DeepLOB: Deep convolutional neural networks for limit order books. IEEE TSP, 67(11), 3001–3012."
     ]
 )
@@ -836,34 +837,19 @@ compile_paper(
 # COMPILE UNIFIED FLAGSHIP MONOGRAPH: Merging All 3 Studies into One Cohesive Paper
 # =============================================================================
 u_abstract = (
-    "We present a unified architectural, continual-learning, mechanistic, and economic audit of the <b>Financial Conformer "
-    "Foundation Model</b>, a <b>1,189,135,740-parameter (1.189B)</b> hybrid convolution-attention network pre-trained across "
-    "14 calendar years (2010–2023; 168 monthly Parquet shards, 15,030 steps) of 1-minute US equity bars and evaluated across "
-    "<b>32,100 stratified contiguous windows</b> spanning Pre-2024 historical crises (March 2020 COVID-19 crash, June 2022 Fed "
-    "rate shock, December 2023 anchor) and strictly out-of-sample 2024, 2025, and 2026 cohorts across 15 liquid tickers. "
-    "Consolidating our complete checkpoint and telemetry lineage into a single self-contained monograph, we establish four primary "
-    "contributions: (1) <b>Base-Rate &amp; Flat-Bar Decomposition:</b> We prove mathematically and empirically that because 6.46%–7.86% "
-    "of 1-minute bars are exact zero-return flat bars (y=0), strictly positive returns occur on only 46.15%–46.51% of 1m bars, "
-    "causing uncalibrated z &gt; 0 thresholding to collapse to all-down predictions (53.49%–53.85% raw accuracy, matching a "
-    "FlatBar-Down heuristic). Removing this static threshold bias via <b>Strict Non-Zero ROC-AUC (NZAUC)</b>, <b>Causal Prior-Calibrated "
-    "Accuracy</b>, and <b>|y|-Quartile Stratification</b> characterizes threshold-independent ranking: 1m All-Bar AUC of <b>0.5375 / 0.5262 / 0.5274</b>, "
-    "1m Strict Non-Zero AUC of <b>0.5177 / 0.5161 / 0.5198</b>, and active two-sided 5m Non-Zero Accuracy of <b>51.74%</b> (2024), <b>49.90%</b> (2025), and "
-    "<b>51.18%</b> (2026); (2) <b>Exact Cosine Geometry, Adam Arithmetic &amp; Pre-2024 Crisis Forgetting:</b> Using the exact float64 "
-    "Hilbert-space Law of Cosines, we show that online walk-forward adaptation shifts global weights by <b>0.5439%</b> (2024, "
-    "cos_exact = <b>0.99998521</b> &gt;= sqrt(1-δ_F^2)) and <b>0.5725%</b> (2025, cos_exact = <b>0.99998361</b>) with nearly orthogonal "
-    "annual update vectors (cos = -0.009270, 90.53°). Decomposing drift into RMS weight norm ||W||_RMS and RMS update ||ΔW||_RMS "
-    "proves that the tokenizer's high plasticity reflects <b>4.00x larger absolute updates</b> (3.030e-4 vs. 7.575e-5) rather than "
-    "small weight norms (||W||_RMS = 0.02574 vs. 0.01432). Running a 960-step control on shuffled labels (AdamW, lr=1e-5, seed 42) confirms Tokenizer update "
-    "displacement (3.031e-4) sits directly on the theoretical Adam noise ceiling (3.098e-4) with an orthogonal update cosine of +0.019184, "
-    "providing no evidence of structured continual drift beyond optimizer Brownian motion. "
-    "Pre-2024 stress tests confirm zero overall forgetting (0.5130 -> 0.5135 AUC; COVID 1m ΔNZAUC = +0.0044 [-0.0112, +0.0200]), while on strictly forward 2026 OOS data, testing on strict non-zero AUC "
-    "yields ΔNZAUC = +0.0013 [-0.0074, +0.0100] (raw p = 0.77, Holm p = 1.0, not significant under multiplicity control); (3) <b>Controlled Mechanistic Ablation:</b> "
-    "Using in-distribution mean-ablation and 25-draw random 341ch/682ch controls, we prove that the 15m peak in uncentered ||Δz||_2 / ||z||_2 is a denominator "
-    "effect from z_bar_{15m} ≈ 0, whereas denominator-free RMS(Δz) decays monotonically and centered L2 sensitivity generally exceeds random "
-    "controls (20.61% ± 0.90% and 30.97% ± 1.07%), while ranking accuracy deltas are indistinguishable from random channels; and (4) <b>Non-Overlapping Execution Economics:</b> "
-    "Under non-overlapping holding periods, 1-bar delayed execution (t+1 entry, y - y_1m), state-dependent turnover, Deflated Sharpe Ratios, and a single frozen 2024 policy, we "
-    "show that net Sharpe at 1.5 bps fee drops to +0.59 in 2024, while frozen transfer to 2025 and 2026 yields negative net Sharpes (-2.38 and -5.10), "
-    "yielding DSR <= 0.028 across all trials."
+    "We conduct an empirical audit of scale, continual walk-forward adaptation, and execution economics in a "
+    "<b>1,189,135,740-parameter (1.189B)</b> financial Conformer pre-trained across 14 calendar years (2010–2023; 168 monthly shards, 15,030 steps) "
+    "of 1-minute US equity bars and evaluated on 32,100 stratified contiguous windows (2020–2026: 7,896 pre-2024 historical crisis stress windows and 24,204 forward out-of-sample windows across 2024–2026). Across out-of-sample cohorts, "
+    "scaling from a 63-parameter linear ridge baseline to 1.189B parameters yields <b>no capacity advantage beyond the empirical noise floor</b> "
+    "(±0.010 AUC), with ridge matching or exceeding the Conformer at 1m (<b>0.5289 vs. 0.5177 NZAUC</b>; a +0.0112 advantage slightly exceeding the empirical noise floor). Decomposing predictions reveals that apparent "
+    "pre-training gains stem from identifying flat, uninformative bars rather than predicting price direction: on strict non-zero returns, "
+    "discrimination is small but positive (<b>0.5177 NZAUC at 1m</b>), while decaying toward chance at longer horizons. In continual walk-forward fine-tuning, "
+    "weight updates exhibit a sharp structural divergence: input tokenizer updates sit at <b>97.9% of the Adam Brownian noise ceiling</b> "
+    "(η*sqrt(K) = 3.10e-4), whereas deep backbone updates are <b>3.4x smaller</b> under real labels than under shuffled noise (24.4% vs. 83.5% of ceiling; "
+    "||ΔW||_RMS = 7.58e-5 vs. 2.59e-4), consistent with weak gradient signals suppressed by Adam damping. Paired block bootstraps across 35 hypothesis tests "
+    "confirm no statistically significant transfer gains after Holm-Bonferroni correction (p >= 0.50). Finally, introducing a 1-bar execution delay eliminates "
+    "apparent 60m profits, cutting net Sharpe from +2.04 to +0.48 in 2024 and yielding a Deflated Sharpe Ratio of DSR <= 0.028 across all horizons. "
+    "These findings establish an empirical null result on parameter scale for intraday return forecasting."
 )
 
 import re
@@ -895,13 +881,24 @@ for item in raw_u_elements:
     elif kind == "p":
         txt = item[1]
         if sec_idx == 5:
-            txt = txt.replace("Table 1 proves", "Table 6 proves")
+            txt = txt.replace("Table 1 establishes", "Table 6 establishes")
+            txt = txt.replace("Table 1 proves", "Table 6 establishes")
+            txt = txt.replace("(Table 1)", "(Table 6)")
+            txt = txt.replace("Table 1", "Table 6")
         elif sec_idx == 6:
             txt = txt.replace("Tables 2 and 3 resolve", "Tables 7 and 8 resolve")
+            txt = txt.replace("(Table 3)", "(Table 8)")
+            txt = txt.replace("(Table 2)", "(Table 7)")
+            txt = txt.replace("Table 3", "Table 8")
+            txt = txt.replace("Table 2", "Table 7")
         elif sec_idx == 7:
             txt = txt.replace("Table 1 and Figure 4", "Table 9 and Figure 4")
+            txt = txt.replace("(Table 1)", "(Table 9)")
+            txt = txt.replace("Table 1", "Table 9")
         elif sec_idx == 8:
             txt = txt.replace("Table 2 and Figure 5", "Table 10 and Figure 5")
+            txt = txt.replace("(Table 2)", "(Table 10)")
+            txt = txt.replace("Table 2", "Table 10")
         u_elements.append(("p", txt))
     else:
         u_elements.append(item)
@@ -909,8 +906,8 @@ for item in raw_u_elements:
 
 compile_paper(
     filename="papers/Paper_Unified_Financial_Conformer_Monograph.pdf",
-    running_header="Airborne AI // Unified 1.189B Financial Conformer Monograph",
-    title="The 1.189-Billion Parameter Financial Conformer: An Empirical Audit of Pre-Training Dynamics, Base-Rate Calibration, Walk-Forward Drift, Controlled Kernel Ablation, and Non-Overlapping Execution Economics",
+    running_header="Airborne AI // A Null Result on Scale in the 1.189B Financial Conformer",
+    title="A Null Result on Scale in Intraday Return Prediction: An Empirical Audit of Pre-Training Dynamics, Base-Rate Calibration, Walk-Forward Drift, Controlled Kernel Ablation, and Non-Overlapping Execution Economics in the 1.189B Financial Conformer",
     keywords="Financial Foundation Models, Conformer Architecture, Base-Rate Calibration, Continual Walk-Forward Learning, Controlled Mechanistic Ablation, Deflated Sharpe Ratio.",
     abstract_text=u_abstract,
     elements=u_elements,
@@ -921,9 +918,6 @@ compile_paper(
         "Su, J., et al. (2024). RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568, 127063.",
         "Zhang, B., & Sennrich, R. (2019). Root mean square layer normalization. NeurIPS, 32.",
         "Shazeer, N. (2020). GLU variants improve transformer. arXiv:2002.05202.",
-        "Nie, Y., et al. (2023). A time series is worth 64 words: Long-term forecasting with transformers (PatchTST). ICLR.",
-        "Das, A., et al. (2024). A decoder-only foundation model for time-series forecasting (TimesFM). ICML.",
-        "Ansari, A. F., et al. (2024). Chronos: Learning the language of time series. TMLR.",
         "Kim, T., et al. (2022). Reversible instance normalization for accurate time-series forecasting (RevIN). ICLR.",
         "Cont, R. (2001). Empirical properties of asset returns: stylized facts and statistical issues. Quantitative Finance, 1(2), 223–236.",
         "Sirignano, J., & Cont, R. (2019). Universal features of price formation in financial markets. Quantitative Finance, 19(9), 1449–1459.",