🏆 Generated Paper Showcase

From a research proposal to a publication-format PDF — end-to-end, machine-checked integrity throughout.

--- Below are **seven papers** generated **end-to-end by spark-to-paper-skills** — each starting from a research proposal. The pipeline autonomously planned the paper, searched and verified real literature, drafted all sections, refined and peer-reviewed the draft, generated editable vector figures, and compiled the final PDF. > 📌 **Six domains** — environmental monitoring, energy forecasting, environmental AI, computer vision / agriculture, clinical AI, and bearing fault diagnosis — demonstrating the pipeline's cross-domain generality. --- ## 🔄 How It Works

**💡**
**Proposal**

➜

**🗺️**
**Plan**
_{blueprint.json}

➜

**📚**
**Cite**
_{≥40 real refs}

➜

**📝**
**Write**
_{all sections}

➜

**⚔️**
**Review**
_adversarial

➜

**🖼️**
**Figures**
_{editable vector}

➜

**📄**
**PDF**
_{compiled paper}

_{Each run traverses 7 focused stages with self-review, adversarial peer-review hardening, vision-critiqued figures, and deterministic quality gates — no fabricated numbers, every citation verified.}

--- ### 📄 Paper I · Bearing Fault Diagnosis

> **Segment-Overlap Leakage in CWRU Bearing Fault Diagnosis with a Compact 1D-CNN**

_{👆 Click to read the full paper}

#### 💡 Idea Bearing-fault diagnosis papers often split heavily overlapping CWRU vibration windows at random, putting near-duplicate segments on both sides of the train-test boundary. This paper holds the model and features fixed while comparing naive random, recording-aware, and cross-load protocols. #### ⚙️ Pipeline Journey | | | |:---|:---| | 🔗 **Stages** | 7 stages (plan → cite → write → refine → review → figure → compile) | | 📚 **References** | 50 cited (all verified via WebSearch + Crossref) | | 📊 **Figures** | 6 auto-generated (editable vector PDFs) | | 📋 **Tables** | 4 result tables | | 📑 **Pages** | 13 pages (official ICML 2025 style) | #### 🎯 Key Contribution Quantifies segment-overlap leakage under controlled protocols: perfect naive accuracy drops once recordings or loads are held out, and the compact CNN no longer dominates classical baselines under honest evaluation.

--- ### 📄 Paper II · Chronic Disease Risk Screening

> **Interpretable, Calibrated Early Screening of Chronic Disease Risk with Feature Selection and Ensembles**

_{👆 Click to read the full paper}

#### 💡 Idea Clinical screening models need calibrated probabilities and defensible explanations, not only high ranking accuracy. This paper evaluates diabetes, heart-disease, and stroke cohorts with leakage-safe preprocessing, feature selection, ensemble baselines, post-hoc calibration, and TreeSHAP audits. #### ⚙️ Pipeline Journey | | | |:---|:---| | 🔗 **Stages** | 7 stages (plan → cite → write → refine → review → figure → compile) | | 📚 **References** | 42 cited (all verified via WebSearch + Crossref) | | 📊 **Figures** | 5 auto-generated (editable vector PDFs) | | 📋 **Tables** | 4 result tables | | 📑 **Pages** | 17 pages (official NeurIPS 2025 style, preprint) | #### 🎯 Key Contribution Recasts chronic-disease screening as calibrated, interpretable risk estimation — showing where feature selection preserves discrimination, recalibration repairs unreliable rare-event probabilities, and accuracy can hide clinically useless stroke screening behavior.

--- ### 📄 Paper III · Leaf Disease Leakage Audit

> **A Compact CNN for Leaf-Disease Classification under Leakage-Aware Evaluation**

_{👆 Click to read the full paper}

#### 💡 Idea PlantVillage leaf-disease results are often suspected of near-duplicate leakage and background shortcuts. This paper audits the Tomato subset with group-aware source-leaf splits, near-duplicate threshold sweeps, masked-background probes, and compact CNN, transfer, and classical baselines. #### ⚙️ Pipeline Journey | | | |:---|:---| | 🔗 **Stages** | 7 stages (plan → cite → write → refine → review → figure → compile) | | 📚 **References** | 52 cited (all verified via WebSearch + Crossref) | | 📊 **Figures** | 6 auto-generated (editable vector PDFs) | | 📋 **Tables** | 4 result tables | | 📑 **Pages** | 17 pages (official NeurIPS 2025 style, preprint) | #### 🎯 Key Contribution Separates suspected leakage from measured effects: near-duplicate leakage is negligible on this subset, while the background shortcut is real but modest, yielding a repeatable leakage-aware evaluation protocol for plant-disease recognition.

--- ### 📄 Paper IV · PM2.5 Forecasting

> **A Leakage-Free Reappraisal of Decomposition-Ensemble PM2.5 Forecasting**

_{👆 Click to read the full paper}

#### 💡 Idea Short-term PM2.5 forecasting guides public early-warning and emission-control decisions. The dominant decomposition-ensemble pipeline is widely reported as SOTA — but most published results suffer from future-information leakage in the train-test split. This paper provides a leakage-free reappraisal. #### ⚙️ Pipeline Journey | | | |:---|:---| | 🔗 **Stages** | 7 stages (plan → cite → write → refine → review → figure → compile) | | 📚 **References** | 45 cited (all verified via WebSearch + Crossref) | | 📊 **Figures** | 6 auto-generated (editable vector PDFs) | | 📋 **Tables** | 4 result tables | | 📑 **Pages** | 10 pages (Traitement du Signal format) | #### 🎯 Key Contribution A leakage-free evaluation framework demonstrating that the reported superiority of decomposition-ensemble methods shrinks dramatically under proper temporal splitting — providing corrected baselines for the forecasting community.

--- ### 📄 Paper V · Electricity-Load Forecasting

> **A Leakage-Free Reappraisal of Decomposition-Ensemble Electricity-Load Forecasting**

_{👆 Click to read the full paper}

#### 💡 Idea Short-term electricity-load forecasting guides unit commitment, demand response, and grid-security decisions. Like PM2.5, the field's decomposition-ensemble results are inflated by data leakage. This paper transfers the leakage-free framework to the energy domain. #### ⚙️ Pipeline Journey | | | |:---|:---| | 🔗 **Stages** | 7 stages (plan → cite → write → refine → review → figure → compile) | | 📚 **References** | 62 cited (all verified via WebSearch + Crossref) | | 📊 **Figures** | 6 auto-generated (editable vector PDFs) | | 📋 **Tables** | 4 result tables | | 📑 **Pages** | 12 pages (Traitement du Signal format) | #### 🎯 Key Contribution Demonstrates methodology transferability across forecasting domains — the leakage-free framework generalizes from environmental to energy time series, revealing consistent overestimation patterns in published decomposition-ensemble results.

--- ### 📄 Paper VI · Air and Water Quality Prediction

> **Interpretable Air and Water Quality Prediction with Feature Selection and Boosted Ensembles**

_{👆 Click to read the full paper}

#### 💡 Idea Environmental-quality models that screen drinking water or estimate hourly pollutant levels are increasingly accurate yet remain hard to trust — they consume every feature, expose no rationale, and ignore missing values and class imbalance. This paper combines feature selection with boosted ensembles for interpretable prediction. #### ⚙️ Pipeline Journey | | | |:---|:---| | 🔗 **Stages** | 7 stages (plan → cite → write → refine → review → figure → compile) | | 📚 **References** | 51 cited (all verified via WebSearch + Crossref) | | 📊 **Figures** | 6 auto-generated (editable vector PDFs) | | 📋 **Tables** | 4 result tables | | 📑 **Pages** | 11 pages (Traitement du Signal format) | #### 🎯 Key Contribution A dual-domain interpretable framework achieving competitive accuracy while exposing feature-level rationale for predictions — bridging the trust gap between black-box accuracy and regulatory transparency in environmental monitoring.

--- ### 📄 Paper VII · Leaf Disease Recognition

> **Sparse Evidence Pooling for Compact Convolutional Leaf-Disease Recognition**

_{👆 Click to read the full paper}

#### 💡 Idea Compact CNNs for phone-based crop disease diagnosis universally end in global average pooling — a read-out assuming the signal is spread evenly across the image. For leaf disease this assumption fails: diagnostic lesions are sparse and localized. This paper introduces sparse evidence pooling as a principled alternative. #### ⚙️ Pipeline Journey | | | |:---|:---| | 🔗 **Stages** | 7 stages (plan → cite → write → refine → review → figure → compile) | | 📚 **References** | 68 cited (all verified via WebSearch + Crossref) | | 📊 **Figures** | 5 auto-generated (editable vector PDFs) | | 📋 **Tables** | 4 result tables | | 📑 **Pages** | 12 pages (Traitement du Signal format) | #### 🎯 Key Contribution Sparse evidence pooling replaces global average pooling with a learnable, sparse read-out that focuses on diagnostically relevant regions — yielding higher accuracy on compact architectures suitable for edge deployment in agricultural settings.

--- ## 📊 Aggregate Statistics

📋 Metric	I	II	III	IV	V	VI	VII	🏆 Total
🏷️ Domain	_{Fault Dx}	_{Clinical AI}	_{CV / Agri}	_{Env. Mon.}	_Energy	_{Env. AI}	_{CV / Agri}	6 fields
📚 References	50	42	52	45	62	51	68	370 cited
📊 Figures	6	5	6	6	6	6	5	40 figs
📋 Tables	4	4	4	4	4	4	4	28 tables
📑 Pages	13	17	17	10	12	11	12	92 pages
🖼️ Vector Figures	✓	✓	✓	✓	✓	✓	✓	All editable
🔒 Integrity Check	✓	✓	✓	✓	✓	✓	✓	All passed

---

🚀 Try It Yourself

Every paper above was generated by installing the plugin and asking Claude:

```bash git clone https://github.com/Albus-White/spark-to-paper-skills.git ~/.claude/skills/spark-to-paper-skills ``` ``` Run ts-paper on this proposal. ```