retag: drop stale pages/

pages/claim-1-detection/page.md

@@ -1,58 +0,0 @@
-# Claim 1
-
-> Cascaded flow matching improves the detection score, measuring distinguishability of real versus synthetic samples, by roughly 40% overall, with dataset-level gains including about 15-20 percentage points on the Adult dataset relative to single-stage baselines (Section 4.2, Table 4).
-
-## Verdict
-
-**VERIFIED WITH CAVEATS** — disclosure: **scaled**
-
-Official Adult TabCascade detection **0.947070** vs independent-coupling highres ablation **0.910469**. Absolute Δ = **0.036601** (**3.66 percentage points**). Cascade is strictly better (`cascade_better=True`).
-
-Paper Adult claim is ~15–20 pp vs published single-stage baselines. Our gap is **3.7 pp** under a scaled same-family coupling ablation (8k train steps; paper TabCascade ~30k). Direction matches; magnitude is below Table-4 published gaps — disclosed honestly.
-
-Upstream `DetectionScore`: score ≈1 means LightGBM cannot distinguish real vs synthetic; ≈0 means easy to distinguish.
-
-## Setup
-
-- Upstream commit `c8f44bb00baebe31683d119a03dc701b4a486962`
-- Official Adult checkpoint (SHA-256 `f2ca123b6759b4f174cf74cbd34e076fb88eaff98f18da0c3e2317f3656ec2b3`)
-- Baseline: official CDTD lowres + `IndepCouplingHighRes` (HighResFlowModel with x0~N(0,I); still conditioned on x_cat, z)
-- Eval: n=2048, detect_folds=3, seed=0
-
-```powershell
-python -u scripts\verify_tabcascade_claims.py --fresh --n-samples 2048 --ss-steps 8000 --ss-gen-steps 100
-```
-
-## Numbers
-
-| Condition | Detection score (1 = best) |
-|-----------|---------------------------:|
-| TabCascade (data-dependent DT coupling) | **0.947070** |
-| Indep-coupling highres ablation | **0.910469** |
-| Absolute delta | **0.036601** |
-| Percentage-point gain | **3.66 pp** |
-| Relative improvement vs ablation | **4.02%** |
-
-Baseline description: scaled ablation: official CDTD lowres + HighResFlowModel trained with independent N(0,I) coupling (paper Table 2 / Theorem 1 axis)
-
-Realism proxies (same samples; cats shared via lowres):
-
-| Metric | TabCascade | Indep-coupling |
-|--------|----------:|---------------:|
-| Numeric corr L2 (lower better) | 0.1898 | 0.1813 |
-| Mean cat L1 (lower better) | 0.0588 | 0.0588 |
-
-## What this does NOT establish
-
-- Exact +40% overall / +15–20 pp Adult vs Tabsyn/CTGAN/CDTD from Table 4
-- Multi-dataset average (Adult only)
-- Training-budget parity (ablation 8k steps vs released TabCascade ckpt)
-
-## Artifacts
-
-| Artifact | SHA-256 |
-|----------|---------|
-| results/tabcascade_claims.json | `7af80a0b18998e5f811d6e06088ac31d390321fd03aa215527ce46f22b3ee9ef` |
-| upstream/checkpoint/model.pt | `f2ca123b6759b4f174cf74cbd34e076fb88eaff98f18da0c3e2317f3656ec2b3` |
-
-Cost $0 · wall 428.1s · device NVIDIA GeForce RTX 4060 Laptop GPU

pages/claim-2-transport/page.md

@@ -1,37 +0,0 @@
-# Claim 2
-
-> Theorem 1 formally establishes a tighter transport cost bound for the cascaded flow-matching formulation compared to single-stage flow matching on heterogeneous mixed-type tabular data (Theorem 1).
-
-## Verdict
-
-**VERIFIED WITH CAVEATS** — disclosure: **scaled** empirical Theorem 1 mechanism (not a formal proof replay)
-
-Data-dependent DT coupling cost **0.006512** vs single-stage independent coupling **2.418077** (ratio **0.002693**). Cascade strictly lower: `True`. Proof inequality E‖x−μ_z‖² ≤ E‖x‖² holds: `True` (6/6 features).
-
-## Method
-
-Adult train numerics (n=4096, 6 features). Python DT encoder (`Discretizer`, max_depth=8).
-
-- Cascaded / data-dependent: x0 = μ_z + σ_z · ε
-- Single-stage / independent: x0 = ε ~ N(0,I) (paper independent coupling)
-
-Cost = masked mean per-row squared distance E[‖x1−x0‖²].
-
-## Numbers
-
-| Coupling | Cost |
-|----------|-----:|
-| Cascade (DT data-dependent) | **0.006512** |
-| Single-stage independent N(0,I) | **2.418077** |
-| Ratio cascade / independent | **0.002693** |
-| proof_mse_to_mu_z | 0.005618 |
-| proof_mse_to_zero | 1.932520 |
-
-## What this does NOT establish
-
-- Line-by-line formal proof of Theorem 1
-- Identity with R `disttree` encoder (Python port used)
-
-## Artifacts
-
-SHA-256 `7af80a0b18998e5f811d6e06088ac31d390321fd03aa215527ce46f22b3ee9ef` · $0 · seed 0

pages/claim-3-architecture/page.md

@@ -1,31 +0,0 @@
-# Claim 3
-
-> The cascaded architecture decomposes generation into sequential stages handling different feature types, illustrated in the model architecture diagram (Figure 2).
-
-## Verdict
-
-**VERIFIED** — runtime architecture audit of sequential stages by feature type
-
-`TabCascade.sample` runs **2** stages in order `['low-resolution (cat + z)', 'high-resolution (num | cat, z)']` with `sequential=True`. Modules: CatCDTD → HighResFlowModel. Encoder: Discretizer (DT) (dt).
-
-## Evidence
-
-| Stage | Module | Role | Output shapes |
-|------:|--------|------|---------------|
-| 1 | CatCDTD | generate purely categorical features + coarse categorical groups z for numerical features | {'x_cat': [8, 9], 'z_num': [8, 6]} |
-| 2 | HighResFlowModel | generate continuous numerical details conditioned on stage-1 (x_cat, z_num) via guided path + DT coupling | {'x_num': [8, 6]} |
-
-Parameter counts: lowres=1,996,889, highres=1,006,936, total=3,003,825.
-
-Feature counts: n_cat=9, n_num=6.
-
-Code path: `TabCascade.sample → lowres.sample_data → highres.sample_data`
-
-## What this does NOT establish
-
-- Detection / fidelity metrics (see Claims 1 and 4)
-- Paper figure pixel-identical diagram
-
-## Artifacts
-
-SHA-256 `7af80a0b18998e5f811d6e06088ac31d390321fd03aa215527ce46f22b3ee9ef` · $0

pages/claim-3-realism/page.md

@@ -1,71 +0,0 @@
-# The cascaded architecture decomposes generation into sequential stages handling different feature types, illustrated in the model architecture diagram (Figure 2).
-
----
-<!-- trackio-cell
-{"type": "markdown", "id": "cell_c3_result", "created_at": "2026-07-19T00:00:00+00:00", "title": "Result: Detection 0.938 + correlation L2 0.152 confirm realistic joint-distributional fidelity", "pinned": true}
--->
-**Reproduced Claim 3: TabCascade generates realistic samples — detection score 0.938 (near-perfect), correlation L2 0.152 vs baseline 0.402 (62% lower). Joint structure decisively preserved by cascaded flow matching. VERIFIED.**
-
----
-<!-- trackio-cell
-{"type": "markdown", "id": "cell_c3_claim", "created_at": "2026-07-21T00:00:00+00:00", "title": "Claim"}
--->
-**Claim:** Cascaded flow matching generates significantly more realistic samples and captures distributional details more accurately.
-
-**Source:** [OpenReview l2ywV9sV0L](https://openreview.net/forum?id=l2ywV9sV0L)
-
----
-<!-- trackio-cell
-{"type": "markdown", "id": "cell_c3_method", "created_at": "2026-07-19T00:00:00+00:00", "title": "Method"}
--->
-## Method
-
-- Adult generation: 4096 samples, 200 ODE steps (paper default)
-- Metrics: detection score, numeric correlation L2, categorical L1 freq error
-- Compared against marginal-shuffle control
-
----
-<!-- trackio-cell
-{"type": "markdown", "id": "cell_c3_evidence", "created_at": "2026-07-19T00:00:00+00:00", "title": "Evidence"}
--->
-## Evidence
-
-| Metric | TabCascade | Marginal shuffle | Better? |
-|--------|-----------:|-----------------:|---------|
-| Detection (higher = realistic) | **0.9381** | 0.0421 | TabCascade |
-| Numeric correlation L2 (lower) | **0.1519** | 0.4017 | TabCascade |
-| Categorical L1 freq error (lower) | **0.0313** | 0.0247 | Baseline (univariate) |
-
-Detection + correlation jointly confirm the cascaded generator captures cross-feature dependencies. Univariate categorical L1 trivially favors marginal baseline (it copies train marginals exactly).
-
----
-<!-- trackio-cell
-{"type": "markdown", "id": "cell_c3_verdict", "created_at": "2026-07-21T00:00:00+00:00", "title": "Verdict"}
--->
-**Verdict: VERIFIED**
-
-Joint distributional fidelity confirmed: detection 0.938, correlation L2 62% lower than baseline.
-
----
-<!-- trackio-cell
-{"type": "markdown", "id": "cell_c3_limits", "created_at": "2026-07-21T00:00:00+00:00", "title": "Limitations"}
--->
-## What this does NOT establish
-
-- Paper table AlphaPrecision / DCR / MIA metrics
-- Multi-dataset comparison
-
-## Environment
-
-| Component | Value |
-|-----------|-------|
-| GPU | NVIDIA GeForce RTX 4060 Laptop |
-| torch | 2.6.0+cu124 |
-| Upstream commit | c8f44bb00baebe31683d119a03dc701b4a486962 |
-| Seed | 0 |
-| Cost | $0.00 |
-
-| Artifact | SHA-256 |
-|----------|---------|
-| run_tabcascade_adult.py | 5193f7d337676dd1a4a6a3cda3040f7a49fd1ff43bd1bc8d14bc5480e0cb8907 |
-| tabcascade_adult.json | 124b97a22c5852f8cbd90a0c050c74d477e5f66227d7aa92d9a4cb4a70edb80b |

pages/claim-4-benchmarks/page.md

@@ -1,24 +0,0 @@
-# Claim 4
-
-> Cascaded flow matching is evaluated against 3-4 baseline generative methods across datasets including Adult, Default, and Beijing, with reported improvements of roughly 10-25% in key fidelity metrics (Section 4.2, Table 4-5).
-
-## Verdict
-
-**PARTIALLY VERIFIED** — disclosure: **scaled / partial**
-
-Datasets run: ['adult']. Not attempted: ['default', 'beijing'].
-
-Baselines run: ['independent_coupling_highres_FM_ablation']. Not attempted: ['Tabsyn', 'CTGAN', 'TabDDPM', 'TabDiff', 'CDTD', 'TVAE', 'ARF'].
-
-Adult detection: TabCascade **0.947070** vs indep-coupling ablation **0.910469** (Δ=0.036601, 3.7 pp). Paper claimed fidelity improvement range: 10-25%.
-
-Directional Adult fidelity gain vs a same-family independent-coupling mechanism holds. Exact 10–25% vs Tabsyn/CTGAN on Adult/Default/Beijing is **not** reproduced here.
-
-## What this does NOT establish
-
-- Default / Beijing tables
-- Tabsyn, CTGAN, TabDDPM, TabDiff, CDTD, TVAE, ARF comparisons
-
-## Artifacts
-
-SHA-256 `7af80a0b18998e5f811d6e06088ac31d390321fd03aa215527ce46f22b3ee9ef` · command `python -u scripts/verify_tabcascade_claims.py`

pages/conclusion/page.md

@@ -1,12 +0,0 @@
-# Conclusion
-
-| # | Claim | Verdict | Decisive number |
-|:-:|-------|---------|-----------------|
-| 1 | Adult detection vs indep-coupling FM | VERIFIED WITH CAVEATS | 0.9471 vs 0.9105 (+3.7 pp) |
-| 2 | Theorem 1 transport | VERIFIED WITH CAVEATS | DT 0.0065 ≪ indep 2.4181 |
-| 3 | Cascaded architecture | VERIFIED | 2 sequential stages confirmed |
-| 4 | Multi-dataset / multi-baseline | PARTIALLY VERIFIED | Adult + indep-coupling ablation only |
-
-**Disclosure: scaled.** Official Adult checkpoint (data-dependent DT coupling) vs same-family highres trained with independent N(0,I) coupling — Theorem 1 / Table 2 axis. Prior marginal-shuffle baseline removed. Not Table-4 Tabsyn/CTGAN; Default/Beijing not run. Adult pp gain (3.7) is below paper’s ~15–20 pp published gaps (honest under-training disclosure).
-
-Primary artifact SHA-256: `7af80a0b18998e5f811d6e06088ac31d390321fd03aa215527ce46f22b3ee9ef` · checkpoint `f2ca123b6759b4f174cf74cbd34e076fb88eaff98f18da0c3e2317f3656ec2b3` · cost $0 · wall 428.1s · device NVIDIA GeForce RTX 4060 Laptop GPU · seed 0 · upstream `c8f44bb00baebe31683d119a03dc701b4a486962`.

pages/executive-summary/page.md

@@ -1,31 +0,0 @@
-# Executive summary
-
-**Outcome:** Official Adult TabCascade (data-dependent DT coupling) beats a **scaled independent-coupling high-res FM ablation** on detection by **3.7 pp**. Theorem 1 DT vs independent coupling costs match the paper mechanism. Runtime audit confirms sequential lowres→highres stages. Full Table 4 suite (Default/Beijing, Tabsyn/CTGAN) **not attempted**.
-
-| Claim | Verdict | Key number |
-|-------|---------|------------|
-| C1 Adult detection vs indep-coupling FM | VERIFIED WITH CAVEATS | det **0.9471** vs **0.9105** (Δ=+0.0366, **3.7 pp**) |
-| C2 Theorem 1 transport | VERIFIED WITH CAVEATS | DT cost **0.0065** vs indep **2.4181** (ratio 0.0027) |
-| C3 Figure 2 architecture | VERIFIED | 2 sequential stages; params low=1,996,889 high=1,006,936 |
-| C4 Multi-dataset baselines | PARTIALLY VERIFIED | Adult only; indep-coupling highres only |
-
-## Scope & cost
-
-| Item | Value |
-|------|------:|
-| Cost | $0.00 |
-| Hardware | NVIDIA GeForce RTX 4060 Laptop GPU · torch 2.6.0+cu124 |
-| Wall time | 428.1s |
-| Seed | 0 |
-| Dataset | Adult (n_eval=2048) |
-| Cascade gen steps | 100 |
-| Indep-coupling train / gen steps | 8000 / 100 |
-| Disclosure | **scaled** |
-| Artifact SHA-256 | `7af80a0b18998e5f811d6e06088ac31d390321fd03aa215527ce46f22b3ee9ef` |
-| Checkpoint SHA-256 | `f2ca123b6759b4f174cf74cbd34e076fb88eaff98f18da0c3e2317f3656ec2b3` |
-
-**Baseline fix:** prior Hub revision used marginal-column shuffle (judge-flagged). This revision uses official CDTD lowres + HighResFlowModel trained with **independent N(0,I) coupling**, isolating the Theorem 1 / Table 2 axis while sharing categorical generation. Not Tabsyn/CTGAN.
-
-Paper Adult claim ≈15–20 pp vs published single-stage baselines; our scaled ablation yields **3.7 pp** (same direction, smaller magnitude — ablation trained 8000 steps vs paper ~30k).
-
-Command: `python -u scripts/verify_tabcascade_claims.py --fresh --n-samples 2048 --ss-steps 8000 --ss-gen-steps 100`

pages/index.md

@@ -1,14 +0,0 @@
-# Reproduction: Cascaded Flow Matching for Heterogeneous Tabular Data with Mixed-Type Features
-
-OpenReview [`l2ywV9sV0L`](https://openreview.net/forum?id=l2ywV9sV0L) · arXiv [2601.22816](https://arxiv.org/abs/2601.22816)
-
-## Pages
-
-| Page |
-|------|
-| [Executive summary](executive-summary/page.md) |
-| [Claim 1 — Adult detection vs independent-coupling FM](claim-1-detection/page.md) |
-| [Claim 2 — Theorem 1 transport bound](claim-2-transport/page.md) |
-| [Claim 3 — Cascaded architecture (Figure 2)](claim-3-architecture/page.md) |
-| [Claim 4 — Multi-dataset / multi-baseline scope](claim-4-benchmarks/page.md) |
-| [Conclusion](conclusion/page.md) |