本地大重构与服务器就绪计划
- Stable ID:
codex-user-prompt:c382ed7dfc607cee
- Source kind:
codex-session-user
- Category:
coding-agent-workflow
- Timestamp:
2026-05-11T12:01:40.780Z
- Semantic hash:
c382ed7dfc607cee35a71b99390eecea2c044095580835c7b06486ad27bdc3f3
- Public handling: selected full-text prompt with secret filtering.
Prompt Text
# 本地大重构与服务器就绪计划
## Summary
明确的新 idea:**ObsShift-BO**
用 DeepSeek/API 与本地 Qwen3-8B 观察器发现低样本蛋白优化中的可复现决策失败模式,把 accepted pain-point ledger 蒸馏成 LoRA/SFT 本地观察器,再把这些痛点转化为可被现代 baseline 证伪的 shift-aware proposal、risk、batch 策略。
Harness engineering 只作为项目执行方式,不作为论文贡献。
默认决策已锁定:
- 清理策略:分层归档,不硬删可复现核心。
- LLM 公平策略:统一 LoRA,统一 Qwen3.5/Qwen3 8B backbone where applicable,baseline 用官方默认/报告超参。
- 复杂任务:默认多 agent,至少 reviewer agent + implementation agent + integration owner。
- 提交策略:按主题分批 commit,验证通过后 push GitHub。
## Key Changes
### 1. 仓库清理与架构重排
- 新建清晰三层结构:
- `kernel/` 或保留现有核心包:data、representation、models、uq、acquisition、proposal、loop、evaluation。
- `protocols/`:observation、finetune、baseline、low_n_audit、method、server。
- `paper/claim layer`:只保留 claim/evidence/reviewer/milestone,不再反向定义代码。
- 分层归档旧内容:
- `config/experiment/day*`, `*_main5`, `*_smoke`, 旧 v2/v3 文档迁入 `archive/legacy_experiments/`。
- 保留 `paper_main_configs.yaml` 作为兼容索引,但新增 semantic protocol aliases。
- 删除或本地归档明显无用文件:旧 zip、旧 version 本地目录、临时 YAML/SVG、重复草稿。
- 测试从“锁文档句子”迁移到“锁协议 schema、runner 输出、server command 可执行性”。
### 2. Observation -> LoRA/SFT -> Method 主线
- Observation tiers 固定:
- Smoke:`1 synthetic task x 8 records x 2 observers`
- Pilot:`4 real tasks x 32 records x 2 observers`
- Main:`12 main tasks x 64 records x DeepSeek+Qwen = 1536 records`
- Appendix:`10 tasks x 32 records x 2 observers = 640 records`
- DeepSeek/API 与 Qwen3-8B runner 都必须支持:
- `--workers`
- rate-limit/retry
- deterministic record ids
- JSONL outputs
- rejected/accepted pain-point ledger
- LoRA/SFT path:
- accepted observations -> instruction JSONL
- Qwen3-8B LoRA observer
- no full finetune in main setting unless all comparable LLM baselines also full finetune
- Method freeze rule:
- 每个 ObsShift-BO 组件必须有 pain-point row。
- 没有 pain-point 对应的组件删除,不允许拼接式 A+B+C。
### 3. Modern Baseline Framework
- Baseline 至少 8 个 family,不再把 acquisition variants 当 8 个 baseline。
- 必须覆盖:
- random/mutation library
- greedy supervised surrogate
- GP/RF/TPE BO
- deep ensemble BO
- conformal/weighted conformal internal controls
- zero-shot PLM
- PLM embedding + small head
- active/few-shot protein optimization
- LLM-directed evolution
- foundation/generative PLM
- structure-conditioned design where fair
- 参考项目仅作设计基准,不抄袭/缝合:
- [ProSpero](https://arxiv.org/abs/2505.22494)
- [Large Language Model is Secretly a Protein Sequence Optimizer](https://arxiv.org/abs/2501.09274)
- [DPLM-2](https://proceedings.iclr.cc/paper_files/paper/2025/hash/57c30b677add9aa78e1745f0643104d0-Abstract-Conference.html)
- [ESM3](https://www.nature.com/nature-index/article/10.1126/science.ads0018)
- [Evo 2](https://www.nature.com/articles/s41586-026-10176-5)
- [ProteinGym](https://papers.nips.cc/paper_files/paper/2023/hash/cac723e5ff29f65e3fcbb0739ae91bee-Abstract-Datasets_and_Benchmarks.html)
- [FLIP/FLIP2](https://flip.protein.properties/)
- Fairness policy:
- official implementation
- adapt only dataset I/O
- use Qwen3.5/Qwen3 8B shared backbone where applicable
- unified LoRA
- baseline official/default hyperparameters
- our method tuned with documented validation budget
### 4. Server-Ready Deliverables
- 生成一个 server run bundle:
- `server/run_all_observation.ps1`
- `server/run_lora_observer.ps1`
- `server/run_low_n_audit.ps1`
- `server/run_modern_baselines.ps1`
- `server/run_obsshift_bo.ps1`
- 每个命令只需要设置环境变量和路径即可运行。
- 每个 runner 输出:
- `run_manifest.json`
- `config_snapshot.yaml`
- `metrics.json/csv`
- `failure_report.md`
- `next_action.json`
- 本地 dry-run 必须全部通过后才能上服务器。
## Test Plan
- Unit/schema tests:
- protocol manifests load
- baseline registry >= 8 families
- observation records schema valid
- LoRA data conversion valid
- server job matrix emits commands
- Dry-run tests:
- DeepSeek dry-run parallel
- Qwen dry-run parallel
- LoRA dry-run
- small-model training smoke
- baseline job generation
- Integration tests:
- one synthetic smoke end-to-end
- one real dataset pilot end-to-end if local data exists
- Full validation before commit:
- `py -3.12 -m unittest discover -s tests -p "test_*.py"`
- dry-run server command generation
- git status clean except intended outputs ignored
## Milestones And Stopping Criteria
- M1 local rebuild complete:旧实验归档,semantic protocols 接管,所有 local dry-run 通过。
- M2 observation complete:DeepSeek + Qwen live main-panel observations 完成,accepted/rejected ledger 固化。
- M3 LoRA observer complete:accepted observations converted,Qwen observer LoRA 成功或明确硬件阻塞。
- M4 baseline complete:至少 8 baseline families 可运行或正式标记 unavailable with reason,其中至少 3 个 PLM/LLM4Protein 已跑。
- M5 method complete:ObsShift-BO 每个组件都有 pain-point 与 falsifying baseline。
- M6 writing-ready:12-task main + 10-task appendix,10 seeds where feasible,CI/paired tests/effect sizes/corrected p-values 完成。
只有 M6 满足时,才能说“实验阶段基本结束,可以进入写作”。
## Assumptions
- 使用“分层归档”而不是硬删除可复现实验。
- 主比较默认统一 LoRA,不混用 full finetune。
- Qwen3.5 8B 和 Qwen3 8B 作为同类 backbone policy;实际服务器可用哪个,就在 config 中锁定哪个。
- Harness engineering 不进入论文 novelty,只进入执行、审计、复现与停止规则。
- 下一次切出执行模式后,按上述顺序分批实现、测试、commit、push。
这些你也别忘了要有,我刚才的说明只是让你在这个基础上改
Reuse Notes
- EXTRACTED: This is a selected Codex prompt or automation prompt from the local Codex corpus.
- INFERRED: Future agents can reuse its structure, constraints, and acceptance criteria when creating similar Codex workflows.