The Doctrine
Rome is a pattern, not a place
Every era has its gatekeepers. In the ancient world, priests controlled access to scripture. In the medieval world, the Church controlled access to knowledge. In the modern world, journals and institutions control access to research.
The pattern is always the same: knowledge exists, but ordinary people are told they cannot be trusted with it. That they need interpreters. That the raw material is too complex, too dangerous, too sacred for common hands.
Tables Turned refuses this premise. The public record is public. PubMed abstracts are free. The question is not access. The question is translation.
The information flood
The modern crisis is not a lack of information. It is a drowning in it. PubMed alone contains over 37 million citations. No human can read them all. No human needs to. What people need is a way to find the relevant research, understand what it says, and make an informed decision.
This is not a problem AI solves by replacing human judgment. It is a problem AI solves by being the translator at the table. You bring the question. The public record provides the evidence. AI reads the abstracts and explains what they say in plain language. You decide what to do.
Receipts only
The core principle is provenance. Every claim in a Tables Turned brief traces back to a specific paper, identified by its PubMed ID (PMID). This is not a suggestion. It is a rule. If a claim cannot be traced to a provided paper, it is marked [UNWITNESSED] and flagged.
This matters because the history of AI-generated content is a history of hallucination. Models invent citations. They blend findings from different papers. They state conclusions that sound authoritative but have no source. Tables Turned treats this as a moral failure, not a technical limitation. Every claim has receipts, or it is marked as unsupported.
How It Works
Step 1: Ask your question
Type a plain-English question and the decision it serves. The decision context matters because it shapes what kind of evidence is relevant. "Does melatonin help kids sleep?" is different when the context is "deciding whether to try it tonight" versus "writing a school paper on sleep disorders."
Step 2: AI-powered search
Your question is sent to Claude, which generates optimized PubMed search terms. This is not a simple keyword extraction. The AI understands medical terminology, knows how to construct boolean search queries, and generates multiple search strategies to maximize coverage.
The search terms are shown to you transparently. You can see exactly what was searched and why. The results are fetched directly from the NCBI E-utilities API, the same public API that powers PubMed itself.
You then review every paper found and choose which ones to include. This is your curation step. You are not a passive recipient. You decide what enters the synthesis.
Step 3: Synthesis
The selected papers (abstracts, titles, authors, journals, years) are sent to Claude with strict instructions:
- Separate what the papers say from what we infer
- Every claim must cite its source by PMID
- Surface contradictions between papers
- Write in plain language, no jargon
- State confidence honestly (strong, mixed, or thin)
- Name what is unknown
- Never hallucinate or invent findings
The synthesis streams in real-time so you can watch it being written.
Step 4: Your brief
The finished brief is rendered with formatted headings, highlighted PMID citations, and clear section breaks. It includes:
- What the papers say — grouped by finding, each cited
- What we infer (with receipts) — numbered claims, each sourced
- What is unknown — gaps the papers do not address
- Confidence — honest assessment with reasoning
Download as a formatted Word document (.docx), markdown (.md), or the raw session data (.json).
Technical Details
Data sources
PubMed / NCBI E-utilities: All paper metadata and abstracts are fetched from the National Center for Biotechnology Information's public API. This is the same database used by researchers, doctors, and institutions worldwide. It is free, public, and requires no account.
AI model
Claude by Anthropic: The synthesis is powered by Claude, Anthropic's AI assistant. You provide your own API key, which means your data is sent directly from your browser to Anthropic's servers. Nothing passes through our servers because we do not have servers. This is a static website.
Privacy
Your API key is stored in your browser's localStorage. It is never sent anywhere except to Anthropic's API. Your questions, papers, and briefs exist only in your browser. When you close the tab, they are gone unless you downloaded them. We do not track you. We do not have analytics. We do not have a database.
Limitations
- Abstracts only. Tables Turned reads abstracts, not full papers. Abstracts are free and public. Full papers are often paywalled. This means nuance in the methods or results sections may be missed.
- AI can err. Despite strict instructions, Claude may occasionally misinterpret an abstract or make an inference that is not fully supported. The PMID citations let you verify every claim yourself.
- Not medical advice. This tool helps you understand research. It does not replace a doctor. It gives you better questions to ask, not answers to follow blindly.
- English only. PubMed abstracts are predominantly in English. Research published in other languages may not appear.
The Lineage
Tables Turned inherits from a long tradition of making knowledge accessible. From the translation of scripture into common tongues, to the creation of public libraries, to the open-access movement in academic publishing, the pattern is always the same: gatekeepers resist, translators persist, and eventually the public record returns to the public.
This tool is one small table flip in that tradition. It is not a replacement for reading. It is not a shortcut for thinking. It is a translator that sits at the table with you, reads the public record aloud, and makes sure every claim has a receipt.
The commons table is set. Pull up a chair.