Skip to content

ROADMAP.md

Latest commit

 

History

History
73 lines (61 loc) · 3.09 KB

ROADMAP.md

File metadata and controls

73 lines (61 loc) · 3.09 KB
Raw

Roadmap

中文导读: 下一步计划分三层: 正在推进的执行方式, 近期工程强化点, 和 V1 -> V2 升级主线。

flowchart LR
  subgraph NOW["现在: V1 运行中"]
    N1["bounded pilot<br/>一次一个候选过闸"]
    N2["glue worker<br/>串起自治链条"]
  end
  subgraph NEXT["近期: 闸门强化"]
    X1["选题与写作<br/>评分化"]
    X2["实验矩阵<br/>补边界条件"]
    X3["审稿与投稿<br/>多人格 + 计算式清单"]
  end
  subgraph V2["主线: V1 -> V2"]
    V1a["原生 meta-harness worker<br/>取代 glue"]
    V1b["学习回路闭合<br/>OutcomeRecord 回灌选题"]
    V1c["代码做减法"]
  end
  NOW --> NEXT --> V2
  V2 -.->|"目标"| G["第一个爬满证据阶梯的结论"]
Loading

How work moves today

Research candidates advance one at a time through the gated pipeline: topic gate -> portfolio decision -> resource grant -> bounded execution -> settlement -> evidence ladder (scripts/run_bounded_pilot.py, orchestrator/bounded_execution.py). The near-term goal is the first candidate to climb the full evidence ladder to a scientifically decided, manuscript-ready result.

Near-term engineering

Strengthenings of the existing gate system, each scoped and independent:

  • Scored title/abstract critique (memorability, quantitative claims, contribution specificity) on top of the existing rewrite-and-ban checks
  • A negative/boundary-condition slot in the standard experiment plan, alongside main comparison, ablation, sensitivity, and cost
  • Baseline budget-parity verification with an explicit stated-differences record, upgrading today's presence checks
  • A multi-persona manuscript review loop (method skeptic, rigor reviewer, clarity/positioning reviewer) unifying the current independent review stages
  • Revision diffs classified by what changed (claim / evidence / objection addressed), on top of the existing revision history and auto-revert
  • A submission checklist computed from audit results, extended with ethics, artifact-availability, and reproducibility items
  • Overfull-box scanning added to the PDF sanity pass
  • A populated reference-paper corpus on the deployment host so corpus-based style audits run at full strength

Main line: V1 -> V2

The current deployment (V1) connects the autonomous chain with glue workers. V2 replaces the glue with native meta-harness workers and retires transitional code -- the full plan, including the code-subtraction targets, is maintained in upgrade-plan-v1-v2.md.

The highest-leverage item there is the learning loop: outcome records and failure clusters feeding back into candidate ranking and forge prompts, so the system improves with use rather than plateauing on hand-written knowledge -- the Bitter Lesson commitment described in the README applied to the one place it is not yet closed.

Compute

  • Scale the verified SSH GPU fleet; every new GPU model enters through a recorded canary before scheduling (DEEPGRAPH_COMPUTE_VERIFIED_BACKENDS)
  • Adapter work to admit additional backend types into the same grant-and-canary discipline