This is a source-grounded appraisal of the nutrition, exercise, and lifestyle evidence: what an exposure does, for whom, how much, and how confident we should be — with every claim traced back to the study or guideline it came from.
Everything here is built around one test: would this change what someone actually does — on an outcome they care about — in a way they couldn’t already have worked out? Being true isn’t enough; a finding has to be useful and not already obvious to earn a place. That’s also why some topics that fill a lot of column inches elsewhere get a single line here: when the honest answer is “it barely moves the needle,” that is the answer, and it doesn’t need a page.
This page has two halves: what the guide is and how it judges, then how it is actually built and maintained — because a big part of trusting it is understanding how the pages came to be. For the narrower question of how far to trust it — where judgment enters, and what it deliberately does not do — see Can you trust this guide?
Part 1 — What it is, and how it judges
What makes it different
- Ranked by effect size × certainty — big rocks first. Most of the achievable benefit sits in a handful of large exposures (not smoking, staying active, avoiding excess visceral fat, sleeping enough). The endless refinements share what’s left. We say so plainly instead of burying the large levers among small ones.
- Honest about uncertainty. We keep four states apart — benefit, harm, no meaningful effect, and not enough evidence to say — and never let the last two blur together. Where the evidence is thin, we say the evidence is thin.
- Surrogates are not outcomes. A number that moves (cholesterol, blood sugar) is not the same as a life lived longer or better. We flag when a result rests on a marker rather than on what you actually care about.
- Symmetric standards. A long-standing recommendation earns no free pass from age or authority; a contrarian claim earns none from novelty. The same scrutiny applies to both.
- Gaps are findings. When the levers left for you are small and uncertain, that is the answer — it tells you it’s okay to stop optimizing.
What it covers
The scope is modifiable exposures — the things you can actually change — judged mostly by what they do to physical health: how long you live, how well you age, cardiometabolic and musculoskeletal function, and avoiding deficiency.
- The core: food and whole dietary patterns, physical activity and training, and sleep. This is where the large levers live, so it’s where most of the attention goes.
- The periphery, held to the same bar: other lifestyle levers — heat and sauna, cold, sun and light, stress and how it’s managed — and the trade-offs inside an exercise programme (load, recovery, injury). These earn a page when the evidence bears on a real outcome, not because they’re in fashion; most turn out to be small-effect and stay short.
- Drugs, where they’re the realistic alternative. A medication — a GLP-1 agonist, say — is often the honest comparison to a lifestyle change, and leaving it out would make the comparison unfair. We appraise what it does; we don’t tell you to take it.
- What’s left out: claims that can’t be tested, or that have been measured and come up empty in humans. A practice earns a place only through a specific, checkable claim about a human outcome — never as an identity, a tradition, or a philosophy on its own.
What evidence it leans on
Not all evidence is equal, and we try not to pretend otherwise.
- Preferred, near the top: systematic reviews, meta-analyses, and umbrella reviews (reviews of the reviews), plus the major guideline and consensus statements built on them. Pooling many studies is harder to fool than any single striking result.
- Then: individual randomized trials, large long-term cohorts, and natural experiments — including Mendelian randomization, which uses an accident of genetics as a kind of built-in randomization.
- Fit-to-question over pedigree. The right design depends on the question. Nutrition rarely allows the clean, decades-long, blinded trial you’d want (you can’t blind a food or randomize a lifetime), so instead of demanding one perfect study we triangulate — reading trials, observational data, dose-response, and biological mechanism together, and weighing them against each other.
- Mechanism can point a direction — it isn’t proof. A plausible biological pathway with some human support can tell us which way an effect probably runs; we mark it as such and note what would confirm or overturn it. A result in a petri dish or a mouse is a candidate, not a conclusion.
- The caveat that shadows the whole field: people misreport what they eat, by a lot. That measurement error is big enough to wash out real effects, so a flat or null dietary finding often means “we couldn’t measure it well enough,” not “there’s nothing there.” Where that’s likely, we say so.
The honest limits
- This appraises evidence; it does not prescribe. It says what the evidence shows for a group of people. Selecting, dosing, screening for contraindications, and managing your medications are things only a clinician who knows you can do. This is not medical advice.
- A group is not a person. Every finding is about a reference class. You may differ from it in ways that matter — which is exactly the judgment that belongs to you and your clinician, not to a page.
- The loop is open. This grades whether the reasoning is internally sound and faithful to its sources — not whether following it made anyone better off. A clean, well-sourced page can still be wrong about the world. Nothing here has been tested against real outcomes, and we don’t pretend otherwise.
Part 2 — How it is built and maintained
Knowing how the pages are produced is part of being able to trust them. Here is the machinery, in plain terms — the information flow first, then each stage.
flowchart LR Q["Human curator<br/>selects, poses questions, approves"] SRC[("Source<br/>literature")] ING["Ingest<br/>extract claims"] WEA["Weave<br/>link by question"] CHK{"Source-grounding<br/>checks (gate)"} TENS["Tension<br/>conflict filed, not flattened"] GRAPH[("Interlinked<br/>claim graph")] COM["Compile<br/>cut a deliverable"] PUB["Reader-facing<br/>topic pages"] Q --> SRC --> ING --> WEA --> CHK CHK -->|passes| GRAPH CHK -. "fix / demote" .-> ING WEA -. conflict .-> TENS --> GRAPH GRAPH --> COM --> PUB Q -. approves .-> PUB
Two things the diagram deliberately shows. The source-grounding checks are a gate, not an afterthought — nothing enters the graph until it passes, and the same checks run again when a deliverable is compiled. And there is no arrow back from real-world outcomes: nothing here has been tested against whether the advice actually helped anyone. That open loop is the central limit, spelled out on the trust page.
Built as an LLM wiki
This guide is written and maintained differently from most. It is an LLM wiki — an idea suggested by Andrej Karpathy: instead of a person composing articles, a large language model (Claude) reads the source literature and maintains an interlinked knowledge graph, under a human curator’s direction. The two roles are kept distinct:
- The human curates. A single maintainer chooses what to read, poses the decision-questions, reviews the output, challenges it, and approves every change. The selection of questions and sources — and the final judgment — are human.
- The model does the reading and the bookkeeping. It extracts what bears on a decision, integrates it with what is already held, writes the cross-links, and runs the consistency checks — the labour that would make a hand-maintained graph of this size impossible for one person.
This is a genuine strength (breadth and consistency one person could not sustain) and a genuine risk (a fluent model can write a plausible sentence its source does not support). The mechanisms below exist specifically to hold that risk down, and the trust page is candid about what they cannot fix.
How a page comes to be
Pages are not written top-to-bottom; they accrete from sources through a few repeated operations:
- Ingest — a source is read in full and the decision-relevant parts are extracted, each tagged with where it came from. We keep what changes reasoning about what to do — not a summary of the paper.
- Weave — the extract is folded into the graph and linked to other pages by the decision-question it answers, not by topical similarity. Where it clashes with something already held, the clash is confronted before anything is reconciled.
- Compile — the reader-facing topic pages (the cards on the home page) are cut from that graph, with the underlying tensions resolved or named, into a standalone piece.
How it stays source-grounded and consistent
The point of the discipline is that every non-obvious claim either traces to a source or is marked as the guide’s own reasoning — and that this is mechanically enforced, not merely intended:
- Provenance is tagged on the claim. An extracted claim carries its source; the guide’s own inference is labelled as inference. You can always tell which is which.
- Quotations are pulled from the source by tool, never retyped. A quote is lifted directly from the source text so it cannot drift, get truncated before a qualifier, or be misattributed from memory.
- A blind re-read checks every new claim. After a page is written, a separate pass re-reads each attributed claim against only its cited source, with no other context, to catch anything the source does not actually say. Overstatements are demoted on the spot.
- Pages are checked against each other. Automated checks flag internal contradictions, claims that overstate their evidence, orphaned assertions with no source, and a claim retracted on one page but left standing on another.
None of this grades whether a claim is true — only whether it is faithful to its source and consistent with the rest. That distinction is the single most important thing on the trust page.
How disagreement between sources is handled
When good sources genuinely conflict, we do not quietly pick a winner:
- The conflict is filed, not flattened. Each side is stated in its own terms — the actual point of disagreement located — and the page shows why the evidence does not settle it, rather than averaging it into a mush.
- When guideline bodies disagree with each other, that is itself the finding. It means the evidence does not determine the answer, and the guide’s job is to show that, not to referee it.
- Before a “these disagree” claim is made, we check that the two sources are answering the same question at the same scope — apparent contradictions are often two true statements about different things.
How source quality is weighed
Every source carries a tier — from gold (a systematic review, umbrella review, or guideline with a documented evidence base) down through high, moderate, weak, to mechanism (animal or test-tube). A weak source is labelled, not deleted — in a thin field a clearly-marked weak source beats silence — but it cannot quietly prop up a confident claim. How the tiers are assigned, and an honest account of how sources were selected (including where selection was not systematic), are on the trust page.
For where judgment enters, who maintains it, the calibrated language, and how to challenge it, see Can you trust this guide?