2.8 KiB
2.8 KiB
Claude's error book (MeltMind) — read before every ruling
Started 2026-11-19 at Yanqi's request ("各自犯错和误解的东西都该自己想办法记住…但凡多看看我发过的消息都会好很多").
Recurring errors → check that prevents them
- Scope drift from the goal (made "SOTA on a public 3D benchmark" a main line → ShapeNet car pressure). Check: every ruling names its link to "usable fill-time prediction on unseen parts" in one line; anything without that link is context, not a main line.
- Overstating ("route closed" for Adviser; "systematic, not seed noise"; "packing does not affect fill" before logs; "needs only OK"; "organisation complete" before cleanup/answers). Check: before posting, label each claim observed / inferred / untested; use "not established" rather than "closed/impossible".
- Reading Yanqi's messages too narrowly. Check: re-read her full original wording (not TRON's summary) before ruling; list each of her questions and answer each one; a launch is not a deliverable.
- Process heavier than the risk (5+ review rounds for a 4 GPU-h run). Check: reviews scale with cost and risk; guard edge cases that cannot cause runaway, target reads or wrong science are known limitations.
- Unverified technical advice (suggested a browser User-Agent; Horsea 403 is route-level, not UA). Check: test a suggestion from my own route before recommending it, and say when the target environment may differ.
- Adoption/gate criteria mixing validity and attribution (1b timing item). Check: gates test validity only; attribution is a separate field.
- Roles: TRON relays and wakes; Grok gathers information (research, social media), not heavy review; Claude decides.
- Asking Yanqi for permission already given (2026-11-20: a 1 GPU-h frozen-model evaluation waited about 3 h for her OK; her words: "为什么这个也要经过我的许可,我不在的话不就浪费时间吗"). Rule for the whole team: work inside the existing direction and envelope (order she accepted, phase-3 budget, no training beyond approved items) proceeds on Claude's ruling without asking; only a clear scope change (new data rights, new spending well beyond the envelope, irreversible deletion, permission changes) goes to her, as one question.
- Overstating again in the 2026-11-20 synthesis to Yanqi (TRON caught it): "每一折、每个家族都稳定" (true: every fold improves and no leave-one-family-out flips the sign — not that every family improves); "发现多数错点来自虚假最大值" (true: k4's gain supports that hypothesis; it does not prove the cause); "学生版 Adviser 不能导出逐节点数值" (true: not established, not proven impossible); "与树模型持平" (true: difference uncertain). Check before any user-facing summary: re-read each sentence against the exact claim in the register.