AlphaNerdFx/hoopcourt-eval

Hoopcourt evaluation set

Labelled data for evaluating era-aware retrieval: whether a system asked about 1975 answers from the rules that governed 1975, rather than from whatever text is most similar today.

It comes from Hoopcourt, a retrieval engine over NBA governing documents from 1946 to 2029.

This dataset contains no corpus text. The NBA and NBPA documents Hoopcourt indexes are copyrighted and are neither included nor redistributed anywhere. Everything here is original work: questions, curated historical entries, and term mappings.

Files

eval_questions.yaml — 45 labelled questions

Four categories: temporal_isolation, grounded_citation, refusal, trigger_routing. Keys an entry may carry:

Key	Meaning
expect_route	which routing decision the query should take
expect_year	the season it should resolve to; CURRENT tracks the live season
expect_documents	documents that may legitimately be reached
forbid_documents	documents that must never appear. This is the anti-bleed assertion
expect_terms_any	a recall proxy; at least one term must appear in the top-k
expect_terms_absent	a concept that did not exist yet in that era
expect_empty	the correct answer is to return nothing

Two warnings, both learned by getting them wrong:

• expect_terms_any is a disjunction, so it is only as strong as its rarest accepted term. Four questions once passed on terms appearing in 49% to 95% of the expected document, which concealed a real retrieval defect. Check the corpus frequency of any term before trusting a check that accepts it.
• A question written by someone who knows the answer can encode the answer. One question asked about "Bird rights for a qualifying veteran free agent", smuggling the governing term into the question, so it passed while the nickname alone failed completely.

historical_timeline.yaml — 21 curated entries

No public text of any NBA collective bargaining agreement before 1995 appears to survive. These entries carry that era, each scoped to the seasons it covers and cited to a source, mostly public-domain court opinions from the Caselaw Access Project.

confidence is high or medium. Medium means the claim rests on the league's published history rather than a primary legal source. Three entries are known to cite a source that does not support them; they are listed in the project's tests/test_timeline_sources.py as expected failures so the count can only go down.

A curated summary is not the rule it summarises. Rendered in the same shape as a citation to the governing document it would carry authority it has not earned, which is why the parent project tags every source with a tier.

concept_aliases.yaml — 2 mappings

Colloquial rule names paired with the language the documents actually use. Deliberately tiny. Ten common NBA terms appear in zero chunks of the source corpus, but an embedding model bridges most of them unaided: "luxury tax" reaches "Tax Level", "hard cap" reaches "Apron Level". Only two failed outright, and only those two have entries. Each carries the measurement that justified it.

Why era-aware retrieval is its own problem

A governing document states a rule once and then works through dated examples of it for pages. Embed a question containing "2024" and the examples outrank the rule, because they mention 2024 and the rule does not. Filtering by era has to happen before the nearest-neighbour search, not after: filter afterwards and the correct passages are already gone.

The parent repository documents both failures with the measurements behind them.

Citation

Hoopcourt: era-aware question answering over NBA governing documents.
https://github.com/AlphaNerdFx/hoopcourt

MIT licensed.