--- name: kol-paper-trial-triangulator description: Use when a medical-affairs or MSL user names a KOL (key opinion leader) plus a context query and wants that author's top-relevant papers each triangulated to the clinical trial it reports — outputs a trust-tagged publication-and-trial-linkage table plus an engagement-dossier workbook, grounded in Amass's author-scoped BioMedCore search and the cross-core publication↔trial graph (referencesTrialCore) so no NCT number is fabricated. license: Apache-2.0 metadata: { author: amass, version: "0.1.0" } --- # KOL paper-trial triangulator For a medical-affairs lead or MSL building a KOL engagement dossier. The single input is an **author name plus a context query** (e.g. `Mikhail Kosiborod · cardiometabolic heart failure`). The skill seeds an author-scoped BioMedCore search, then for each returned paper walks the cross-core `referencesTrialCore` edge to the trial that paper reports, and rolls the result into a per-paper publication-and-trial-linkage table plus a downloadable `.xlsx` engagement dossier. The hallucination delta: plain Claude, asked "which trials has this KOL's work on this topic reported on?", invents plausible NCT numbers and mis-attributes papers; Amass returns the author filter as a search parameter and the paper→trial link as a first-class graph edge, so every NCT is a returned field, never a guess. ## When to invoke - A medical-affairs / MSL user names a clinician-researcher and a therapeutic-area query and wants their recent relevant work mapped to the trials it reports ("what trials has Dr. X's HFpEF work reported on?", "build me a KOL dossier for Y on Z"). - Any author-scoped literature pull where the value is the paper→trial linkage, not just the bibliography — i.e. you need the cross-core spine per paper, not a flat reading list. - Works for any KOL + topic, not just the anchor: validated on Scott D. Solomon · heart failure preserved ejection fraction trial (a distinct trial portfolio — TOPCAT, PARAGON-HF, DELIVER). Do **not** invoke for an exhaustive bibliometric census of an author (search returns no total at any `limit`, so this is a relevance sample, not the full oeuvre — see Failure modes), nor when the user has a trial ID and wants its describing papers (that is the inverse trial→paper direction — walk `referencesBiomedCore` from a TrialCore record instead). ## Inputs 1. `kol` — the author name, passed verbatim to `authorNames` (e.g. `Mikhail Kosiborod`). Required. 2. `context_query` — a few distinguishing topic tokens, passed to `query` (e.g. `cardiometabolic heart failure`). Required — this disambiguates the author and ranks the sample. Strip punctuation/operators; the search is a bag-of-tokens match, no booleans/quotes/wildcards. 3. *(optional)* `min_jufo` — `minJournalQualityJufo` (0–3) to trust-gate the seed search. Default unset; set `2` for domain-leading-or-better. (Used on the Scott D. Solomon second-input run.) Anchor example: `kol = "Mikhail Kosiborod"`, `context_query = "cardiometabolic heart failure"`. ## The Amass MCP calls (exact sequence) 1. **Seed (author search — the SEED STEP ONLY).** `search_amass_biomedcore_records(query=, authorNames=[])` → up to 10 papers. A two-part name must be carried by a single author (order ignored), but it is still a token match with **no prefix expansion and no identity resolution** — a namesake matches. Verify the KOL actually appears in each returned record's `authors` list before keeping the row (on the anchor, all 10 carried Kosiborod). **Sample, not census** — this is the top-10 by relevance to the query, NOT the author's full bibliography. `limit` accepts 1–50 and defaults to the 10 this skill runs at, and no search returns a total — never claim exhaustiveness at any limit. To trust-gate, add `minJournalQualityJufo` — but confirm the returned `journalQualityJufo` on each row rather than assuming it as a hard gate. 2. **Triangulate (per-paper cross-core fan-out — the HEADLINE step).** For each returned paper: `get_amass_biomedcore_record(type="amassId", value=)` and read its `referencesTrialCore` array (the Amass trial ids of the trial(s) that paper reports). This is the load-bearing move; the `authorNames` seed only got you here. - **Rate-limit batching.** Pace calls to roughly one every 2 seconds; for larger author corpora, batch in chunks of ~40–55 calls and pause ~10 s between batches. On HTTP `429`, read `Retry-After`, back off, then resume. - **Metadata-only recovery on overflow.** A landmark paper's record can exceed the per-call token budget (its citing-paper list can run to thousands). If a fetch overflows, recover by reading just the identity metadata + `referencesTrialCore` rather than dropping the row. **On the anchor this was NOT triggered** — all 10 fetched clean, including the DELIVER NEJM primary (PMID 36027570, 2,390 citations). It is documented as the safety net for still-larger landmarks. 3. **(optional) Resolve the trial title.** For each non-empty `referencesTrialCore` id, you may `get_amass_trialcore_record(type="amassId", value=)` (or `type="nctId"`) to attach the trial's `briefTitle` / `acronym` / sponsor / phase / status. Keep it to field-grounded identity; the reciprocal `referencesBiomedCore` edge on the trial should list the paper back, confirming the link. ## Output template A header line naming the input honestly (`Author X · query "…" — top-10 by relevance, not the full oeuvre`), then one table sorted however the persona prefers (the anchor keeps the seed-search relevance order), with columns: `Title (verbatim) | Journal | Date | Cites | JuFo | Retracted | referencesTrialCore (NCT)` Each row's NCT cell is the literal `referencesTrialCore` value (or "— none (meta-analysis / review)" when the edge is empty). Then a one-line **Verdict** stated as a checkable field claim: "N of 10 papers carry exactly one `referencesTrialCore` link; M carry none." Close with the artifact emitted: an `.xlsx` engagement dossier — one row per paper with columns `paper_PMID, paper_AMBC, title, journal, date, journalQualityJufo, citationCount, isRetracted, referencesTrialCore_NCTs`. ## Failure modes & recovery - **Empty `referencesTrialCore` on a paper.** Not a failure — meta-analyses and reviews legitimately carry zero (they aggregate many trials, name none as a single primary). Render "— none (meta-analysis / review)"; never invent a trial to fill the cell. - **Token-budget overflow on a landmark paper.** Recover with a metadata-only read; keep the row. Documented fallback; not triggered on the anchor. - **Author returns thin or off-topic results.** `authorNames` is a token-match with no identity resolution; a common name can pull a namesake's papers. If the full name is thin, retry with the citation form (`Kosiborod MN`) or the surname alone, the widest match. Verify the KOL in each `authors` list; if the corpus is review/consensus-heavy, triangulation will be thin — report that honestly rather than forcing trial links. - **JuFo gate under weak token match.** If you trust-gate the seed with `minJournalQualityJufo` and the query has no strong token match, results may not all honor the filter — confirm the returned `journalQualityJufo` on each row before relying on it as a gate. - **429 rate limit.** Read `Retry-After`, back off, resume the fan-out. - **Do not guess → "not in abstract".** Any abstract-derived detail absent from the returned text is reported as that literal absence. Trial fields such as `primaryOutcomeMeasures` and `whyStopped` are quoted verbatim from the fetched trial record, never paraphrased. ## Guardrail This dossier names a **living physician** — medium stakes. Render **only field-grounded rows** — title (verbatim) + PMID + journal + date + JuFo + `isRetracted` + the `referencesTrialCore` NCT id(s). **Do NOT synthesize a subjective characterization** of the named person: no "thought leader", "sponsor-aligned", "influential", or any inferred standing or intent. Frame the output strictly as a **"publication + trial-linkage record for author X (top-10 by relevance to the context query; not the full oeuvre)"**. The verdict is a checkable field claim — a count of linked vs. unlinked papers — not an opinion about the person.