TextInsightBench / docs /VERIFICATION.md
erwinmsmith's picture
Rebuild open exploration tasks and corpus-grounded evaluation
149da17 verified
|
Raw
History Blame Contribute Delete
1.2 kB

Verification scope

The release was checked with local automated tests and an all-50-task abstaining agent smoke run. This verifies loading, corpus hashes, task dispatch, output contracts and reporting, not real-agent mining quality or empirical difficulty.

Synthetic tests exercise agent-selected group/date comparisons, forbidden filters, overlapping groups, minimum population sizes, exact partitions, Simpson-style reversals, missing metadata, counterexamples, quotation offsets, score bindings, null reference coverage and bounded reproducible semantic packets. No paid agent benchmark or independent validation set was added.

python -m unittest discover -s tests -v
tib verify --data data/participant --with-learning

The data builder deterministically selects disjoint IDs from an already curated pool, preserves original text, enriches released metadata and filters selected documents out of the remaining pool. It records exact input shard hashes. Current-snapshot disjointness does not erase historical public exposure. The design increases the required exploration and analysis workload; actual difficulty and discriminative power still require empirical agent results.