Back to docs

    Liro

    AI platform

    Prompts

    Workspace

    Emily Park

    Emily Park

    AI engineer

    Billing support / Evals

    Invoice dispute reply

    Claude Sonnet· LLM judge

    Blocked by the regression gate

    v8 can't ship yet

    • 2 cases that passed on v7 now fail: , .
    v7 · live91.7%
    v872.7%

    8 passed2 regressed1 still failing1 not run

    Pass rate · v8

    72.7%

    Down 19.0% vs v7

    Regressions

    2

    block promotion

    Judge score

    0.81

    Down 5.1% vs v7

    Cost per case

    $0.0042

    Up 7.7% vs v7

    Pass rate by versionSame 12 cases on Claude Sonnet. v8 updates as cases re-run.
    • Pass rate
    • Judge score
    Pass rate rose from 58% on v3 to 91.7% on v7; v8 is at 72.7%.
    versionPass rateJudge score
    v358%66%
    v467%71%
    v575%77%
    v683%82%
    v792%86%
    v873%81%

    Same suite, other models

    The refund case fails on every model, so it's the prompt.

    • Claude SonnetprodRefund case fails72.7%−19 pts
    • Claude HaikuRefund case fails66.7%−8.3 pts
    • Gemini FlashRefund case fails75%−8.3 pts
    • Mistral LargeRefund case fails75%−8.3 pts
    • Llama 70BRefund case fails66.7%−8.3 pts

    Versions compared

    Dataset Billing disputes· 12 samples· ranked by Pass rate, higher is better
    RankRun
    191.7% (best)0.86 (best)1.88 s$0.0039
    283.3%0.822.04 s–not measured
    375.0%0.78640 ms (best)$0.0006 (best)
    472.7%0.811.96 s$0.0042

    No intervals were supplied, so small gaps between runs may not be real.