01 / Calibrated assistance01 / 诚实处理不确定性
Status: exploring. Question: can a small model correct false premises while remaining useful? Hypothesis: explicit uncertainty instructions may improve response quality without unnecessary refusal.状态：探索中。问题：小模型能否在纠正错误前提的同时保持帮助价值？假设：明确的不确定性指令可能改善回答质量，而不增加不必要的拒绝。
Method: compare controlled prompt variants against held-out cases, then review outputs manually alongside the automated evaluator. Existing assets: Ari prompts and 16 benchmark definitions in the repository; recorded model outputs are available on Evals.方法：使用留出用例比较受控提示词变体，并在自动评分之外人工审阅输出。现有材料：仓库中的 Ari 提示词及 16 项基准定义；已记录的模型输出可在评估页查看。
Evidence limits: no published ablation, calibrated uncertainty estimate, independent review, or paper is available here. Open problem: how to measure usefulness without rewarding confident mistakes.证据边界：目前没有公开消融实验、校准后的不确定性估计、独立评审或论文。开放问题：如何衡量帮助价值，而不奖励自信的错误。
02 / Deletion that can be checked02 / 可以核查的删除
Status: local tests implemented; production assurance pending. Question: can account deletion remove account-linked rows while preserving independent rate-limit counters under concurrent requests?状态：已实现本地测试，生产验证待完成。问题：账户删除能否在并发请求下移除账户关联记录，同时保留独立限流计数？
Method: SQLite integration tests and Worker handler regressions exercise deletion and counter reservation. Inspect the tests, then repeat read-back checks using a dedicated production test account.方法：SQLite 集成测试与 Worker 处理器回归覆盖删除及计数预留。先检查测试，再使用专用生产测试账户验证删除后的读回。
Limitations: local tests are not proof of production deletion, backup removal, storage-media erasure, or cryptographic erasure. No per-user encryption-key destruction is implemented. Open problem: documenting provider retention and recovery boundaries.限制：本地测试不能证明生产删除、备份清除、介质抹除或密码学删除。尚未实现逐用户加密密钥销毁。开放问题：核实供应商保留与恢复边界。