Explore

Operate ScienceClaw's program-level self-evolution - turn a finished, replay-verified canvas session into a Skill/Operator candidate, gate it on user-registered validation tasks, and leave promotion (and rollback) to the user, with receipts.

Identify research gaps, synthesize cross-disciplinary insights, and generate novel hypotheses. Use when: user asks about unexplored areas, cross-field connections, or new research directions. NOT for: routine literature review or data analysis.

Classify scientific content by discipline, methodology, topic, and quality. Use when: user asks to categorize papers, methods, or research outputs. NOT for: simple keyword tagging or non-scientific content.

Solve a data-analysis or scientific-computing request as a typed, executable workflow graph on the ScienceClaw canvas, using the scientific tool library and the evolved Skill/Operator program. Use for any task that has input files and a checkable deliverable; answer plain questions directly.

Use when a task asks to predict a phonon property of inorganic crystals from their structure, for example the frequency of the last peak of the phonon density of states in cm^-1 (Matbench phonons), given structures (lattice, species, fractional coordinates) and a labelled training set; the deliverable is a float vector in input order, judged by mean absolute error. Covers composition/geometry descriptors, log-target tree regressors and pretrained universal interatomic potentials (SevenNet, CHGNet) as phonon feature extractors. Corresponds to FoR51 of the companion ScienceClaw-Eval benchmark.

Use when a task gives arguments (conclusion, stance, premise) with a training set labelled for the 20 human-value categories of the ValueEval taxonomy (SemEval-2023 Task 4 / Touche23-ValueEval) and asks for a 0/1 multi-label matrix (arguments by 20 values), judged by the official F1 built from macro precision and macro recall. Corresponds to FoR50 of the companion ScienceClaw-Eval benchmark.

Use when a task gives SMT-LIB v2 scripts (quantifier-free nonlinear integer or integer-real arithmetic, QF_NIA / QF_NIRA, SMT-COMP style single-query instances) and asks for the satisfiability verdict of each one (sat, unsat or unknown) as a list in input order, judged by agreement with the status published with the benchmark. Corresponds to FoR49 of the companion ScienceClaw-Eval benchmark.

Use when a task asks to locate the clauses of contracts or NDAs (split into candidate spans) that support or contradict fixed natural-language hypotheses, ContractNLI-style: inputs are contract spans, hypotheses and annotated training contracts; the deliverable is, per (contract, hypothesis) pair, an Entailment or Contradiction label plus one relevance score per span, judged by mean average precision. Corresponds to FoR48 of the companion ScienceClaw-Eval benchmark.

Use when a task gives a Universal Dependencies treebank (annotated training sentences with upos, head, deprel) and tokenised target sentences, and asks for a dependency tree (head and relation per word) for every target sentence, scored by labeled attachment score LAS (CoNLL 2018 UD parsing with gold tokenization, French GSD/PUD; corresponds to FoR47 of the companion ScienceClaw-Eval benchmark).

Use when a task gives Python programming problems (a function stub with docstring or a task description, an entry-point name, a few public assert tests) and asks for one complete program per problem that must pass hidden unit tests, scored by execution pass@1 (HumanEval / MBPP style code generation; corresponds to FoR46 of the companion ScienceClaw-Eval benchmark).

Use when a task gives small culturally specific images with language metadata (Indigenous languages of the Americas) and a few captioned training examples per language, and asks for one caption string in the item's own language per image, scored by mean sentence chrF++ against the reference captions (AmericasNLP 2026 cultural image captioning; corresponds to FoR45 of the companion ScienceClaw-Eval benchmark).

Use when a task gives a covariate table plus, for each of many simulated observational datasets, a binary treatment vector and an outcome vector, and asks for one average treatment effect on the treated (SATT) per dataset without access to the counterfactuals, scored by RMSE relative to the outcome standard deviation (ACIC 2016 causal inference challenge; corresponds to FoR44 of the companion ScienceClaw-Eval benchmark).

Use when a task gives noisy OCR text of historical newspapers, periodicals or books (several languages, with metadata and a set of labelled OCR/ground-truth pairs) and asks for corrected transcriptions, one string per unit, scored by character match error rate against the gold text (HIPE-OCRepair OCR post-correction; corresponds to FoR43 of the companion ScienceClaw-Eval benchmark).

Use when a task gives hourly ICU records (vitals, labs, demographics, one stay per patient) with sepsis labels and asks for causal hour-by-hour sepsis alarms on new stays, scored by normalized clinical utility (PhysioNet/CinC 2019 early sepsis prediction; corresponds to FoR42 of the companion ScienceClaw-Eval benchmark).

Use when a task asks for 30-day-ahead probabilistic daily forecasts of dissolved oxygen (mg/L) and water temperature (degC) at freshwater monitoring sites from multi-year daily histories (NEON aquatics, neon4cast ecological forecasting), delivered as a normal predictive distribution (mu and sigma per variable and lead day) and judged by CRPS (FoR41 of the companion ScienceClaw-Eval benchmark).

Use when a task asks for unsupervised anomalous sound detection for one machine type in the first-shot setting (only normal source-domain training clips, no labelled anomalies, unlabeled test clips) and wants one anomaly score per test clip, judged by the DCASE 2024 Task 2 score (harmonic mean of source AUC, target AUC and pAUC); machine condition monitoring from audio with log-mel or pretrained AST / CLAP embeddings (FoR40 of the companion ScienceClaw-Eval benchmark).

Use when a task asks to predict whether students answer held-out diagnostic multiple-choice questions correctly, from a large student-by-question correctness matrix and a small per-student budget of adaptively chosen revealed answers (about 10), delivered as a 0/1 prediction per target cell and judged by accuracy; Eedi / NeurIPS 2020 Education Challenge Task 4 adaptive testing with item response theory (FoR39 of the companion ScienceClaw-Eval benchmark).

Use when a task asks to forecast annual country-level economic series several years ahead (for example GDP per capita levels or unemployment rates, horizon 4) for many economies, from a panel of past series and covariate indicators such as growth and inflation, delivered as level forecasts and judged by mean sMAPE (World Bank WDI-style macro panel forecasting; FoR38 of the companion ScienceClaw-Eval benchmark).

Use when a task asks for a 24-hour-ahead forecast of a global gridded surface field (2 m temperature on a 64x32 equiangular ERA5 / WeatherBench 2 grid, 6-hourly) from a training record and the four preceding fields of each initialisation, delivered in kelvin and judged by latitude-weighted RMSE (FoR37 of the companion ScienceClaw-Eval benchmark).

Use when a task asks to separate stereo music mixtures into four stems (vocals, drums, bass, other) and deliver the estimated stems, optionally with training mixtures plus reference stems, judged by BSSEval-v4 source-to-distortion ratio (SDR, dB), as in MUSDB18 music source separation (FoR36 of the companion ScienceClaw-Eval benchmark).

Use when forecasting monthly tourism, visitor or other seasonal business and service series (Monash Tourism Monthly, 24-month horizon, period 12) from each series' history, delivering non-negative forecasts per series scored by mean MASE. Corresponds to FoR35 of the companion ScienceClaw-Eval benchmark.

Use when predicting a binary molecular property such as HIV replication inhibition (OGB ogbg-molhiv, MoleculeNet-style activity) from SMILES strings and labelled molecules, delivering one probability-like score per query molecule evaluated by ROC-AUC under a scaffold split. Corresponds to FoR34 of the companion ScienceClaw-Eval benchmark.

Use when forecasting day-ahead hourly electricity load of individual buildings (residential and commercial smart-meter data, BuildingsBench-style) from the previous 168 hours and optionally the building's own earlier history, delivering a 24-hour forecast in kWh per window scored by balanced CVRMSE. Corresponds to FoR33 of the companion ScienceClaw-Eval benchmark.

Use when segmenting the hippocampus into anterior and posterior parts in small 3-D T1-weighted brain MRI crops (Medical Segmentation Decathlon Task04) from a set of labelled volumes, delivering one integer label volume per image scored by Dice. Corresponds to FoR32 of the companion ScienceClaw-Eval benchmark.

Use when ranking or predicting the fitness effect of single amino-acid substitutions in a protein (deep mutational scanning, ProteinGym-style assays) from labelled variants of the same assay and the wild-type sequence, delivering one score per query variant evaluated by per-assay Spearman correlation. Corresponds to FoR31 of the companion ScienceClaw-Eval benchmark.

Use when segmenting top-down field or UAV RGB images of crops (sugar beet, PhenoBench-style) into soil / crop / weed semantics plus crop-plant and crop-leaf instances (hierarchical panoptic segmentation), from a few labelled images, delivering per-image integer label maps scored by PQ+. Corresponds to FoR30 of the companion ScienceClaw-Eval benchmark.

Index of the 23 discipline skills (FoR30-FoR52) that mirror the disciplines of the companion ScienceClaw-Eval benchmark: find which discipline skill fits a scientific task (plant panoptic segmentation, protein fitness prediction, hippocampus segmentation, building load forecasting, molecule property prediction, music source separation, sepsis early warning, French dependency parsing, contract evidence retrieval, ...). Use to pick the right discipline skill and the scilib tools it names.

Single-cell RNA-seq analysis with scanpy and anndata. Use when: (1) scRNA-seq preprocessing and QC, (2) clustering and cell type annotation, (3) differential expression analysis, (4) trajectory and pseudotime analysis, (5) UMAP/tSNE visualization. NOT for: bulk RNA-seq differential expression (use DESeq2/edgeR), protein structure analysis (use pdb-structure), or imaging data.

Reflect on completed research tasks to improve future performance. Use when: a research task has just been completed and the agent should evaluate its own process, store lessons learned, or retrieve past reflections before starting new work. NOT for: active research execution or data analysis.

Guides research ethics compliance including IRB protocol preparation, informed consent document drafting, research integrity standards, data management plans, and ethical considerations for human/animal subjects; trigger when users discuss IRB, ethical approval, consent forms, or responsible conduct of research.

Promotes research reproducibility through guidance on pre-registration, open data/code sharing, replication study design, computational reproducibility, and open science best practices; trigger when users discuss replication, open science, pre-registration, data sharing, or research transparency.

Molecular chemistry operations via RDKit. Use when: user asks about molecular structures, SMILES, chemical properties, or fingerprints. NOT for: reaction databases or wet lab protocols.

Designs and analyzes quantum computing solutions including quantum circuit construction, algorithm implementation, error correction, and quantum advantage assessment; trigger when users discuss qubits, quantum gates, quantum algorithms, or quantum hardware.

Materials science computation with pymatgen. Use when: (1) crystal structure creation and manipulation, (2) phase diagram construction, (3) electronic structure analysis, (4) symmetry and space group operations, (5) VASP input/output parsing. NOT for: molecular chemistry (use rdkit-chemistry), protein structure (use biopython-bio), or general numerical computation (use scipy-analysis).

Search PubMed/MEDLINE for biomedical literature via NCBI E-utilities API. Use when: (1) searching medical/biomedical papers, (2) finding clinical studies, (3) querying with MeSH terms, (4) retrieving abstracts by PMID. NOT for: non-biomedical papers (use arxiv-search or semantic-scholar), full-text access (PubMed provides abstracts), or social science literature.

Search and retrieve chemical compound data from PubChem's PUG REST API (110M+ compounds). Use when the user needs compound properties, molecular structures, similarity searches, substructure searches, or bioactivity data. NOT for protein structures (use pdb-structure), NOT for drug-target interactions (use chembl-drug), NOT for gene-disease associations (use open-targets).

Conduct psychological research analysis including mental health, cognitive science, and behavioral studies

Analyzes protein 3D structures, performs homology modeling, interprets AlphaFold predictions, conducts molecular docking, and evaluates protein-ligand interactions; trigger when users ask about PDB files, folding, binding sites, or structural biology.

Analyze political data, fact-check claims, and study policy impacts using evidence-based methods

Physics problem solving including classical mechanics, electromagnetism, thermodynamics, quantum mechanics, optics, and computational physics. Use when user asks to solve physics problems, simulate physical systems, derive equations, or do unit conversions. Triggers on "physics problem", "Newton's law", "electromagnetic", "quantum", "thermodynamics", "optics", "wave equation", "Schrödinger", "relativity", "unit conversion", "circuit analysis".

Extract text and tables from PDF files, fill forms, merge documents. Use when working with PDF files or when the user mentions PDFs, forms, or document extraction.

Query the RCSB PDB API for protein 3D structures, experimental metadata, and structure files. Use when the user needs crystal or cryo-EM structure data, PDB entries, resolution info, or structure file downloads. NOT for protein sequences/annotations (use UniProt), gene data (use NCBI), or pathway info (use KEGG).

Conducts patent landscape analysis including prior art searches, patent claim interpretation, freedom-to-operate assessment, and intellectual property strategy for scientific inventions; trigger when users discuss patents, prior art, IP protection, or technology licensing.

Write scientific papers following IMRaD structure with proper citations. Use when: user asks to draft or improve a scientific paper, write specific sections, or format citations. NOT for: non-academic writing or presentation slides.

Read, summarize, and critically analyze scientific papers. Extract key findings, methodology, limitations, and contributions. Use when user shares a paper (PDF/URL/DOI), asks to summarize a paper, critique methodology, extract data from a paper, compare papers, or do a critical review. Triggers on "summarize this paper", "analyze this study", "what does this paper say", "critique this methodology", "extract findings from".

Open academic metadata via OpenAlex API. Use when: user needs author profiles, institution data, concept mapping, or open citation data. NOT for: full-text search or downloading papers.

NLP and corpus analysis via NLTK. Use when: user asks about tokenization, POS tagging, parsing, sentiment, or corpus statistics. NOT for: deep learning NLP or transformer models.

Natural language processing for research including text mining, sentiment analysis, topic modeling, named entity recognition, text classification, and corpus analysis. Use when user needs to analyze text data, extract information from documents, do sentiment analysis, topic modeling, or text classification for research purposes. Triggers on "text mining", "sentiment analysis", "topic modeling", "NER", "named entity", "text classification", "word embeddings", "LDA", "corpus analysis", "word frequency", "TF-IDF".

Supports neuroscience research including brain imaging analysis (fMRI, EEG), neural circuit modeling, cognitive experiment design, and neurological disorder investigation; trigger when users discuss brain regions, neural signals, cognitive tasks, or neuroimaging data.

Query NCBI E-utilities for GenBank sequences, gene info, SNPs, ClinVar variants, and literature links. Use when the user needs nucleotide/protein sequences from GenBank, gene summaries, variant data, or cross-database links. NOT for protein annotations (use UniProt), 3D structures (use PDB), or pathway mapping (use KEGG).

Multi search engine integration with 17 engines (8 CN + 9 Global). Supports advanced search operators, time filters, site search, privacy engines, and WolframAlpha knowledge queries. No API keys required.

Machine learning pipeline for scientific research including data preprocessing, feature engineering, model selection, training, evaluation, and interpretation. Covers supervised/unsupervised learning, deep learning, cross-validation, hyperparameter tuning, and model explainability. Use when user asks to build a predictive model, classify data, cluster samples, do feature selection, or apply ML to research data. Triggers on "machine learning", "classification", "clustering", "random forest", "neural network", "deep learning", "predict", "feature selection", "cross-validation", "train model".

Medical question answering using structured biomedical knowledge bases and clinical datasets

Scientific visualization via Matplotlib. Use when: user asks for plots, charts, or data visualization. NOT for: interactive dashboards or web-based charts.

Mathematical computation including symbolic math, numerical methods, linear algebra, calculus, differential equations, optimization, and mathematical modeling. Uses Python with SymPy, NumPy, SciPy. Use when user asks to solve equations, compute integrals/derivatives, do matrix operations, solve ODEs/PDEs, optimize functions, or build mathematical models. Triggers on "solve equation", "integral", "derivative", "matrix", "eigenvalue", "differential equation", "optimization", "linear algebra", "symbolic math", "proof".

Orchestrates a materials screening workflow from database search through property filtering to stability assessment and ranking. Use when identifying candidate materials for batteries, catalysts, semiconductors, or other applications. NOT for molecular chemistry or biological compound analysis.

Analyzes material properties including crystal structures, phase diagrams, mechanical/thermal/electronic properties, and supports materials discovery through computational approaches; trigger when users discuss alloys, ceramics, polymers, nanomaterials, or materials characterization.

Query the Materials Project API v3 for crystal structures, band gaps, formation energies, and thermodynamic stability of 150k+ inorganic materials. Use when: (1) searching materials by chemical formula, (2) looking up material properties by MP ID, (3) filtering materials by band gap, energy, or density, (4) finding stable phases for a composition. NOT for: organic molecules (use pubchem-compound), proteins (use uniprot-protein), drug compounds (use chembl-drug).

Comprehensive multi-database scientific literature search orchestrating Semantic Scholar, OpenAlex, arXiv, PubMed, and CrossRef. Use when: (1) systematic literature review, (2) finding all relevant papers on a topic, (3) checking state of the art, (4) building comprehensive bibliographies. NOT for: single-database queries (use specific search skills), data analysis (use code-execution).

Analyze language structures, typological features, and semantic change across languages