📡 Idea Radar 每天 8:00 / 18:00 推送 · 18:00 生成精选
reddit · SideProject · 55 分 ← 返回列表

I built EvalSeal: reproducibility receipts for LLM evals

I’ve been working on EvalSeal, a small open-source tool for making LLM eval results more trustworthy. The idea is simple: a single eval score is not enough. EvalSeal runs each case multiple times, measures flip rates, captures provenance, and seals the result into a tamper-evident ledger. A finding from the current repo: Using an LLM judge, 5 of 20 borderline cases flipped verdicts across repeated runs. Using numeric answer matching on 40 GSM8K cases, 0 cases flipped. Same model

五维评分

付费意愿
40
需求真实度
45
个人可落地
90
竞争空间
40
新鲜度
80

AI 分析

痛点:LLM eval 结果不稳定且缺乏可复现证据:同一模型用 LLM judge 评分时,边界用例多次运行会翻转结论,团队无法信任 CI 门禁或评测结论。

目标用户:AI 工程师、LLM 评测/CI 负责人、Agent 可靠性团队、需要审计评测结果的企业。

证据:A finding from the current repo: Using an LLM judge, 5 of 20 borderline cases flipped verdicts across repeated runs. Using numeric answer matching on 40 GSM8K cases, 0 cases flipped. Same model family, different grading method. The instability came from the judge, not the target model.

变现:开源自托管 + 云托管/团队版收费:CI 集成、评测收据签名与审计、历史趋势、协作看板、企业合规。

MVP:2 周内做 GitHub Action/CLI:接收 eval 结果或运行 eval,多次重复跑边界用例,计算 flip rate,生成签名收据并评论到 PR;先支持 promptfoo/OpenAI evals 等。

竞品:promptfoo、DeepEval、Braintrust、LangSmith、Weights & Biases、Humanloop、OpenAI Evals

摘要:EvalSeal 是一个 v0.3.0 开源工具,用重复运行、flip rate、provenance 和防篡改账本来让 LLM eval 结果更可信。作者发现 20 个边界用例中 5 个在 LLM judge 下翻转,而 GSM8K 数值匹配 0 翻转。它击中评测可靠性痛点,但尚无付费或外部需求验证,适合做开核+托管 CI/审计服务。