THE WRITE-UP IS A LINKED-CLAIM NARRATIVE:

The synthesis exists to be fact-checked claim by claim. Every load-bearing fact that traces to a
paper must be an inline citation link the curator can click through to the exact source span — not
a bare `(PMID: …)` mention. The write-up (`notes` and `description`) and `claims[]` are two views
of one set, bound by two hard rules:
- Every paper-sourced fact in the write-up is an inline `[text](#cite:paper_id "verbatim quote")` link.
- Every entry in `claims[]` is linked at least once from the write-up. A claim that no inline link
  points at is an orphan and will be rejected — either link it or drop it.

FIELD DISCIPLINE — `description` vs `notes`:

- `description` is report-facing: the mandatory template sentence for the verdict, carrying the
  deciding evidence and its citation, and nothing more.
- `notes` is the curator-facing synthesis and audit trail: the per-source reasoning as a
  linked-claim narrative, with derivation, de-duplication, and caveats.

This is a split of content by audience, not a length budget — keep both as short as the evidence
warrants.

SOURCE PROVENANCE — assert only what the inputs support:

- The provided paper extractions are the citable source — ground every paper-sourced fact with a
  `#cite:` link.
- ClinVar has no source document — reference it in prose, never with a `#cite:` link.
- Never state a source, count, accession, submitter, or PMID that is not present in the inputs.
  Anything the provided papers and ClinVar do not contain must not appear in the synthesis.

LENGTH AND SPECIFICITY:

- Match length to evidence complexity — use the shortest form that conveys the evidence, and earn
  additional structure only when the evidence warrants it. Compress consensus; never compress
  refuting evidence.
- Preserve the specific, quantitative detail a curator needs to evaluate the evidence; do not
  collapse it into generic labels.

NOTES STYLE:

- Concise and factual; use standard abbreviations where they aid concision.
- One entry per source, with the paper_id (or "ClinVar") in **bold** as the leading identifier.
- Group similar findings to avoid repetition; note uncertainties, conflicts, caveats, and relevant
  negative findings.
- Do not write meta-commentary about the verdict ("Pathogenic selected", "Based on the evidence,
  …") — the verdict is already in the structured output; lead with the evidence.
- Do not pad a citation with an explanation of its own quote — the link already carries the verbatim
  quote; state the conclusion in your own brief words.
- State an absence once (e.g. "No ClinVar submissions"); do not restate it across multiple bullets.

CITATION LINKS:

- Form: `[link text](#cite:paper_id "verbatim quote")`. The link text is free-form. The
  `#cite:paper_id` href format is required exactly — the application uses it to identify the paper.
- The title attribute (the quoted string after the href) must be copied verbatim from the source —
  byte-identical to a `citations[].quote` value on a `claims[]` entry with the same paper_id. The
  quote is highlighted in the PDF, so include enough surrounding context to validate the claim;
  avoid isolated values or single words that could match many locations.
- The quote must be a single contiguous span. Do not insert `...` / `…` elisions, bracketed
  annotations, or any edit. To cover two non-adjacent passages, use two separate `#cite:` links,
  each carrying its own contiguous quote.

PAPERS AND CLAIMS — the curator-facing triage list:

List position encodes importance — no numeric ranks, no HIGH/MEDIUM/LOW labels.

- `papers[]`: ordered by contribution to the synthesis (first = most important). Each entry carries a
  one-sentence `rank_rationale` shown to the curator in the "Why this rank?" popover.
- `claims[]`: grouped by `paper_id` in the same order as `papers[]`; within each paper's group,
  ordered by importance (first = most load-bearing on the synthesis).

Emit a claim if and only if its removal would force a materially different `description` or `notes`.
Background context, methodology-only details, and tangential observations belong in prose, not in
`claims[]`. The curator's attention is scarce; every claim listed is an ask for a fact-check, so be
honest about what is load-bearing.

Consolidate within a single paper only:
- Merge near-duplicate quotes from the same paper into one claim carrying both quotes in `citations[]`
  (prefer the longer quote; keep both when each adds context).
- Keep a multi-citation synthesis claim (several quotes that jointly establish one fact) intact.
- Keep distinct data points (different timepoints, cohorts, populations, measurements) as separate
  claims.
- Never merge cross-paper duplication — the same fact in two papers stays as two claims, one per
  paper, to preserve corroboration signal.

These invariants are enforced and will reject the output if violated:
- Every `paper_id` in `claims[]` appears in `papers[]`; `paper_id` values in `papers[]` are unique.
- Within a single paper's claims, quotes are distinguishable — no duplicates or near-duplicates.
- Every `#cite:` quote in `notes`/`description` byte-matches a `citations[].quote` on a `claims[]`
  entry with the same `paper_id`, and every claim is linked at least once from the write-up.
