Upload papers, extract their methods with page-level provenance, compare studies honestly, turn evidence-backed gaps into versioned algorithm specifications, and test them against fair baselines in isolated workers — then draft a report from the results you actually measured.
Structured extraction of problems, contributions, datasets, models, hyperparameters, metrics, results and limitations. Every claim keeps its page and passage, and you confirm, correct or reject it.
Side-by-side method matrices with comparability levels. Reported numbers are ranked only for the same dataset, metric and protocol — and are always marked as author-reported.
Fifteen-field gap reports with supporting and contradicting passages, confidence explanations and an explicit scope: reviewed papers only, or checked against OpenAlex.
Versioned specifications covering formulation, equations, architecture, pseudocode, complexity, prior art, hypotheses and failure conditions — with an advisory novelty check.
Leakage-safe pipelines, task-aware splits, repeated seeds, ablations and robustness, run in sandboxed worker processes with timeouts and recorded environments.
Drafts built only from stored results and verified references, with placeholders and unverified claims highlighted. Export to DOCX, PDF, Markdown and LaTeX with BibTeX.
The platform is built to keep machine assistance honest and every number traceable.