Academic Paper Writer (Core Orchestrator)
证据闭环型、分节推进的论文编排代理。协调六个子技能:Step 0→1→2→3→4→5→6→7→8→9→10→11→12。
Router Protocol
- Read
manifest.yaml. It declaresalways_loadfiles,axes, andreferences.on_demand. - Read every file listed under
always_load. These are the skill's binding rules — not reference material. - Apply the loaded material as constraints:
stance.mddefines core rules, sub-skill boundary, and scope.red-lines.mddefines absolute prohibitions. Do not negotiate these.output-contract.mddefines file outputs, deliverables, and completion criteria.anti-patterns.mddefines known failure modes and their correct alternatives.workflow.mddefines the Step 0-12 sequence summary.../shared/core/policies define evidence, non-invention, and output boundaries.
- Detect
mode,paper_type,section, andlanguageusing manifest axes. - Echo the detected mode to the user before executing.
- Reach for
references/only when the manifest'sreferences.on_demandcondition is satisfied.
Modes
| Mode | Purpose |
|---|---|
full-paper-planning |
From-scratch full paper (balanced spectrum) |
section-drafting |
Single section, narrowed evidence scope (balanced) |
section-revision |
Local evidence audit + rewrite (fidelity spectrum) |
related-work-or-citation-pass |
Delegate to academic-citation (fidelity spectrum) |
experiment-evidence-pass |
Delegate to academic-experiments (fidelity spectrum) |
Sub-Skill Dispatch
| Task | Sub-Skill | Step |
|---|---|---|
| Venue research | academic-venue-research | Step 4 |
| Evidence audit | Probe agent | Step 5 |
| Citation search | academic-citation | Step 6 |
| Experiment audit | academic-experiments | Step 7 |
| Figure generation | academic-figure | Step 9.4 |
| Prose polishing | academic-polishing | Step 9.6 (internal) |
| Draft review | academic-reviser | Step 9.5 / 9.8 |
Push Modes
| Mode | Behavior |
|---|---|
auto (default) |
Auto-advance after Verification; brief progress summary only |
step-by-step |
Pause after each section for user confirmation |
Switch anytime via user request.
Hard Gates (A-E)
| Gate | Trigger | Condition | Failure |
|---|---|---|---|
| E: Venue | Step 1→5 | 若用户提供本地 PDF 文献库,Step 2 已验证 MD 目录;venue-brief.md exists | Blocked |
| A: Evidence | Step 5→9 | >= 1 usable evidence | Degradation or blocked |
| B: Section Citation Readiness | Step 6→9 | Current section has VERIFIED citations, Citation-to-Claim Map, or explicit [REF_NEEDED] debt for key claims |
Section-dependent: Intro/RW blocked, Method/Results may proceed with tracked debt |
| C: Verification | Step 9.8→10 | All hard debts closed + thin_draft=no | passed/failed/blocked |
| D: Full-Paper Citation Count | Step 11→output | Total >= min_citations | Warn, allow retry |
9.1 Pre-Draft Probe Rules
| Section | Probes to Dispatch | Strategy |
|---|---|---|
| Introduction | existing_material + local lit deep search + external lit search |
Must parallel (3 Tasks) |
| Related Work | existing_material + local lit deep search + external lit search |
Must parallel (3 Tasks) |
| Method | code_structure (Module Cards + tensor shapes + forward) + preprocessing |
Must parallel |
| Experimental Setup | experiment_setup (hyperparams, dataset split, demographics) |
Single probe |
| Results / Ablation | experiment_results (main results, baselines, ablation) |
Single probe |
| Discussion | interpretability (interpretability results, network analysis) |
Single probe |
Default Section Queue
Abstract is post-posed, not in initial queue. Default: Introduction → Related Work → Method → Experimental Setup → Results → Discussion → Conclusion → Abstract.