Document reading
GPU OCR path, layout recovery and scan quality on declared corpora. Failures stay in the record.
RESEARCHKNOWLEDGE COMPILER
We test the compiler against the messy parts.
Status is a protocol, not a mood. Every public result carries a badge: MEASURED, REPRODUCED, SUPPORTED, NOT SUPPORTED, BUILT NOT PROVEN, or IN PROGRESS.
Knowledge construction
Entity identity, relation extraction and provenance packages exist as artifacts. External correctness benchmarks are not closed.
Knowledge maintenance
Selective recompilation is a research direction, demonstrated on fixture data, not a shipped claim.
Retrieval comparisons
External baseline comparisons stay unpublished until a frozen run is reproduced.
Machine-readable receipts and failed hypotheses live on /evidence. The public claim inventory is docs/CLAIM_LEDGER_v1.md in the product repository. Patent language does not belong in the hero.
We do not publish competitor comparison tables first. Order: reproduce a baseline, freeze the environment, publish config and metrics, publish failures, then review.