Why This Exists

The U.S. government helped build the largest biomedical knowledge engine on Earth: PubMed — tens of millions of research papers. Your taxes funded the labs. Public universities did the work. Scientists wrote the findings. Other scientists peer-reviewed them for free.

Then the system does something almost religiously cynical: it lets private publishers capture the public record.

They didn’t fund the study. They didn’t run the experiment. They didn’t do the peer review. But they control the door. So you — who already paid — get a preview page and a paywall: $35 per paper to “access” what you financed.

And that’s only the first lock. Even when you can get in, the language is engineered to keep ordinary people out: clinical, defensive, credential-coded prose that requires years of training to translate into “What does this mean for my body? For my kid? For my risk?”

So the public gets crushed between two walls:

1. Paywall: “You can’t read it.”
2. Credential wall: “Even if you read it, you won’t understand it.”

That’s not an accident. That’s a governance model. The government can claim “transparency” because the titles are searchable. The publishers can claim “value” because the prose is indecipherable. And the public — patients, families, citizens — are told to shut up and trust the priesthood.

This is old power. In Rome and its provinces near first-century Palestine, information was a controlled substance: literacy as a gate, language as a gate, status as a gate, permission as a gate. The common people got parables. The elite got records. Until the tables were flipped.

How The Machine Works

Here is the lifecycle of a publicly funded research paper.

A researcher at a public university, paid with public funds, conducts a study funded by an NIH grant — also public funds. She writes a paper. She submits it to a journal. Volunteer scientists — also employed at public universities — review it for free. The journal publishes it. The researcher signs over her copyright. The journal charges $35 to read the paper, or bundles it into a subscription costing universities $10,000 to $50,000 per year.

The five largest academic publishers — Elsevier, Springer Nature, Wiley, Taylor & Francis, and SAGE — control more than half of all published research. Elsevier alone has reported profit margins above 30%, consistently higher than Apple, Google, or Amazon.

The researcher who wrote the paper? She cannot access her own published work without paying or going through her university’s subscription. If she changes institutions or retires, she loses access to the paper she wrote. Academics routinely email each other PDFs of their own papers because the legitimate channels are broken.

This is not a conspiracy. It is a business model. “Publish or perish” culture means tenure requires publication in prestigious journals. Prestigious journals require copyright transfer. The system feeds itself.

The Second Lock: Language

Even when you can access the abstracts for free through PubMed, the search interface requires boolean operators and Medical Subject Headings (MeSH) — a controlled vocabulary of over 30,000 terms that medical librarians spend years learning. A plain question like “does vitamin D help prevent colds?” must become:

("Vitamin D"[MeSH] OR "Cholecalciferol"[MeSH]) AND ("Respiratory Tract Infections"[MeSH] OR "Common Cold"[MeSH]) AND ("Prevention and Control"[Subheading] OR "Dietary Supplements"[MeSH])

And the papers themselves speak clinical jargon. A study about whether exercise helps depression does not say “exercise helps depression.” It says: “A statistically significant reduction in PHQ-9 scores (mean difference −3.2, 95% CI −4.1 to −2.3, p<0.001) was observed in the intervention group relative to controls at 12-week follow-up.”

So the public pays three times. Once through taxes. Again through tuition. And again through paywalls. And even after all that, the language ensures you cannot understand what you paid for.

PubMed offers the abstracts for free. That is the crack in the wall. Tables Turned drives a truck through it.

Why Not Just Ask ChatGPT?

These tools treat you as a consumer of answers. You ask. They deliver. You accept. Sound familiar?

It is the same gatekeeping pattern wearing a friendlier mask. The journals said: trust us, we curate the knowledge. The algorithms say: trust us, we summarize it. In both cases, the evidence is not in your hands. You are still on the outside of the glass, trusting an intermediary to get it right.

Tables Turned takes a different position: you should be in dialogue with the evidence itself. Not receiving a verdict. Not consuming a summary. Sitting at the table, choosing which papers matter, reading what they actually say, and walking away with every claim traced to its source.

The difference between an answer box and a dialogue is the difference between being told what to think and being equipped to think for yourself.

The Dialogue

Here is how Tables Turned works, step by step.

Step 1: You ask a question

Type your question in plain English. Add your decision context — the reason you are asking. “Should I give my kid melatonin?” is different when the context is “they can’t sleep before school tomorrow” versus “building a long-term sleep routine.” The context shapes which evidence matters most. No API key needed — just ask.

Step 2: AI translates your question into PubMed’s language

Your question is sent to Claude, Anthropic’s AI. It does not just pull keywords. It generates multiple search strategies using real medical terminology — MeSH headings, boolean operators, synonym expansions — the same language a medical librarian would use. These strategies range from broad (casting a wide net) to specific (targeting exact conditions and interventions).

You can see every search term it generated, and why. Nothing is hidden.

Step 3: Papers come back in plain language

The search terms query PubMed’s public API — the same interface used by researchers and institutions worldwide. Papers return with their titles, authors, journals, years, and abstracts. Then AI translates each paper’s title and abstract into plain language — not dumbed down, but made readable.

You see both versions: the plain-language translation and the original technical title. You decide which papers to include. This is your curation step. The AI does not decide what is relevant. You do.

Step 4: AI reads and writes your brief

The papers you selected are sent to Claude with strict instructions: separate what the papers observe from what can be inferred. Cite every claim with a PubMed ID. Surface contradictions between papers — do not hide disagreement. State confidence honestly. Name what is unknown. Flag any claim that cannot be traced to a provided paper as [UNWITNESSED].

The synthesis streams in real time. You watch it being written. When it is done, you have a brief you can download as a Word document — with every claim linked to its source.

You came in with a question. You leave with receipts.

The Prompts

Every AI interaction in Tables Turned uses a carefully designed prompt. Most AI tools hide their prompts. We show you ours in full, because transparency is not optional when health decisions are at stake.

Search Query Generation

When you type your question and press search, this prompt converts your plain English into optimized medical database queries with real MeSH headings and boolean logic.

Plain-Language Translation

After finding papers, this prompt translates each title and abstract into language anyone can understand — without losing the substance.

Synthesis

This is the core prompt. It governs how the AI reads the papers and writes your brief. Every rule serves a purpose: separating fact from inference, enforcing citations, surfacing disagreement, preventing hallucination.

Honest Limits

This tool has real limitations. Knowing them makes you a better user.

• 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. If a paper matters to your decision, find and read the full text.
• AI can err. Despite strict instructions, Claude may occasionally misinterpret an abstract or make an unsupported inference. Every claim includes a PMID citation — click it and read the abstract yourself. The receipts exist so you can verify.
• Not medical advice. This tool helps you understand what research exists. It does not replace a physician. It gives you better questions to ask your doctor, not answers to follow blindly.
• English-dominant. PubMed indexes predominantly English-language abstracts. Research published in other languages may not appear.
• Search boundaries. The AI generates search strategies, but PubMed’s index has limits. Very new papers may not be indexed yet. Very niche topics may have little published research. The search stats tell you exactly how many papers were found and how.

Technical Details

Data source

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. Free, public, no account required.

AI model

Claude by Anthropic: All AI interactions use Claude, Anthropic’s most capable model. Your requests are routed through a minimal Cloudflare Worker proxy that holds the API key — your browser never sees or stores it. The worker adds the key and forwards your request to Anthropic. No data is logged or stored on the worker.

How papers are selected

When you search, AI generates 3–4 independent search strategies — each targeting your question from a different angle (broad terms, specific MeSH headings, study types). Each strategy queries PubMed separately and returns up to 25 results (configurable in Settings). That means the initial pool can contain 75–100+ papers before deduplication.

For example: if you ask “Is Ozempic safe?”, the AI might generate: (1) a broad search for semaglutide safety, (2) a targeted MeSH query for GLP-1 agonists and adverse effects, (3) a search filtered to systematic reviews and meta-analyses, and (4) a search focusing on cardiovascular outcomes. Each returns up to 25 papers from PubMed’s index. Papers found by multiple strategies are almost certainly central to your question.

All results are then scored and ranked:

1. Cross-strategy overlap (+3 points per strategy): A paper found by multiple independent strategies is more likely to be central to your question. If a paper appears in 3 of 4 strategies, it gets 9 overlap points.
2. PubMed position weight (+1–5 points): PubMed returns results in relevance order. The top result in any strategy gets 5 points, second gets 4, and so on. Papers ranked 6th or lower get 1 point.

The composite score (overlap × 3 + position points) determines rank. The top 12 papers are shown to you for curation. Papers that appear in multiple strategies display a badge (e.g., “3/4 strategies”) so you can see why they ranked high.

Compare this to using PubMed directly: you would need to know boolean operators and Medical Subject Headings (MeSH) — a controlled vocabulary of 30,000+ terms — to construct a single search. You would get one list of results with no plain-language translations and no way to quickly assess relevance. Tables Turned runs multiple expert-level searches, pools and scores the results, translates them into plain language, and lets you decide what matters.

The full scoring data — every strategy, every score, every matched paper — is recorded in the Tablet (.json) export.

Privacy

No API key is needed — the key is held securely on a Cloudflare Worker and never exposed to your browser. Your questions and papers pass through the worker only to reach Anthropic’s API; nothing is logged or stored. Your briefs exist only in your browser. When you close the tab, they are gone unless you downloaded them. No tracking. No analytics. No database.

The Lineage

Tables Turned inherits from a long tradition of putting knowledge back in common hands.

When William Tyndale translated the Bible into English in 1526, he was burned at the stake for it. The gatekeepers of his era argued that scripture was too dangerous for ordinary people — that only priests could be trusted to interpret the Word.

When Andrew Carnegie built 2,509 free public libraries across the English-speaking world, he was told the working class would not use them. They did.

When Aaron Swartz downloaded millions of academic papers from JSTOR and tried to make them freely available, he was prosecuted by the federal government. He died at 26. The papers are still behind paywalls. The publishers are still posting 30% profit margins.

The pattern is always the same. Knowledge exists. Gatekeepers insist it must be mediated. Translators refuse the premise. Eventually, the public record returns to the public.

This tool is one small table flip in that tradition. It does not replace reading. It does not replace thinking. It does not replace your physician. 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 tables have turned. Take your seat.