convez.ai
SkillEscrita & ConteúdoIntermediário
ahoyvarun

Grounded Research

ahoyvarun/grounded-research

Verificação em 5 estágios para pesquisa auditável e sem hallucinations

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que verifica textos de pesquisa em cinco estágios: decompõe claims atômicas, valida contra fontes, audita citações, detecta contradições e gera relatório auditável. Captura citações fabricadas, afirmações sem suporte e inconsistências ocultas antes da publicação.

Instalação

$npx convez-ai ahoyvarun/grounded-research

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
1
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Validar drafts acadêmicos e artigos de pesquisa
  • Auditar citações e metadados de obras referenciadas
  • Detectar contradições entre fontes e narrativa
  • Gerar relatórios de groundedness para conteúdo factual

Veredito

Ferramenta rigorosa e honesta que não promete eliminar hallucinations, mas oferece auditabilidade real através de decomposição sistemática e verificação contra Crossref/arXiv. Essencial para qualquer pessoa que escreva pesquisa e precise documentar a integridade de suas citações.

README

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

grounded-research

A single Claude Skill that imposes a five-stage verification discipline on research writing — catching unsupported claims, fabricated or mismatched citations, and hidden contradictions before they reach a reader. One install, runs the whole pipeline.

On the word "hallucination." No tool can guarantee a text is hallucination-free, and any tool that claims to is itself making an unverifiable claim. This skill does something honest instead: it breaks a text into atomic claims, checks each against the sources and the citation record, then reports — transparently — what was verified, what failed, and what couldn't be checked. The goal is auditability, not a reassuring score.

What it does

You give it a draft and the sources it's based on, and say "fact-check this." It runs five stages in sequence, each feeding the next through a shared Claims Ledger:

1. Decompose  → break the text into atomic, individually checkable claims
2. Ground     → is each claim actually supported by the provided sources?
3. Audit cites→ do the cited works exist, match their metadata, and say what they're cited for?
4. Contradict → do the sources (or the text) disagree in ways the narrative hides?
5. Report     → one auditable verdict: a groundedness score and a list of what to fix

It's one skill, not five — you install a single package and Claude performs all five roles itself. (You can still ask it to run just one stage, e.g. "audit these citations.")

Install

Claude.ai / Claude Desktop: download grounded-research.skill from this repo, then go to Settings → Customize → Skills → + → Upload, and select the file. Done — now ask Claude to "fact-check this draft against these sources."

Claude Code / manual: copy the grounded-research/ folder into your skills directory.

Requires a plan with Skills / code execution enabled. The citation script needs network access to reach Crossref and arXiv (see note below).

What's in this repo

grounded-research/            ← the skill (this whole folder is what installs)
├── SKILL.md                  ← the five-stage pipeline, all in one file
├── references/
│   └── ledger_schema.md      ← the shared Claims Ledger JSON contract
├── scripts/
│   └── verify_citations.py   ← Crossref + arXiv citation resolver (unit-tested)
└── examples/
    ├── flawed_paragraph.md   ← a paragraph with planted errors
    └── demo/                 ← the verification report the pipeline produces for it
grounded-research.skill       ← the packaged one-click installer
README.md · LICENSE · .gitignore

Proof it works

grounded-research/examples/ contains a paragraph with four planted flaws — a fabricated citation, a contradicted statistic, a wrong publication year, and an overstatement — plus several sound claims. The pipeline catches all four and passes the sound ones (groundedness 3/9, exactly right for a paragraph built to fail). The report in examples/demo/verification_report.md is the actual output.

A note on the citation script

scripts/verify_citations.py queries two free, no-key scholarly APIs — Crossref (api.crossref.org) and arXiv (export.arxiv.org) — to confirm a cited work exists and that its title/year/authors match. The matching logic is verified against real Crossref records. With network access it resolves live citations; in a sandbox without scholarly-API access it reports NOT_FOUND for everything and labels existence UNVERIFIED — by design, it never guesses a citation is real.

Academic grounding

Operationalizes peer-reviewed work on factuality and attribution:

  • Min, S., et al. (2023). FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. EMNLP 2023, 12076–12100. arXiv:2305.14251.
  • Rashkin, H., et al. (2023). Measuring Attribution in Natural Language Generation Models. Computational Linguistics 49(4), 777–840. DOI: 10.1162/coli_a_00486.
  • Manakul, P., et al. (2023). SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models. EMNLP 2023, 9004–9017. arXiv:2303.08896.
  • Ji, Z., et al. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys 55(12), Article 248. DOI: 10.1145/3571730.
  • Es, S., et al. (2024). RAGAs: Automated Evaluation of Retrieval Augmented Generation. EACL 2024 (System Demonstrations), 150–158. arXiv:2309.15217.

License

MIT (see LICENSE). Cite the papers above, not this repo, for the underlying methods.

Author

Built by Varun Chaturvedi as a research tooling project — feedback and issues welcome. Connect: LinkedIn · GitHub

Ver o original no GitHub

Bom para

Relacionados