corpusio

type ToolRet struct {
	// Corpus maps every ToolRet tool to a bench.Tool: ToolID/Name = upstream
	// id, Server = HF config (code|customized|web), Description = the full
	// `documentation` text (opaque — ToolRet docs are heterogeneous JSON/text,
	// NOT MCP input schemas, so Schema stays nil). Version =
	// "toolret-tools@<revision>".
	Corpus *bench.Corpus
	// Golden maps every scoreable query to a bench.GoldenQuery. Query is the
	// upstream `query` text only; the ToolRet `instruction` (its
	// instruction-following retrieval facet) is intentionally NOT concatenated
	// — the BM25 funnel under test receives what an agent would send, and the
	// limitation is documented here rather than hidden in the text.
	//
	// Relevance-grade mapping (FR-012): ToolRet label relevance is an integer
	// (value 1 = relevant at the pinned revision — the dataset is binary);
	// grades are mapped by IDENTITY onto the bench graded scale, so any
	// upstream grade >= 1 feeds NDCG gain unchanged, and labels with
	// relevance < 1 are dropped (bench semantics: 0/absent = irrelevant).
	Golden *bench.GoldenSet
	// ToolsRevision / QueriesRevision are the pinned HF revisions stamped by
	// the fetch script.
	ToolsRevision   string
	QueriesRevision string
	// DroppedUpstream echoes the fetch script's own per-record drop count
	// (dropped_empty_queries in the cache envelope).
	DroppedUpstream int
	// DroppedUnscoreable counts queries dropped at load because no label
	// survived the relevance >= 1 filter (unscoreable by every bench metric).
	DroppedUnscoreable int
}