rekursiv-ai/trackinizer/websearch

WebSearch — authoring expectations

A WebSearch records a web/paper search and the references it surfaced, so the search is reproducible and its findings are linked.

Per-kind expectations doc (SoT). Fields owned by types/inquiries.py; grammar by ../SKILL.md.

Completeness bar

• query = the search string exactly as issued (reproducibility).
• provider = the engine (google, duckduckgo, arxiv, scholar, s2,

crossref); empty only when genuinely untracked.

Findings are edges, not a column

Papers/pages a search surfaced are recorded as produced_by edges (WebResult/Paper → WebSearch: the finding is the younger child of the search), NOT a field. Found artifacts are many-to-one (two searches surfacing the same Paper share one node, each with its own produce edge).

Expectations

• A WebSearch can itself cite a claim: websearch S proves belief P with positive

valence ("a standard search confirms this is widely reported") or negative ("the position appears in no major index").

• Record the query verbatim; a paraphrase breaks reproducibility.

Common mistakes

• Storing findings as prose instead of produced_by edges to real artifact rows.
• Paraphrasing the query.
• Omitting the provider when it is known.