Praxist Scientific Research

Praxist Scientific Research

Use this skill to build a source-backed research context for a Praxist task project. This is a planning and context-gathering skill, not an experiment runner.

Boundaries

• Do not start, stop, resume, or mutate a Praxist run.
• Do not edit Praxist core or generic plugins.
• Do not require additional API keys. Prefer public no-key sources and Codex web search when available.
• Do not treat papers, database entries, blog posts, benchmark pages, or open-access text as measured task performance.
• Do not download datasets, checkpoints, simulators, packages, licensed assets, or new runtime environments just because a source mentions them. Use those sources to improve the best solution possible under the task's current local data, simulator, dependencies, evaluator, hardware, and runtime. Record missing resources only as task-local notes or user-facing requirements.
• Keep domain-specific search policy and conclusions in the task project, usually under assets/literature/, description.md, role skills, or task prompt files.

Source Priority

1. Local project docs, papers, README files, benchmark docs, existing logs, and task-owned assets/literature/.
2. Public no-key literature/database tools when tool_server:literature_lookup is available in the standard or task-declared Praxist tool set.
3. Agent-host web search/open web browsing for official benchmark pages, dataset cards, standards, methods, and recent papers.
4. Optional credentialed or licensed sources only when the user explicitly provides them and the task requires them.

Workflow

1. Identify the task or research project root. Prefer the current directory when the user does not provide a path.
2. Read local task/project material first:
   • task.yaml, description.md, README.md, prompt_task.jinja2;
   • assets/literature/, dataset metadata, baseline summaries, result logs;
   • domain papers or PDFs supplied by the user.
3. Determine the domain and research objective. Keep categories broad enough to preserve diversity: machine learning, biology, chemistry, medicine, robotics/control, physics/materials, quantitative systems, software systems, or generic optimization.
4. If tool_server:literature_lookup is available, use:
   • literature_source_guide(domain, objective) to pick source families;
   • literature_search(query, sources, max_results) for compact public records;
   • literature_resolve(identifier) for DOI/PMID/arXiv/OpenAlex records.
   • literature_open_access_text(identifier_or_url, max_chars) when an open-access full text or PDF provenance record is needed;
   • scientific_database_search(query, sources, max_results) for public scientific databases such as Europe PMC, UniProt, or ClinicalTrials.gov.
5. If the tool is not available, use agent-host web search for official and primary sources where current information matters.
6. Produce compact notes with:
   • source title, URL or identifier, authors/organization, year/date;
   • what claim or design hint the source supports;
   • how reliable or directly applicable it is;
   • whether it affects metrics, benchmarks, constraints, high-value directions, or prior-art risk.
   • whether the source assumes resources not present locally; if yes, write a current-resource adaptation and a missing-resource note, not an install or download plan.
7. Save notes only when the user asked for task updates or task initialization is in progress. Recommended paths:
   • assets/literature/research_directions.md;
   • assets/literature/source_notes.md;
   • assets/literature/prior_art_risks.md.

Output Shape

For direct analysis, report:

• recommended metrics and why;
• benchmark/evaluation standards and official sources;
• high-value research directions, grouped by mechanism family and intervention surface;
• prior-art risks and leakage risks;
• open-access/provenance notes for sources used in task prompts;
• open questions that should become Praxist validation candidates.

For task files, write concise task-owned text. Do not create duplicate leaderboards, fake baselines, or Praxist runtime artifacts.

Evidence Language

Use this distinction consistently:

• Literature signal: a paper, database record, benchmark page, or source note suggesting a hypothesis or constraint.
• Open-access provenance: retrieval URL, timestamp, content hash, and credential scope showing where a public source came from.
• Measured task fact: evaluator output or structured finding produced by the task harness during a Praxist run.

Only measured task facts should drive performance claims. Literature signals can drive search directions and skepticism.