Update docs with latest architecture and results

docs/README.md

@@ -16,3 +16,8 @@ Tools:
 - `example/run_all_full.py` for one-command full pipeline + diagnostics.
 Notes:
 - If `use_quantile_transform` is enabled, run `prepare_data.py` with `full_stats: true` to build quantile tables.
+
+Current status (high level):
+- Two-stage pipeline (GRU trend + diffusion residuals).
+- Quantile transform + post-hoc calibration enabled for continuous features.
+- Latest metrics (2026-01-27 21:22): avg_ks ~0.405 / avg_jsd ~0.038 / avg_lag1_diff ~0.145.

docs/decisions.md

@@ -71,3 +71,10 @@
   - `example/export_samples.py`
   - `example/prepare_data.py`
   - `example/config.json`
+
+## 2026-01-27 — Full quantile stats in preparation
+- **Decision**: Enable full statistics when quantile transform is active.
+- **Why**: Stabilize quantile tables and reduce CDF mismatch.
+- **Files**:
+  - `example/prepare_data.py`
+  - `example/config.json`

docs/experiments.md

@@ -27,3 +27,8 @@ YYYY-MM-DD
 - Config: `example/config.json` (two-stage residual diffusion; user run on Windows)
 - Result: 0.7096230 / 0.0331810 / 0.1898416
 - Notes: slight KS improvement, lag-1 improves; still distribution/temporal trade-off.
+
+## 2026-01-27
+- Config: `example/config.json` (quantile transform + calibration, full stats)
+- Result: 0.4046 / 0.0376 / 0.1449
+- Notes: KS and lag-1 improved significantly; JSD regressed vs best discrete run.

docs/ideas.md

@@ -11,3 +11,6 @@
 
 ## Two-stage training with curriculum
 - Hypothesis: train diffusion on residuals only after temporal GRU converges to low error.
+
+## Discrete calibration
+- Hypothesis: post-hoc calibration on discrete marginals can reduce JSD without harming KS.

report.md

@@ -34,6 +34,10 @@ One command pipeline:
 ```
 python example/run_all.py --device cuda
 ```
+Full pipeline + diagnostics:
+```
+python example/run_all_full.py --device cuda
+```
 
 Pipeline stages:
 1) **Prepare data** (`example/prepare_data.py`)
@@ -63,6 +67,11 @@ Defined in `example/hybrid_diffusion.py`.
 - Transformer encoder (self‑attention)
 - Post LayerNorm + residual MLP
 
+**Current default config (latest):**
+- Backbone: Transformer
+- Sequence length: 96
+- Batch size: 16
+
 **Outputs:**
 - Continuous head: predicts target (`eps` or `x0`)
 - Discrete heads: logits per discrete column
@@ -83,6 +92,8 @@ trend = GRU(x)
 residual = x - trend
 ```
 
+**Two-stage training:** temporal GRU first, diffusion on residuals.
+
 ---
 
 ## 5. Diffusion Formulations / 扩散形式
@@ -145,6 +156,7 @@ Key steps:
 - Streaming mean/std/min/max + int-like detection
 - Optional **log1p transform** for heavy-tailed continuous columns
 - Optional **quantile transform** (TabDDPM-style) for continuous columns (skips extra standardization)
+- **Full quantile stats** (full_stats) for stable calibration
 - Optional **post-hoc quantile calibration** to align 1D CDFs after sampling
 - Discrete vocab + most frequent token
 - Windowed batching with **shuffle buffer**
@@ -189,8 +201,8 @@ Metrics (with reference):
 - 输出 `example/results/cdf_<feature>.svg`（真实 vs 生成 CDF）
 - 统计生成数据是否堆积在边界（gen_frac_at_min / gen_frac_at_max）
 
-Recent run (user-reported, Windows):
-- avg_ks 0.7096 / avg_jsd 0.03318 / avg_lag1_diff 0.18984
+Recent runs (Windows):
+- 2026-01-27 21:22:34 — avg_ks 0.4046 / avg_jsd 0.0376 / avg_lag1_diff 0.1449
 
 ---
 
@@ -226,9 +238,9 @@ Recent run (user-reported, Windows):
 ---
 
 ## 13. Known Issues / Current Limitations / 已知问题
-- KS may remain high → continuous distribution mismatch
+- KS can remain high on a subset of features → per-feature diagnosis required
 - Lag‑1 may fluctuate → distribution vs temporal trade-off
-- Continuous loss may dominate → needs careful weighting
+- Discrete JSD can regress when continuous KS is prioritized
 - Transformer backbone may change stability; needs systematic comparison
 
 ---
@@ -237,6 +249,7 @@ Recent run (user-reported, Windows):
 - Compare GRU vs Transformer backbone using `run_compare.py`
 - Explore **v‑prediction** for continuous branch
 - Strengthen discrete diffusion (e.g., D3PM-style transitions)
+- Add targeted discrete calibration for high‑JSD columns
 
 ---