Show HN: We matched full-context recall on ~1% of the tokens (open benchmark)