Sync README.md

README.md

@@ -58,64 +58,17 @@ While traditional academic financial machine learning has historically relied on
 
 ---
 
-## 🔬 Empirical Breakthrough Telemetry
+## 🔬 Verified Pre-Training Telemetry (855 Logged Steps, 2010–2023)
 
-Financial time series exhibit extreme entropy with very low Signal-to-Noise Ratios ($SNR < 0.05$). Under standard semi-strong market efficiency, a binary directional bet has a theoretical random entropy baseline of:
+Financial time series exhibit extreme entropy with very low Signal-to-Noise Ratios ($SNR < 0.05$). Under standard semi-strong market efficiency, the weighted binary cross-entropy floor ($0.5 \ln 2$) is **`0.34657`** ($\ln 2 \approx 0.69315$). Across 3 full epochs ($15,030$ steps; $855$ logged telemetry checkpoints in [`logs/pretraining_telemetry_verified.json`](logs/pretraining_telemetry_verified.json)) on NVIDIA A100-SXM4-80GB hardware:
 
-$$\mathcal{H}_{\text{random}} = -\ln(0.5) \approx \mathbf{0.6931}$$
-
-Classical quant funds (e.g., Renaissance Technologies, Two Sigma) built quantitative dynasties on directional edges of **51.5% to 53.5%**. During active 3-Epoch training on NVIDIA A100-SXM4-80GB hardware, the ULTRON 1.189B Conformer achieved:
-
-- **Total Loss Floor**: Decreased monotonically to **`0.3273`**.
-- **Directional Loss**: Penetrated down to **`0.6545`** (decisively breaking below the 0.6931 random barrier).
-- **Multi-Horizon Directional Accuracy**:
-  - **Overall Aggregate**: **`58.9% – 69.6%`**
-  - **1-Minute Horizon**: **`75.0% – 87.5%`** (Peaks at `100.0%`)
-  - **5-Minute Horizon**: **`62.5% – 75.0%`** (Peaks at `100.0%`)
-  - **15-Minute Horizon**: **`75.0%`**
-  - **1-Hour Horizon**: **`75.0% – 87.5%`**
-- **Information Coefficient (IC)**: Estimated $> 0.15$ across tactical holding horizons.
-
-<div align="center">
-  <img src="assets/training_telemetry.png" alt="ULTRON Pre-Training Telemetry" width="95%"/>
-</div>
-
----
-
-## 📚 Multi-Paper Research Agenda
-
-The findings and empirical weight trajectories from this foundation training run form the core of a three-paper academic and quantitative research portfolio:
-
-```
-                            ┌──────────────────────────────────────────────┐
-                            │  1.189B Conformer Pre-Training (2010-2023)  │
-                            │   90 GB OHLCV-1m | 3 Full Foundation Epochs  │
-                            └──────────────────────┬───────────────────────┘
-                                                   │
-         ┌─────────────────────────────────────────┼─────────────────────────────────────────┐
-         ▼                                         ▼                                         ▼
-┌─────────────────────────────────┐       ┌─────────────────────────────────┐       ┌─────────────────────────────────┐
-│            PAPER 1              │       │            PAPER 2              │       │            PAPER 3              │
-│  Scaling Laws & Multi-Scale     │       │   Zero-Shot Alpha & The OOS     │       │   Mechanistic Interpretability  │
-│  Inductive Biases in Financial  │       │   Frontier (2024-2026 Quarantined│      │   of Multi-Scale Financial      │
-│  Time-Series                    │       │   Regimes)                      │       │   Conformer Latents             │
-├─────────────────────────────────┤       ├─────────────────────────────────┤       ├─────────────────────────────────┤
-│ • Power-law scaling (1.189B)    │       │ • Zero-shot cross-regime alpha  │       │ • Inception Conv kernel anatomy │
-│ • Multi-scale Inception [3-127m]│       │ • Untouched 2024-2026 evaluation│       │   (3m noise vs 127m liquidity)  │
-│ • ReVIN non-stationarity proof  │       │ • Sharpe, Sortino & Max DD      │       │ • Attention head specialization │
-│ • Multi-horizon co-optimization │       │ • Execution slippage dynamics   │       │ • Latent regime phase shifts    │
-└─────────────────────────────────┘       └─────────────────────────────────┘       └─────────────────────────��───────┘
-```
-
-1. **Paper 1: Scaling Laws & Neural Architectures in Financial Time-Series**
-   - Proves empirical neural scaling laws across 1.189 Billion parameters on continuous financial OHLCV. Demonstrates that combining multi-scale temporal convolutions with self-supervised masked candlestick objectives ($15\%$ masking) overcomes noise memorization.
-2. **Paper 2: The Out-of-Sample Frontier: Zero-Shot Alpha Across Unseen Regimes (2024–2026)**
-   - Zero-shot evaluation on pristine, held-out market regimes (post-2023 inflation, mega-cap tech dispersion, aggressive central bank rate cycles). Compares Epoch 1 vs Epoch 2 vs Epoch 3 checkpoints across real-world Sharpe ratio, Sortino ratio, and transaction-cost-adjusted execution alpha.
-3. **Paper 3: Mechanistic Interpretability of Conformer Multi-Scale Latents**
-   - Direct dissection of the 20 Conformer attention heads and multi-scale convolution filters. Proves that lower kernel scales ($K=3, 7$) act as microstructure bid-ask absorption filters, while higher scales ($K=63, 127$) identify institutional order-flow imbalances and macro continuation.
+- **Epoch 1 (Steps 320–4,970, $n=183$)**: Mean Joint Loss **`0.3448 ± 0.0079`** (Min **`0.3176`**), Mean BCE **`0.6896`**, Mean Directional Acc **`53.24%`** (Max Batch **`71.40%`**).
+- **Epoch 2 (Steps 5,030–9,900, $n=204$)**: Mean Joint Loss **`0.3447 ± 0.0071`** (Min **`0.3242`**), Mean BCE **`0.6894`**, Mean Directional Acc **`53.25%`** (Max Batch **`75.00%`**).
+- **Epoch 3 (Steps 10,360–15,030, $n=468$)**: Mean Joint Loss **`0.3450 ± 0.0067`** (Min **`0.3234`**), Mean BCE **`0.6899`**, Mean Directional Acc **`52.66%`** (Max Batch **`75.00%`**).
+- **Monotonic Loss Variance Contraction**: Batch loss standard deviation contracts by **`-15.2%`** across epochs ($0.0079 \to 0.0071 \to 0.0067$), while global Frobenius parameter shift decelerates from **`10.0001%`** (`Ep1→Ep2`) to **`5.7305%`** (`Ep2→Ep3`).
 
 <div align="center">
-  <img src="assets/multi_horizon_forecast.png" alt="Multi-Horizon Forecast" width="95%"/>
+  <img src="papers/figures/fig1_pretraining_curves.png" alt="ULTRON Pre-Training Telemetry" width="95%"/>
 </div>
 
 ---
@@ -126,112 +79,79 @@ Permanent weight archives are versioned on Hugging Face ([`Ultron28/ohlcv-transf
 
 | Checkpoint Name | Size (BF16) | Status | Description |
 | :--- | :--- | :--- | :--- |
-| **`checkpoint_1b_epoch1.pt`** | **4.79 GB** | ✅ **Archived & Verified** | Full Epoch 1 Milestone (Steps 1–5,010 across 168 shards) |
-| **`checkpoint_1b_epoch2.pt`** | **4.79 GB** | ✅ **Archived & Verified** | Full Epoch 2 Milestone (Steps 5,010–10,100 across 168 shards) |
-| **`checkpoint_1b_epoch3.pt`** | **4.47 GB** | ✅ **Archived & Verified** | Full Foundation Pass Complete (Steps 10,100–15,030 across 168 shards) |
-| **`checkpoint_1b_latest.pt`** | **4.47 GB** | ⚡ **Rolling Live** | Terminal foundation checkpoint weights |
-
----
-
-## 📐 Model Architecture Specifications
-
-| Hyperparameter | Value | Description |
-| :--- | :--- | :--- |
-| **Total Parameters** | **1,189,135,740** | **1.189 Billion Parameters** |
-| **Model Dimension ($d_{\text{model}}$)** | **2,048** | Latent embedding space per temporal token |
-| **Layers ($N_{\text{layers}}$)** | **20** | Deep alternating Conformer blocks |
-| **Attention Mechanism** | **Grouped-Query Attention (GQA)** | 16 Query Heads, 8 Key/Value Heads (Head Dim = 128) |
-| **Positional Encoding** | **RoPE (Rotary Position Embeddings)** | Dynamic rotary complex embeddings |
-| **FFN Dimension ($d_{\text{ff}}$)** | **5,504** | SwiGLU Gated Linear Activation |
-| **Context Length ($T$)** | **2,048 Minutes** | ~34 trading hours (covers full intraday + overnight gap) |
-| **Convolution Tokenizer** | **6 Multi-Scale Kernels** | Overlapping $[3, 7, 15, 31, 63, 127]$ with Stride $S=4$ |
-| **Tokens per Context** | **512 Tokens** | 4× temporal downsampling with full receptive fidelity |
-| **Normalizer** | **Reversible Instance Normalization (RevIN)** | Dynamic non-stationary distribution tracking |
-| **Optimizer** | **8-Bit AdamW (bitsandbytes)** | Static VRAM: **2.38 GB** (BF16 weights) |
-| **Hardware** | **NVIDIA A100-SXM4-80GB** | Native Tensor Core BF16 (100% compute saturation) |
+| **`checkpoint_1b_epoch1.pt`** | **4.79 GB** | ✅ **Archived & Verified** | Full Epoch 1 Milestone (Step 5,010 across 168 shards) |
+| **`checkpoint_1b_epoch2.pt`** | **4.79 GB** | ✅ **Archived & Verified** | Full Epoch 2 Milestone (Step 10,020 across 168 shards) |
+| **`checkpoint_1b_epoch3.pt`** | **4.47 GB** | ✅ **Archived & Verified** | Full Epoch 3 Foundation Pass (Step 15,030 across 168 shards) |
+| **`checkpoint_1b_latest.pt`** | **4.47 GB** | ✅ **Foundation Release** | Terminal 14-year (2010–2023) foundation weights |
+| **`checkpoint_1b_wf2024.pt`** | **4.43 GB** | ✅ **Walk-Forward 2024** | Online continual adaptation across 2024 shards (`lr=1e-5`, `0.5439%` Frobenius drift) |
+| **`checkpoint_1b_wf2025.pt`** | **4.43 GB** | ✅ **Walk-Forward 2025** | Online continual adaptation across 2024+2025 shards (`lr=1e-5`, `0.7860%` cumulative drift) |
 
 ---
 
-## 💻 Zero-Shot Inference in PyTorch
-
-```python
-import torch
-from huggingface_hub import hf_hub_download
-from model_1b import FinancialConformer1B
-from config_1b import Model1BConfig
-
-device = "cuda" if torch.cuda.is_available() else "cpu"
-
-# 1. Download model weights from Hugging Face Model Hub
-print("⚡ [ULTRON] Fetching 1.189B Model Weights...")
-weights_path = hf_hub_download(
-    repo_id="Ultron28/ohlcv-transformer",
-    filename="checkpoints_1b/checkpoint_1b_latest.pt"
-)
-
-# 2. Instantiate 1.189B Conformer Architecture
-cfg = Model1BConfig()
-model = FinancialConformer1B(cfg).to(device)
-
-# 3. Load checkpoint state
-checkpoint = torch.load(weights_path, map_location=device, weights_only=False)
-model.load_state_dict(checkpoint["model_state"])
-model.eval()
-print("✅ [ULTRON] 1.189B Conformer Foundation Model Loaded Successfully.")
+## 📐 Exact Model Architecture Audit (`1,189,135,740` Parameters)
 
-# 4. Inference on 2,048-minute candlestick window (B=1, T=2048, F=6)
-# Features: [LogReturn, NormRange, NormBody, UpperWick, LowerWick, NormVol]
-x = torch.randn(1, 2048, 6, device=device)
-
-with torch.no_grad():
-    predictions = model(x)
-    return_forecast = predictions["pred_returns"]       # Shape: [1, 7] (1m, 3m, 5m, 10m, 15m, 30m, 60m)
-    dir_probabilities = predictions["pred_directions"] # Shape: [1, 7] (Upward probability in [0, 1])
-
-print(f"1-Minute Directional Upward Probability: {dir_probabilities[0, 0].item():.2%}")
-print(f"1-Hour Directional Upward Probability:   {dir_probabilities[0, -1].item():.2%}")
-```
+| Subsystem / Module | Configuration & Shapes | Exact Parameters | % of Total |
+| :--- | :--- | :---: | :---: |
+| **Inception Conv1d Branches ($K=3,7,15,31,63,127$)** | $6 \times [\text{Conv1d}(8 \to 341, K_i, s=4) + \text{bias}]$ | `673,134` | `0.057%` |
+| **Inception Linear Projection & RMSNorm** | $\mathbf{W}_{\text{proj}} (2046 \times 2048) + \text{RMSNorm}(2048)$ | `4,192,256` | `0.353%` |
+| **20× GQA Attention (`q, k, v, out` + Norm)** | $H_q=16, H_{kv}=8, d_{\text{head}}=128$, RoPE | `251,699,200` | `21.167%` |
+| **20× Depthwise Separable Conv Module** | Pointwise GLU ($2048 \to 4096$), $\text{DWConv}_{K=7}$, GroupNorm-32 | `251,985,920` | `21.191%` |
+| **20× SwiGLU Feed-Forward Network** | $d_{\text{model}}=2048, d_{\text{ff}}=5504$ (`w_gate, w_up, w_down`) | `676,372,480` | `56.879%` |
+| **Final Backbone RMSNorm & Dual MLP Heads** | `final_norm` ($2,048$) + `return_head` & `dir_head` ($2\times 2,105,351$) | `4,212,750` | `0.354%` |
+| **Total Financial Conformer (1.189B)** | **20 Conformer Blocks, $L=2048 \to T=512$ Tokens** | **`1,189,135,740`** | **`100.000%`** |
 
 ---
 
-## 📑 Research Papers & Technical Manuscripts
+## 📑 Tier-1 Research Papers & Technical Manuscripts
 
-This research is documented across three formal academic manuscripts authored by **Suryaansh Prithvijit Singh** (Lead Researcher, Airborne AI / AirborneHRS):
+This research is documented across three formal academic manuscripts authored by **Suryaansh Prithvijit Singh** (*Lead Researcher, Airborne AI / AirborneHRS* &bull; `Suryaansh@airbornehrs.in`):
 
-| # | Paper Title | Formats | Focus Area |
+| # | Paper Title | Formats | Core Empirical & Mathematical Contributions |
 | :-: | :--- | :--- | :--- |
-| **1** | **Scaling 1.189-Billion Parameter Financial Conformers: Empirical Foundations and Multi-Horizon Directional Alpha Across 14 Years of Market Microstructure** | [📄 PDF](papers/Paper_1_Financial_Conformer_Foundation_Model.pdf) &bull; [📝 Markdown](papers/Paper_1_Financial_Conformer_Foundation_Model.md) | Foundation architecture, 14-year pre-training dynamics (168 shards), multi-horizon scaling laws, and zero-shot OOS testing on 2024–2026. |
-| **2** | **Continuous Market Adaptation Without Catastrophic Forgetting: Online Walk-Forward Learning in Multi-Billion Parameter Financial Foundation Models** | [📄 PDF](papers/Paper_2_Continual_Walk_Forward_Online_Adaptation.pdf) &bull; [📝 Markdown](papers/Paper_2_Continual_Walk_Forward_Online_Adaptation.md) | Online walk-forward adaptation (`wf2024`, `wf2025`), preserving 14-year historical priors while capturing new volatility regimes. |
-| **3** | **From Microstructure Waveforms to Executable Alpha: Mechanistic Frequency Decomposition and Transaction Friction Dynamics in Neural Forecasting** | [📄 PDF](papers/Paper_3_Mechanistic_Alpha_and_Execution_Friction.pdf) &bull; [📝 Markdown](papers/Paper_3_Mechanistic_Alpha_and_Execution_Friction.md) | Inception kernel ablation ($K \in [3, 7, 15, 31, 63, 127]$), dollar-neutral portfolio slippage stress testing (0, 2, 5, 10 bps). |
+| **1** | **Scaling 1.189-Billion Parameter Financial Conformers: Empirical Foundations, Pre-Training Dynamics, and Multi-Horizon Directional Alpha Across 14 Years of Market Microstructure** | [📄 PDF](papers/Paper_1_Financial_Conformer_Foundation_Model.pdf) &bull; [📝 Markdown](papers/Paper_1_Financial_Conformer_Foundation_Model.md) | Exact 1.189B architecture equations, 855-step pre-training telemetry ($-15.2\%$ loss variance contraction), 95,040-window OOS benchmark (2024–2026), microstructure flat-bar audit, and multi-epoch scaling vs. 4 classical baselines (30 refs). |
+| **2** | **Continuous Market Adaptation Without Catastrophic Forgetting: Online Walk-Forward Learning and Frobenius Manifold Stability in Multi-Billion Parameter Financial Foundation Models** | [📄 PDF](papers/Paper_2_Continual_Walk_Forward_Online_Adaptation.pdf) &bull; [📝 Markdown](papers/Paper_2_Continual_Walk_Forward_Online_Adaptation.md) | Trust-region walk-forward adaptation (`wf2024`, `wf2025`), exact 308-tensor Frobenius weight drift & cosine similarity audit ($0.5439\%$ global drift, $0.999920$ CosSim), plasticity hierarchy, and forward ROC-AUC/Rank IC transfer (23 refs). |
+| **3** | **From Microstructure Waveforms to Executable Alpha: Mechanistic Receptive-Field Decomposition and Causal Transaction Friction Dynamics in Billion-Parameter Neural Forecasting** | [📄 PDF](papers/Paper_3_Mechanistic_Alpha_and_Execution_Friction.pdf) &bull; [📝 Markdown](papers/Paper_3_Mechanistic_Alpha_and_Execution_Friction.md) | 7-horizon continuous $\ell_2$ & discrete threshold kernel ablation ($K \in \{3..127\}$), Frequency-Horizon Specialization Law, and strictly causal chronological execution backtest with rolling $z$-score confidence hurdles across $0, 1, 2, 5\text{ bps}$ slippage (22 refs). |
 
 <div align="center">
-  <img src="papers/figures/fig2_oos_directional_accuracy.png" alt="OOS Directional Accuracy" width="85%"/>
+  <img src="papers/figures/fig2_oos_directional_accuracy.png" alt="OOS Directional Accuracy and Baselines" width="95%"/>
 </div>
 
 ---
 
-## 🏆 Out-of-Sample Empirical Benchmark (2024 – 2026)
+## 🏆 Out-of-Sample Empirical Benchmark & Scientific Audit (2024 – 2026)
+
+All empirical metrics are 100% deterministic and reproducible via [`logs/oos_evaluation_results.json`](logs/oos_evaluation_results.json), [`logs/pretraining_telemetry_verified.json`](logs/pretraining_telemetry_verified.json), and [`logs/tier1_verified_audit.json`](logs/tier1_verified_audit.json) (generated by [`scripts/tier1_scientific_audit.py`](scripts/tier1_scientific_audit.py)).
 
-Evaluated across **27 strictly held-out out-of-sample monthly Parquet shards** (95,040 independent test windows):
+### 1. 27-Shard Zero-Shot Out-of-Sample Benchmark (95,040 Test Windows)
+| Forecast Horizon | 2024 Dir Acc (%) | 2024 Rank IC | 2025 Dir Acc (%) | 2025 Rank IC | 2026 Dir Acc (%) | 2026 Rank IC | 3-Year Weighted Acc |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| **1-Minute ($\tau=1$)** | **54.22%** | $+0.0106$ | **54.30%** | $+0.0046$ | **53.90%** | $+0.0074$ | **54.22%** |
+| **3-Minute ($\tau=3$)** | **52.15%** | $+0.0065$ | **51.65%** | $-0.0003$ | **51.91%** | $+0.0108$ | **51.90%** |
+| **5-Minute ($\tau=5$)** | **51.39%** | $+0.0144$ | **51.09%** | $-0.0005$ | **51.39%** | $-0.0012$ | **51.26%** |
+| **15-Minute ($\tau=15$)** | **50.52%** | $+0.0148$ | $49.81\%$ | $+0.0057$ | **50.12%** | $+0.0014$ | **50.16%** |
+| **30-Minute ($\tau=30$)** | **51.17%** | $+0.0122$ | **50.06%** | $+0.0120$ | $48.21\%$ | $-0.0009$ | **50.35%** |
+| **60-Minute ($\tau=60$)** | **51.49%** | $+0.0228$ | **50.73%** | $+0.0241$ | $47.96\%$ | $+0.0053$ | **50.76%** |
+| **1-Day ($\tau=1440$)** | **53.00%** | $+0.0269$ | **52.10%** | $+0.0198$ | $47.71\%$ | $+0.0028$ | **52.01%** |
+| **7-Horizon Mean** | **51.99%** | **$+0.0155$** | **51.39%** | **$+0.0093$** | **50.17%** | **$+0.0037$** | **51.52%** |
 
-### 1. Static Zero-Shot Directional Accuracy & Rank IC
-| Forecast Horizon | 2024 Held-Out Acc | 2024 Rank IC | 2025 Held-Out Acc | 2025 Rank IC | 2026 Frontier Acc | 2026 Rank IC |
-| :--- | :--- | :--- | :--- | :--- | :--- | :--- |
-| **1-Minute** | **54.22%** | +0.0106 | **54.30%** | +0.0046 | **53.90%** | +0.0074 |
-| **3-Minute** | 52.15% | +0.0065 | 51.65% | -0.0003 | 51.91% | +0.0108 |
-| **5-Minute** | 51.39% | +0.0144 | 51.09% | -0.0005 | 51.39% | -0.0012 |
-| **15-Minute** | 50.52% | +0.0148 | 49.81% | +0.0057 | 50.12% | +0.0014 |
-| **30-Minute** | 51.17% | +0.0122 | 50.06% | +0.0120 | 48.21% | -0.0009 |
-| **60-Minute** | 51.49% | +0.0228 | 50.73% | +0.0241 | 47.96% | +0.0053 |
-| **1-Day** | 53.00% | +0.0269 | 52.10% | +0.0198 | 47.71% | +0.0028 |
-| **Overall Average**| **51.99%** | — | **51.39%** | — | **50.17%** | — |
+### 2. Exact Frobenius Weight Drift & Online Walk-Forward Adaptation
+- **Frobenius Plasticity Hierarchy (`Ep3 → WF-2024`)**: Global Frobenius drift is **`0.5439%`** (Cosine Similarity **`0.999920`**), with the 20-layer Conformer backbone shifting by only **`0.5291%`** while the Inception Tokenizer (**`1.1773%`**) and Directional Head (**`1.0087%`**) recalibrate to current regimes.
+- **Forward ROC-AUC & Rank IC Expansion**: `WF-2024` improves 2025 OOS Strictly Non-Zero Accuracy (`51.27% → 51.32%`), Balanced Accuracy (`50.26% → 50.31%`), and ROC-AUC (**`0.5125 → 0.5154`**), while `WF-2025` elevates 2025 ROC-AUC to **`0.5167`** (`60m` Rank IC **`+0.0542`**, `1d` Rank IC **`+0.0826`**).
+
+<div align="center">
+  <img src="papers/figures/fig3_walk_forward_comparison.png" alt="Frobenius Drift and Walk-Forward ROC-AUC" width="95%"/>
+</div>
 
-### 2. Continual Walk-Forward Learning Retention
-- **Online Adaptation on 2024 (`checkpoint_1b_wf2024.pt`) Tested on 2025**: **51.40%** directional accuracy (strictly exceeds zero-shot baseline of 51.39% with zero catastrophic forgetting).
-- **Online Adaptation on 2025 (`checkpoint_1b_wf2025.pt`) Tested on 2026**: **53.90%** 1-minute accuracy, **+0.56** gross Sharpe ratio.
+### 3. Mechanistic Kernel Ablation & Strictly Causal Execution Backtest
+- **Frequency-Horizon Specialization**: Masking short receptive fields (`K=7` or `Micro K=3,7`) induces **`18.23%–22.33%`** mean continuous logit $\ell_2$ perturbation and **`49.53%–63.60%`** return forecast $\ell_2$ perturbation (peaking at **`36.11%`** on 15m logits while leaving 1d logits nearly invariant at **`5.54%`**). Masking macro `K=127` drives the largest single-branch accuracy drop (**`-0.11%`**), and masking `Meso K=15,31` drives the largest ROC-AUC drop (**`-0.0019`**).
+- **Post-Friction Net Sharpe Survival Under Causal Confidence Filtering**:
+  - **2024 `5m` ($\tau_{\text{conf}}=2.0\sigma$, $99.3\%$ turnover saved)**: **`+6.24 bps/trade`** gross return (breakeven **`3.12 bps`**), Net Sharpe **`+4.18` (0 bps), `+2.84` (1 bps), `+1.50` (2 bps)**.
+  - **2025 `60m` ($\tau_{\text{conf}}=1.5\sigma$, $94.4\%$ turnover saved)**: **`+5.14 bps/trade`** gross return (breakeven **`2.57 bps`**), Net Sharpe **`+0.80` (0 bps), `+0.49` (1 bps), `+0.18` (2 bps)**.
+  - **2025 `1d` ($\tau_{\text{conf}}=0.0\sigma$ unfiltered / $1.0\sigma$ filtered)**: Unfiltered **`+10.67 bps/trade`** (breakeven **`5.34 bps`**), Net Sharpe **`+1.11` (0 bps), `+0.90` (1 bps), `+0.69` (2 bps), `+0.07` (5 bps)**; at $\tau_{\text{conf}}=1.0\sigma$, gross return jumps to **`+19.01 bps/trade`** (breakeven **`9.51 bps`**).
+  - **2026 `15m` ($\tau_{\text{conf}}=2.0\sigma$, $98.9\%$ turnover saved)**: **`+4.63 bps/trade`** gross return (breakeven **`2.31 bps`**), Net Sharpe **`+0.91` (0 bps), `+0.52` (1 bps), `+0.12` (2 bps)**.
 
 <div align="center">
-  <img src="papers/figures/fig4_kernel_ablation.png" alt="Inception Kernel Ablation" width="85%"/>
+  <img src="papers/figures/fig5_slippage_friction_curve.png" alt="Causal Confidence Hurdle and Post-Friction Net Sharpe" width="95%"/>
 </div>
 
 ---
@@ -241,13 +161,13 @@ Evaluated across **27 strictly held-out out-of-sample monthly Parquet shards** (
 If you utilize this foundation model, dataset pipelines, or academic papers in your quantitative research or trading algorithms, please cite:
 
 ```bibtex
-@misc{singh2026ultronconformer,
+@article{singh2026financialconformer,
   author       = {Suryaansh Prithvijit Singh},
-  title        = {ULTRON: A 1.189 Billion Parameter Multi-Scale Financial Conformer for Deep Microstructure Forecasting Across 14 Years of US Equities},
+  title        = {Scaling 1.189-Billion Parameter Financial Conformers: Empirical Foundations, Pre-Training Dynamics, and Multi-Horizon Directional Alpha Across 14 Years of Market Microstructure},
   year         = {2026},
-  publisher    = {GitHub & Hugging Face},
-  journal      = {Hugging Face Model Hub},
-  howpublished = {\url{https://github.com/Ultron09/ohlcv-transformer}}
+  institution  = {Airborne AI / AirborneHRS},
+  note         = {Lead Researcher, Suryaansh@airbornehrs.in},
+  url          = {https://github.com/Ultron09/ohlcv-transformer}
 }
 ```
 
@@ -255,7 +175,7 @@ If you utilize this foundation model, dataset pipelines, or academic papers in y
 
 <div align="center">
 
-**Architected by Suryaansh Prithvijit Singh // ULTRON Cybernetic Operating Systems**  
-*Connect: [GitHub (@Ultron09)](https://github.com/Ultron09) • [Hugging Face (@Ultron28)](https://huggingface.co/Ultron28) • `connect.singha@gmail.com`*
+**Architected by Suryaansh Prithvijit Singh // Lead Researcher, Airborne AI / AirborneHRS**  
+*Connect: [GitHub (@Ultron09)](https://github.com/Ultron09) • [Hugging Face (@Ultron28)](https://huggingface.co/Ultron28) • `Suryaansh@airbornehrs.in`*
 
 </div>