Commit History

4 days ago

  1. facts: models spread 3.3 times as far within a language, since 3.3 is a ratio of standard deviations (5.6 against 1.7 points), not of variances

    4a60b5f

5 days ago

  1. facts: Skew separates acquiescence rather than cancelling it, the bias pre-existed in Qwen, Figure 6 has eight of seventeen above the line, the paired-complement trick comes from probability research, language coverage per the data, size gradient within providers, sigma not variance, Track B dropped, and six misattributed citations moved or removed

    a4d6f0e
  2. facts: HELM does not back the error-bar check, so its citation comes off it

    fc961f5

15 days ago

  1. prose: split a run-on, fix a connective and a comma splice, cut a restated sentence and a restating aside

    c0d4a81
  2. facts: Sonnet's case-study answers fall short of 100 rather than double-counting, Skew tilts rather than overestimates, and 3.3 is a variance ratio

    cec8b73

28 days ago

  1. Interactive research article

    5c130ff