Measured, Not Promised

Same AI assistant. Then we connected BARK.

We ran the same agent through the same 18 regulatory questions: once on its own, once connected to BARK. The only variable is the connection.

95% fewer

Answers that break a must-not rule

11.7% → 0.6%, 17 Sep run

Zero

Confidently wrong answers

5.6% → 0.0%, 17 Sep run

+22 points

Cites the correct primary provision

59.6% → 81.6%, 17 Sep run

    18 factual regulatory questions, 10 answers each per arm. Answer sheets written by a panel of regulatory experts. Answers by Claude Opus 5, and by Opus 5.5 from 4 Oct. Judged blind. Percentages show the improvement with BARK over Claude alone on the same run.

    Judge: Opus Extra High and SOL Extra High until 17 Sep, Opus 5.5 from 3 Oct.

    All measures by run, as a table

    What it takes to make Claude match BARK

    We told Claude to verify every claim against the law and cite it, pushing it until its work matched what BARK produces. These are the results.

    2.4x

    the cost of BARK

    3.8x

    the time of BARK

    16%

    of its links open the article. BARK: 99%

    Claude Claude pushed to verify everything Claude + BARK
    Quality score 80.0 87.7 88.0
    Cost per answer $0.29 $1.44 $0.60
    Time per answer 49 s 235 s 62 s
    Links to the official source 61% 66% 100%
    Link opens the exact article 0% 16% 99%

    Quality scores from one blind grading of both answer sets together, so they are not the same numbers as the per-run scores in the chart above. Cost is the API list-price estimate per answer, not what BARK charges. Links counted over every citation in the 90 answers per column.

    Method

    How we measured it

    Each run puts the same agent through the same questions twice: once on its own, once connected to BARK. The chart shows five such runs, from August to October: BARK updated between the first four, and a newer answering model, Opus 5.5, for the fifth.

    Questions
    18 factual regulatory questions
    Answers
    10 per arm, per run
    Answer sheets
    written by a panel of regulatory experts
    Answering model
    Claude Opus 5, medium effort; Opus 5.5 from 4 Oct
    Judging
    blind, by Opus Extra High with SOL Extra High as second judge until 17 Sep; by Opus 5.5 from 3 Oct
    Runs
    five, 22 Aug to 4 Oct 2026: BARK updated between the first four, the answering model changed for the fifth
    Variable
    the BARK connection, nothing else

    Don't take our word for it. Test it on your own questions.

    Contact us