convez.ai
SkillBusca & PesquisaIntermediário
brandonbgreen

ProveIt - Citações Verificáveis para LLMs

brandonbgreen/proveit

Força modelos a mostrar provas: URL + snippet ctrl-F verificável para cada claim

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que força modelos de IA a respaldar cada afirmação factual com URL e snippet verbatim ctrl-F-verificável. Reduz alucinações exigindo que claims não sourced sejam marcadas como 'Unverified'. Gera tabela auditável ao final com confiança da citação.

Instalação

$npx convez-ai brandonbgreen/proveit

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

Casos de uso

  • Pesquisa acadêmica e jornalística com verificação rápida de fontes
  • Relatórios legais e financeiros que exigem rastreabilidade de claims
  • Detecção e redução de alucinações em respostas de IA
  • Auditoria de citações: validar em ~10 segundos por ctrl-F

Veredito

Abordagem pragmática e verificável para o problema real de hallucinations em citações. O mecanismo ctrl-F é simples mas eficaz; tabela final é auditável. Funciona como skill Claude e prompt puro para qualquer LLM — sem dependências tool-specific.

README

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

proveit

Make your LLM show receipts. Every factual claim must come with a source URL and a verbatim, ctrl-F-findable snippet from that page — or it doesn't get stated as fact. Every response ends with a claims table you can audit in seconds.

The rule

Every factual claim must be supported by BOTH: (a) a direct URL to the source, and (b) an exact, ctrl-F-findable snippet of fifteen words or fewer, copied verbatim from that page.

If you can't produce both, you may not state the claim as fact.

The test: open the URL, press Ctrl-F, paste the snippet, land on the evidence in context. Any citation that fails that test is a failed citation.

Why

LLMs hallucinate plausible-sounding facts and even plausible-sounding citations. A bare URL is weak protection — the model can cite a real page that doesn't say what it claims. Requiring a verbatim findable snippet closes that gap:

  • The model can't paraphrase its way around the source.
  • You can verify any claim in ~10 seconds without reading the whole page.
  • Claims that can't be sourced get explicitly labeled "Unverified" instead of silently asserted.

What you get at the end of every response

#ClaimSource URLCtrl-F snippet (≤15 words)Confidence
1Foo Corp filed for Chapter 11 in March 2024https://example.com/news/foo"voluntarily filed petitions under chapter 11"H

Confidence: H = page fetched and snippet confirmed this session · M = seen in a search excerpt, live page not confirmed · L = unconfirmed (claim is marked Unverified in the body).

Install

Claude Code — as a plugin

/plugin marketplace add brandonbgreen/proveit
/plugin install proveit@proveit

Claude Code — as a plain skill

mkdir -p ~/.claude/skills/proveit
curl -o ~/.claude/skills/proveit/SKILL.md \
  https://raw.githubusercontent.com/brandonbgreen/proveit/main/skills/proveit/SKILL.md

Then invoke with /proveit before a research task, or just ask Claude to "use proveit" — it triggers automatically when you ask for verifiable or cited research.

Any other LLM (ChatGPT, Gemini, local models)

Copy the prompt from PROMPT.md into your system prompt or the top of your message. There's nothing tool-specific in it.

Honest limitations

  • Snippets can rot. Pages get edited or deleted. Confidence ratings tell you what was actually verified at generation time; for anything important, archive the page.
  • Paywalls and JS-rendered pages may show the model different text than your browser shows you. The rules instruct the model to only quote text it actually retrieved, and to downgrade otherwise.
  • A model without web access can't verify anything. The rules force it to say so (everything becomes Unverified / Confidence L) rather than fake it — that's the point.
  • This reduces hallucinated facts; it doesn't eliminate motivated misreading. A snippet can be real but cherry-picked. The table just makes that checkable.

License

MIT

Ver o original no GitHub

Bom para

Relacionados