Generate a grounded translational research brief from ANY clinical or biomedical observation — a disease's drug response/resistance, a drug's adverse effect/toxicity, or any mechanism (not limited to any one domain).

ASYNC (submit + poll): the full pipeline (decompose -> hypothesize -> retrieve -> grade -> entail -> novelty-check -> design -> assemble) runs ~60-120s, longer on a cold first call — longer than an MCP client will wait on one call. So this tool does NOT block: it STARTS the run and returns immediately with a job handle. You MUST then poll get_experiment_result(job_id) every few seconds until status is "done" (the full TransBrief is in result) or "error". The finished brief has decomposed biological axes, up to 3 falsifiable hypotheses each graded against real PubMed evidence with an auditable novelty verdict, and ONE runnable computational experiment (top_experiment) naming a concrete public dataset plus a claude_science_prompt ready to run in Claude Science.

Args: observation: A free-text clinical/biomedical observation (3-8000 characters) — any disease, drug response/resistance, adverse effect, or mechanism. Examples: "58F, resistant hypertension despite ACEi + CCB + thiazide; elevated hs-CRP" or "30M on amiodarone for AF, developed neutropenia". focus_drug: Optional drug name to focus the analysis on. Omit ("") to let the pipeline infer relevant drugs from the observation itself.

Returns: Immediately: {"job_id", "status": "running", "poll_tool": "get_experiment_result", "message"}. Poll get_experiment_result(job_id) for the outcome — on "done" its result is the full TransBrief (carrying the fixed research-only disclaimer, BUILD_SPEC.md §0.5; never diagnosis, drug selection, or dosing), on "error" its result is {"error", "message", "status_code"} (e.g. no/invalid ANTHROPIC_API_KEY). Obviously invalid input is rejected inline here (no job) as that same error shape.

Look up PubMed-grounded mechanistic evidence for ANY clinical, pharmacological, or mechanistic question (utility / fallback tool — a lighter-weight sibling of generate_experiment, not limited to any one domain).

ASYNC (submit + poll): runs the SAME full TransBench pipeline as generate_experiment (not a separate, cheaper retrieval path), so it is just as slow (~60-120s) and is likewise non-blocking — it STARTS the run and returns a job handle immediately. You MUST then poll get_experiment_result(job_id) until status is "done"/ "error". On "done", result is the smaller, evidence-focused projection: for each hypothesis its novelty verdict and every retrieved evidence item (each with a resolvable citation, an entailment verdict, and an evidence grade), plus the deduplicated reference list, any contradictions surfaced, and the uncertainty note. Omits axes/ top_experiment/run_manifest.

Args: question: A free-text clinical/pharmacological/mechanistic question (3-8000 characters), any domain.

Returns: Immediately: {"job_id", "status": "running", "poll_tool": "get_experiment_result", "message"}. Poll get_experiment_result(job_id) — on "done" its result is the grounded-evidence projection (carrying the fixed research-only disclaimer), on "error" its result is {"error", "message", "status_code"}. Obviously invalid input is rejected inline (no job).

Poll for the result of a run started by generate_experiment or search_grounded_evidence (both are async: they return a job_id and run the ~60-120s pipeline in the background). Call this every ~5s with that job_id until it is no longer "running".

Args: job_id: The job_id returned by generate_experiment / search_grounded_evidence.

Returns: - still working: {"status": "running", "job_id"} — poll again. - finished ok: {"status": "done", "job_id", "result": <the full TransBrief, or the evidence projection>}. - failed: {"status": "error", "job_id", "result": {"error", "message", "status_code"}}. - unknown/expired job_id: {"error": "unknown_job", "message", "status_code": 404} — finished results are retained ~30 min (MCP_JOB_TTL_SECONDS) after completion, then evicted.