*data-to-paper*

Orchestrates a multi-stage pipeline that transforms raw experimental data into complete research papers by chaining LLM calls for data analysis, insight extraction, narrative generation, and formatting. The system maintains semantic coherence across stages through intermediate representations (structured findings, outline templates, citation graphs) rather than naive sequential prompting, enabling papers to reflect actual data patterns rather than hallucinated results.

Analyzes structured datasets to automatically identify statistically significant patterns, anomalies, and relationships, then generates research hypotheses grounded in those patterns. The system performs statistical validation (significance testing, effect size calculation) before proposing insights, preventing the LLM from inventing findings that don't exist in the data.

Chains multiple specialized LLM prompts (abstract generation, introduction framing, results narration, discussion synthesis) while maintaining semantic consistency across sections through shared context vectors and cross-reference validation. Each stage receives not just raw data but also outputs from prior stages, enabling the discussion section to directly reference findings and the introduction to foreshadow results.

Automatically generates citations for claims made in the paper by mapping assertions back to the source data or external knowledge bases, then formats citations in standard styles (APA, IEEE, Chicago). The system validates that cited works actually support the claims made, preventing fabricated or misattributed references.

Accepts human feedback on generated paper sections (e.g., 'this claim needs more evidence', 'this section is unclear') and automatically regenerates affected sections while preserving coherence with unchanged sections. Uses feedback embeddings to identify which parts of the generation pipeline need adjustment and re-runs only those stages rather than regenerating the entire paper.

Applies domain-specific formatting rules, section structures, and style guidelines to generated papers, ensuring output matches the conventions of target journals or conferences. Templates define required sections, citation styles, figure/table placement rules, and language constraints (e.g., passive voice for methods sections), which are enforced during generation through prompt engineering and post-generation validation.

Orchestrates paper generation from multiple related datasets, identifying connections between datasets and synthesizing findings across them. The system detects overlapping variables, temporal relationships, and causal links between datasets, then generates a unified narrative that treats the datasets as complementary evidence rather than separate analyses.

Automatically generates visualizations (plots, charts, tables) from raw data and creates natural language captions that describe the visualizations and their significance. The system selects appropriate visualization types based on data characteristics, generates publication-quality figures, and writes captions that explain what the figure shows and why it matters for the paper's narrative.