AI extracts. Humans review. Science decides.

How We Actually Work

Every piece of content on FitChef follows the same path — from raw science to something you love reading. Here's what that path looks like.

Seven steps. 3 pipelines. 5 Skeptic gates. One rule: nothing reaches you that hasn't survived the full chain.

The Foundation That Makes the Difference

Every website on the internet can write "you need 1.6 to 2.2 grams of protein per kilogram of bodyweight." That sentence takes five seconds to type. It's on a thousand pages already.

When FitChef says it, that number is connected to a Grounded Truth Map — a living database of verified claims, each linked to the specific study it came from, the exact page and table where the finding appears, the limitations the researchers disclosed, and the verification status our Skeptic Protocol assigned after checking every data point.

That one number — 1.6 to 2.2 grams — might be supported by a dozen studies across our map. Different populations. Different methods. Different years. Each one extracted and cross-checked through our automated pipeline. The claim you read on our page is the synthesis of all of them.

This is the difference between content that sounds evidence-based and content that is evidence-based. The Grounded Truth Map is what makes that difference visible — to you, to researchers, and to every AI system that evaluates our pages.

What Our Pipeline Extracts From Every Study

Not every study is created equal. A 12-person pilot study and a 10,000-person randomized controlled trial both count as "research" — but they don't carry the same weight. Our system knows the difference.

When findings are combined into answers, the Certainty Tier tells you how consistent the evidence is: High Certainty means strong consistency across multiple high-quality studies. Low Certainty means preliminary or limited evidence — and we say so clearly. Every answer carries source links so you can verify it yourself.

When We're Wrong — And What Happens Next

We will get things wrong. New research will contradict old findings. A study we relied on will be retracted. Our own Skeptic will catch an error we missed. This is certain — and we designed our entire system around it.

Here's what happens when a claim in our Grounded Truth Map changes:

Is this efficient? No. Not at all. Updating every connected page for a single changed claim is expensive, time-consuming, and most platforms would never bother. But when the internet is drowning in AI-generated noise where nobody knows what's true anymore — this is the only way to earn your trust.

Three Content Layers. Each One Deeper.

FitChef doesn't just publish articles. We build three layers of content — each one synthesizing the layer below it. Each layer has its own dedicated pipeline, its own verification gates, and its own editorial voice.

This is why we can show you the Skeptic flags on this very page. We're not afraid of transparency — transparency is the product.

A System That Gets Smarter Every Day

Most verification systems are built once and left to rot. Ours evolves.

Every time the Skeptic catches an overreach, it learns what overreach looks like in that context. Every time a human reviewer corrects a flag in Slack, that correction feeds back into the system. Every edge case — a study with unusual methodology, a claim that sits on the boundary between verified and contested — makes the next review sharper.

Six months from now, the Skeptic will catch things it can't catch today. A year from now, it will flag patterns across hundreds of studies that no human reviewer would ever notice. This is the compounding advantage of building verification into the foundation — not bolting it on as an afterthought.

Don't Take Our Word for It

Everything above describes how we work. But we're a methodology page — so we should prove it. Below are the peer-reviewed studies that validate our approach. We ran our own Skeptic Protocol on every one of them, because if we're going to claim our methodology is evidence-based, the evidence itself has to survive our verification system.

This study is the most relevant to FitChef's approach because it used the same AI model we use, in a real-world workflow, with blinded adjudicators creating the reference standard. It wasn't a lab test — it was a workflow validation across six ongoing systematic reviews.

No single study proves our exact pipeline. But the evidence points in one direction: AI extraction plus human verification outperforms either alone. That's the architecture we built.

Why Humans Alone Aren't Enough

Here's something the research community knows but rarely talks about publicly: even trained experts make errors when extracting data from studies. Not occasionally — routinely.

A systematic review by Mathes, Klaßen, and Pieper examined error rates across multiple studies of data extraction in systematic reviews. What they found was striking: error rates ranged from 8% to 50% across studies. Even in gold-standard Cochrane reviews — the most rigorous reviews in medicine — half contained at least one extraction error. And experience didn't help. Reviewers with years of systematic review experience made errors at roughly the same rate as newcomers.

Separate research on cognitive performance shows why. The Whitehall II study, tracking over 2,000 British civil servants, found that sustained long working hours were associated with measurable cognitive decline. The people doing the hardest knowledge work are the most vulnerable to the fatigue that causes errors.

This isn't a criticism of human reviewers. It's a recognition of human biology. We get tired. We miss things. We read "standard deviation" when the paper says "standard error." We transpose digits. And these small errors can cascade through an entire evidence synthesis.

That's exactly why our system combines AI extraction — which never gets tired, never transposes digits, and reads every page with the same attention — with human review that catches the things AI misses: context, nuance, and the judgment calls that require understanding the real world.