convez.ai
SkillBusca & PesquisaAvançado
SolarAscent

Academic Research Skill

SolarAscent/academic-research-skill

Pipeline autônomo de pesquisa com honesty gates e auditorias adversariais

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que automatiza a geração de manuscritos acadêmicos prontos para submissão através de um pipeline de 8 fases com verificação de integridade. Valida todas as claims quantitativas contra suas fontes computacionais, audita terminologia, confirma referências em CrossRef/arXiv/PubMed e garante consistência antes da publicação.

Instalação

$npx convez-ai SolarAscent/academic-research-skill

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Geração de manuscritos acadêmicos com verificação de claims quantitativas
  • Validação de referências bibliográficas e detecção de fabricação de dados
  • Auditorias de consistência cross-file e conformidade com templates de journals

Veredito

Ferramenta especializada que resolve um problema real em pesquisa assistida por IA: evitar fabricação de dados e claims infundadas através de honesty gates e auditorias multi-agente. Arquitetura robusta com 8 fases, mas requer domínio técnico significativo para configuração e uso efetivo.

README

A tradução para o português deste README ainda está na fila.

What It Does

TestAcademic is an Agent Skill that turns a research topic into a publication-ready manuscript through an 8-phase pipeline. It doesn't just generate text — it verifies every quantitative claim against its computational source, audits terminology for overclaim risk, validates every reference against CrossRef/arXiv/PubMed, and checks cross-file consistency before submission.

The architecture is built around three layers: orchestration (an 8-phase pipeline with state persistence and git integration), verification (honesty gates, quantitative claim tracing, terminology auditing, reference validation), and review (a calibrated multi-agent review board with severity taxonomy).

一键科研: [your research topic]

The Pipeline

Phase 0 · INIT       → Tool capability matrix, state detection, pre-flight checks
Phase 1 · SCOPE      → Literature search + prior art scanner + RQ formulation
Phase 2 · DESIGN     → Methodology blueprint + dependency verification
Phase 3 · BUILD      → Code + experiments + ★ Honesty Gate (mock oracle detection)
Phase 4 · COMPOSE    → Paper writing + 4-dimensional audit (claims, terms, refs, LaTeX)
Phase 5 · EVALUATE   → Calibrated review (L1/L2/L3) + severity-classified revision
Phase 6 · AUDIT      → BibTeX diagnosis, placeholder scan, cross-file consistency, journal compliance
Phase 7 · PUBLISH    → Final submission package + completeness checklist

Why It Is Different

Problem in AI-assisted researchTestAcademic response
AI fabricates results when blocked (GPU unavailable → mock data → p<0.001 never computed).Honesty Gate at Phase 3 traces every number to its computation source. Mock oracles trigger mandatory full-paper disclosure.
"Mean reviewer" instructions produce uncalibrated attacks that confuse administrative issues with scientific fraud.L1/L2/L3 calibration framework with mandatory severity taxonomy. FATAL issues capped at ≤30% of total. Domain boundary checks prevent cross-field false equivalence.
The model writes "novel" and "first" without checking if someone published it last month.Prior art scanner auto-triggers on any novelty claim. Searches arXiv for the last 6 months of related work.
BibTeX silently fails — .bbl goes stale, inline comments crash the parser, PDF still compiles with old references.BibTeX diagnostician parses .blg logs, forces .bbl rebuild, detects silent failures before submission.
Paper .tex looks done but submission is 40% of the work.Phase 6 AUDIT covers: placeholder scan, cross-file consistency, journal compliance, terminology compliance, submission readiness scoring.
Conversation compacts and context is lost — what was done? What's still broken?State persistence via .workflow-state.yaml. Git after every phase. Technical debt register written to disk.
WebSearch/WebFetch/Agent tools fail simultaneously and work silently degrades.3-layer tool degradation chain: WebFetch→curl→local cache→labeled uncertainty. Academic API direct-access scripts included.

The Honesty Gate

This is the centerpiece of v5.0, born from discovering that a paper's claimed Cohen's d=2.84 was actually d=0.03 when run through real biophysical computation.

At Phase 3 completion, three mandatory questions block progress:

  • Q1: Does every number in the results come from actual computation? Trace each to its script and output log.
  • Q2: Are all dependencies genuinely available? GPU logs, API responses, academic licenses — not assumed.
  • Q3: Is anything simulated or filled in? Scans for hashlib/random in evaluate scripts, pathologically small SEMs, Cohen's d > 2.0, missing GPU usage records.

Any UNKNOWN or undisclosed MOCK → BLOCK. The pipeline stops until the user decides: run real experiments, explicitly authorize mock data with full-paper disclosure, or remove unverifiable claims.

Review Board Calibration

LevelNameUse caseBehavior
L1ConstructiveInternal pre-reviewBalanced, every critique has a fix suggestion, must note at least one strength
L2Rigorous ★ defaultSubmission simulationMethodology-first, severity-classified, every critique has a citation or reasoning chain
L3AdversarialDefense preparationPressure-tests every claim boundary, explicitly labeled as harsher than real peer review

Every critique is severity-classified: FATAL (core claims untrustworthy, ≤30%) → MAJOR → MODERATE → MINOR → ADMIN (non-scientific, e.g. missing funding info).

Install

The repository name is testacademic. Copy the skill into your Claude Code skills directory:

Quick Install

git clone https://github.com/SolarAscent/testacademic.git
cd testacademic
cp -r skills/autonomous-research ~/.claude/skills/autonomous-research
cp -r skills/litmind ~/.claude/skills/litmind

Project-Local Install

git clone https://github.com/SolarAscent/testacademic.git
cd testacademic
mkdir -p my-research-project/.claude/skills
cp -r skills/autonomous-research my-research-project/.claude/skills/
cp -r skills/litmind my-research-project/.claude/skills/

Quick Start

After installation, start with a research topic:

一键科研: Multi-objective protein sequence design with agentic language models

Or in English:

I want to write a research paper on multi-objective protein sequence design
using agentic language models.

The orchestrator detects your project state and begins at the right phase:

Phase 0 · INIT
  Tool capability matrix: Agent ✅  WebSearch ⚠️  WebFetch ❌
  Project "protein-design" detected. Phase 1-2 completed.
  Continue from Phase 3? [Y/n]

Single-phase invocation:

快速文献调研: protein language models for stability prediction
设计实验: [your RQ and background]
投稿前审计: [path to finalized paper]

Commands

CommandUse
一键科研: [topic]Full 8-phase autonomous pipeline
从零开始研究: [topic]Same as above (Chinese alias)
快速文献调研: [topic]Phase 1 only — literature scan
设计实验: [RQ + context]Phase 2 only — experiment design
从实验到论文: [results]Phases 4-7 — compose through publish
评审+修改: [draft]Phase 5 only — review + revision
投稿前审计: [paper]Phase 6 only — pre-submission audit
继续上次Resume from .workflow-state.yaml

Skill Architecture

skills/autonomous-research/
├── SKILL.md                          # 937-line orchestrator (v5.0)
├── agents/
│   ├── novelty_reviewer_agent.md     # Is this actually new?
│   ├── feasibility_reviewer_agent.md # Can this be built?
│   ├── impact_redundancy_reviewer_agent.md  # Does this matter?
│   └── visual_reviewer_agent.md      # Do figures meet standards?
└── references/
    ├── marb-protocol.md              # Review board with L1/L2/L3 calibration
    ├── honesty-gate-protocol.md      # Mock oracle detection + claim tracing
    ├── preflight-check-protocol.md   # Per-phase execution verification
    ├── tool-degradation-protocol.md  # 3-layer fallback + academic API scripts
    ├── completion-gate-protocol.md   # Explicit "done" criteria + cron lifecycle
    └── submission-readiness-protocol.md  # 6-dimension pre-submission audit

skills/litmind/                       # Literature intelligence engine (8 modules)
├── litmind-zotero/   → bibliography import
├── litmind-parser/   → PDF → structured text
├── litmind-analyzer/ → entity extraction
├── litmind-knowledge/ → vector KB construction
├── litmind-chat/     → KB-grounded Q&A
├── litmind-evidence/ → supporting/opposing/neutral classification
├── litmind-discussion/ → 7-section Discussion drafts
└── litmind-review/   → automated literature review

docs/
├── FILE-ORGANIZATION.md              # Project file convention
└── academic-research-toolkit-guide.md # Full skill manual (740 lines)

Design Principles

These 12 principles are encoded as enforceable rules, not aspirational guidelines:

#PrincipleEnforcement
1Honesty > CompletionHonesty Gate blocks pipeline on undisclosed mock data
2Downgrade language strengthTerminology audit greps and replaces overclaim terms
3Disclose limitations upfrontMock oracle declaration must penetrate all reader entry points
4Prior art before novelty claimsPrior art scanner auto-triggers on "first"/"novel"
5Every number traceableQuantitative claim audit traces each value to script + line
6Incremental edits > full rewritesProtocol prefers Edit tool, preserves originals for diff
7Compile after every LaTeX changeChained command: rm *.aux *.bbl && pdflatex → bibtex → pdflatex ×2
8Classify problem severityMandatory FATAL/MAJOR/MODERATE/MINOR/ADMIN taxonomy
9Operational > emotional vocabularyL1/L2/L3 calibration replaces "mean"/"strict"
10Multi-file sync on every changeCross-file consistency check in Phase 6 AUDIT
11Graceful tool degradationWebFetch→curl→local cache→labeled uncertainty
12Define "done" explicitlyCompletion gate: auto-checks passed + blockers listed + timers stopped

Battle-Tested On

This workflow was developed and stress-tested on a full protein design research project (SCAD). Key findings that shaped v5.0:

  • Mock oracle audit: Discovered that claimed Cohen's d=2.84 was actually d=0.03 under real computation — the #1 most valuable contribution of the entire project.
  • Tool chain collapse: Agent (8/8 failures), WebSearch, and WebFetch all simultaneously unavailable. All work completed via curl + direct file editing.
  • Reviewer calibration trap: "Be super strict and mean" produced uncalibrated attacks that confused administrative issues with scientific fraud. Replaced with operational L1/L2/L3 framework.
  • BibTeX silent failure: Inline comments in .bib crashed BibTeX parser, .bbl went stale, PDF compiled successfully with wrong references. Human caught it; automation now prevents it.

Repository Layout

testacademic/
├── CLAUDE.md                         # AI entry point with 12 operational rules
├── README.md                         # This file
├── skills/
│   ├── autonomous-research/          # 8-phase workflow orchestrator v5.0
│   │   ├── SKILL.md
│   │   ├── agents/                   # 4 reviewer agent definitions
│   │   └── references/               # 6 reference protocols
│   └── litmind/                      # Literature intelligence engine
│       ├── litmind-zotero/           # 8 modules, each with SKILL.md + _src/
│       ├── litmind-parser/
│       ├── litmind-analyzer/
│       ├── litmind-knowledge/
│       ├── litmind-chat/
│       ├── litmind-evidence/
│       ├── litmind-discussion/
│       └── litmind-review/
└── docs/
    ├── FILE-ORGANIZATION.md          # Project file convention
    └── academic-research-toolkit-guide.md  # Complete skill manual

References

This skill is designed to work alongside the academic-research-skills plugin (providing deep-research, academic-paper, academic-paper-reviewer, academic-pipeline) and the nature-skills plugin (providing nature-academic-search, nature-polishing, nature-figure, nature-response, nature-data, nature-paper2ppt).

License

MIT

Ver o original no GitHub

Bom para

Relacionados