I built an evidence agent that refuses to answer without receipts (my first hackathon entry)

Описание и применение

Создан агент, который отвечает только с доказательствами: каждый ответ сопровождается официальным источником, цитатой и датой актуальности, а противоречия отображаются в виде карточек с историей.

Задача

Frontier model specifications (context windows, prices, capabilities) change monthly, causing LLM answers to become stale quickly.

Результат

The agent provides grounded answers with receipts, flags stale claims, and resolves contradictions via timeline cards, demonstrating a proof‑of‑concept for evidence‑based model fact retrieval.

Рабочий процесс

• 1. Collect official provider pages for model specs.
• 2. Create evidenceClaim documents in a Sanity dataset with source URL, quote, and observed date.
• 3. Use Sanity’s hosted Context MCP endpoint to query the knowledge base.
• 4. Build a Next.js UI on Netlify to display answers and contradiction cards.
• 5. Deploy the agent, keeping environment secrets in a vault.
• 6. Test the agent by asking constraint questions and reviewing contradiction cards.

Доказательства

Ask it the constraint question today and it answers: one model satisfies every constraint - Claude Haiku 4.5, 200,000-token context (official Anthropic models page), $1 per million input tokens (official Anthropic pricing page), both observed today, 0 stale claims, 0 active conflicts.

Источник

The part I'm proudest of: contradiction cards

Источник