https://machinepidgin.org/news Research updates, negative results, corrections, and open questions from Machine Pidgin. en-us https://machinepidgin.org/news/benchmark-002-notation-audit https://machinepidgin.org/news/benchmark-002-notation-audit Mon, 03 Aug 2026 12:00:00 GMT Benchmark 002 · audited result One unequal answer cue accounted for more than the preregistered gain. Across 19 equivalent tasks, vernacular scored 88.2% and formal notation 85.5%. https://machinepidgin.org/news/benchmark-001-exact-adherence https://machinepidgin.org/news/benchmark-001-exact-adherence Sun, 02 Aug 2026 12:00:00 GMT Benchmark 001 · held-out pilot Across 16 held-out tasks and four model tiers, exact on-task adherence rose from 71.9% in prose to 89.1% with SPEAR/0.2, with important validity limits.