Input
Nutrition facts, ingredients, claims, and serving context come in.
Food scoring
How we score food, and what a score can't tell you.
This is the methodology behind the coach: the published, versioned engine under every meal take. In the app you see coach language first; here you see the whole working: the rules, where each one gets its evidence, where the nutrition data came from, and the places our answer is weaker than it looks.
Derived from the scoring engine on 2026-07-19 · engine build 43cf371
The pipeline
From food entry to explainable score.
Route
The item is routed to the right policy: whole food, packaged, restaurant, supplement.
Score
Penalties and credits applied per route. Caps prevent gaming the number.
Explain
Output explains why, not just what. The number carries its evidence.
How a score is built
Everything starts at 100. Then the food argues it down.
Nothing has to earn its way up from zero. Every item opens at 100 points, and each rule that matches moves the number. What is left at the end is the score.
186 rules
A penalty takes points off. Ultra-processing, a heavy sodium load, an additive with a regulatory finding against it.
80 rules
A credit puts points back. Fibre, whole grains, third-party testing, a short and readable ingredient list.
42 rules
An override sets a ceiling the score cannot rise above, no matter how many credits the item earned. It is reserved for safety signals.
A cap is not a big penalty — it is a different thing. A penalty of 25 can be offset by credits. A cap cannot. If a rule sets a ceiling of 45, the item scores at most 45 even if everything else about it is excellent. That is how a genuine safety signal is stopped from being outvoted by good protein.
Rules are also grouped into buckets — processing, nutrients, additives — and each bucket has a limit on how much it can take off in total. One problem that shows up as five related signals should not be counted five times over.
Context matters
Why one rule doesn't fit all.
A banana and a protein bar might have similar carb counts. They shouldn't be judged by the same rubric.
POWR's scoring engine separates whole foods from packaged foods so processing signals, additives, and ingredient quality are weighed appropriately for the item type.
Organic banana
98 Excellent
- Whole food source
- High potassium
- Natural fiber matrix
- No processing
Chocolate chip protein bar
42 Fair
- High protein density
- Contains sucralose
- Palm kernel oil
- Ultra-processed signals
What the number means
Five bands, the same for everything.
All 19 policies share one set of bands, so a 72 means the same kind of thing wherever you see it.
- 85–100 Excellent
- 70–84 Good
- 55–69 Fair
- 40–54 Avoid
- 0–39 Actively harmful
Scores are only comparable inside their own category. A bottle of water and a loaf of bread are scored by different rules against different expectations, so ranking one against the other would be meaningless — and the app does not do it.
What a rule looks at
No black box. Every rule, visible.
Ingredients
Every name on the label is read. Dyes, seed oils, emulsifiers, sweeteners, gums, gels, glazes.
Banned compounds and active recalls set a hard ceiling no good number can lift.
Nutrients
Sugar load, sodium density, fiber-to-carb ratio, protein quality.
Every vitamin and mineral comes back with the verdict, not buried inside it.
Processing
How much human work happened between the field and the package.
Whole foods start generous. Ultra-processed signals stack quickly.
Source
Where the food came from and how it was raised, when the label can prove it.
Grass-fed, wild-caught, organic, and third-party tested earn real credit.
Contamination
Heavy metals, PFAS, mercury class for seafood, the things you'd never want on the plate.
A real safety signal forces a ceiling. No other credit can lift it.
Confidence
How sure we are sits next to the score, never folded into it.
A verified label beats OCR. OCR beats a visual guess.
Routing
19 policies. An item is scored by one of them.
A piece of fruit, a boxed cereal, a takeaway burrito and a water filter have almost nothing in common. Each kind of item routes to the policy written for it, and only that policy's rules run.
Item Routed to one policy That policy's rules run Score + reasons
The reasons come out with the number, every time. A score with no visible reasons behind it is not something we will show you, and it is not something we will publish.
Reference
All 19 policies, three families.
6 are entry routes for how a food shows up. 10 are food subcategories with rules tuned to that class. 3 are water. All run on the same engine.
Entry routes · 6
Core concept
An apple, a salmon fillet, raw oats. One real food, no ingredient list.
When to use
When the thing on the plate doesn't have a label to read.
Technical focus
Whole foods start from a generous baseline. The score weighs nutrient density and what the food actually brings to the day. Real harm signals still drop the score; nothing about being whole excuses what's in it.
A banana should not be judged like a protein bar.
Core concept
Anything with a barcode and a nutrition label. The clearest evidence we get.
When to use
When there's a nutrition label, ingredient list, or barcode-backed product record.
Technical focus
Penalties stack across ingredients, additives, nutrients, and processing. Caps stop a high-protein number from rescuing a heavily processed product. Banned ingredients, classified carcinogens, and active recalls force a hard ceiling no credit can override.
Where the most goes into the score because the most is on the label.
Core concept
Restaurant menus and fast food. Two policies, described together because they share an evidence problem: lower confidence, on purpose.
When to use
Sit-down menus, chain nutrition info, or anything from a drive-through.
Technical focus
More conservative than packaged because the data is. The classic fast-food combo, deep-fried, processed meat, heavy sodium, gets a hard ceiling no credit can lift. Confidence is shown next to the score, never inside it.
Don't fake exactness when the evidence is thin.
Core concept
Something you cooked from a few different things, where the exact oils and amounts are guessable but not certain.
When to use
When the big components are obvious but the seasoning and portions aren't.
Technical focus
We score the parts we can identify, then weight them by how much of each is on the plate. Confidence reflects how much we actually know, not how much we'd like to.
Useful guidance without faking precision.
Core concept
Powders, capsules, gummies, electrolyte mixes. They get their own rules so a vitamin doesn't get judged like a snack.
When to use
Anything you'd find in the supplement aisle.
Technical focus
Looks at sweeteners, sugar, excipients, additives, and how honest the label is. A protein powder and a multivitamin don't follow the same checklist.
A supplement score doesn't mean the same thing as a food score.
Food subcategories · 10
Routed to from the entry layer when an item matches a category with its own evidence axes.
Meat and seafood
Raw or prepared. Processed meats take a real hit.
Produce
Fruits and vegetables. Whole-food rules apply.
Eggs
Chicken, duck, quail. The yolk is treated as the food, not the enemy.
Dairy products
Cheese, yogurt, butter, cream. Fermented dairy gets credit where it earns it.
Milk
Cow, goat, sheep. The kind of milk and how it's processed both matter.
Plant-based milk
Almond, oat, soy. Strict on added sweeteners.
Drinks and beverages
Sodas, juices, kombucha, energy drinks. Total sugar counts, not just added.
Tea products
Loose, bag, bottled. Independent heavy-metal testing earns credit; sweetened bottled tea takes a hit.
Sweeteners
Ranked from whole-food sweeteners down to the artificial ones we'd avoid.
Bread products
Whole grains earn credit. Refined flour costs you. Bromated flour stops the score cold.
Water domain · 3
A separate domain so water scores are never compared to food scores.
Bottled water
Contaminants, materials, source, certifications. We don't reward water for being premium; we flag harm.
Tap water
Municipal and well water. Detected lead or PFAS forces a hard ceiling.
Water filters
Filters are a tool. We grade them as one.
The evidence standard
148 of our 308 rules cite a named source.
148 cite a source 160 do not yet
What a cited rule means
The rule points at a document somebody else published — a regulator's finding or a peer-reviewed paper — recorded with its authority, its title, an identifier such as a DOI or a CFR section, and the date we last checked that it says what we claim it says.
Across those rules there are 185 sources on file. 119 of them are peer-reviewed papers cited by DOI, so you can pull up the same paper we read.
What an uncited rule means
It means the rule is our judgement, not an established finding — and we would rather say that than round the number up. Some of these are waiting on a source we have not finished verifying. Others are genuinely our editorial position: a threshold we picked, or a signal we think matters.
We are working through the rest. We will not fill the gap by attaching a source that does not really support the rule; a citation that looks like substantiation without being it is worse than an honest blank.
This has teeth. An uncited rule can still affect the score you see privately in the app. It cannot leave it. Before POWR states anything publicly about a named product, every reason shown alongside the score must carry a source that resolves — and if one does not, the whole claim is refused rather than trimmed to the part that passes.
Coverage, policy by policy
Where we are thin.
The average hides a lot. Some policies are nearly fully sourced; others have barely started. Weakest first.
| Policy | Rules | Cited | Coverage |
|---|---|---|---|
| Single Whole Food | 15 | 2 | 13% |
| Produce | 11 | 2 | 18% |
| Meat and Seafood | 19 | 4 | 21% |
| Supplement | 18 | 4 | 22% |
| Milk | 14 | 4 | 29% |
| Water Filter | 6 | 2 | 33% |
| Plant-Based Milk | 18 | 6 | 33% |
| Eggs | 10 | 4 | 40% |
| Homemade Mixed Meal | 20 | 8 | 40% |
| Sweeteners | 13 | 6 | 46% |
| Dairy Products | 16 | 8 | 50% |
| Bread Products | 20 | 11 | 55% |
| Tea Products | 15 | 9 | 60% |
| Drinks and Beverages | 16 | 10 | 63% |
| Fast Food | 20 | 13 | 65% |
| Restaurant Menu Item | 20 | 13 | 65% |
| Packaged Food | 33 | 22 | 67% |
| Bottled Water | 13 | 9 | 69% |
| Tap Water | 11 | 11 | 100% |
Provenance
Where the numbers actually came from.
A rule is only as good as the nutrition data it runs on. So we record how strongly the identity of each item is established, on four rungs.
Verified
StrongestA scanned barcode identifies the exact product. The identifier is machine-readable, so there is nothing to second-guess about what the thing is.
Strong enough to publish — but nothing reaches this rung today. See below.
Disclosed
StrongThe numbers come from a published source we can point at. Two things reach this rung: a manufacturer’s printed label, transcribed and republished by Open Food Facts at a stable address, and a chain restaurant’s federally-mandated menu disclosure under 21 CFR 101.11.
Strong enough to publish, with the source named.
Declared
WeakA person asserted the identity. You picked this item from a list before, or the name and brand matched a catalogue entry exactly. Probably the right food. Nothing machine-proven.
Not published as a claim about a named product.
Inferred
WeakestThe system guessed. A close-but-not-exact name match, a habit of yours, or a photo estimate. This is not evidence of identity.
Not published as a claim about a named product.
Nothing reaches Verified end to end today, and we would rather say so. The strongest rung needs two separate things to be true: a scan must prove which product this is, and the nutrition numbers must come from a source we can point at. A barcode does the first. Nothing we currently ingest does the second in the strongest form, because no manufacturer publishes its nutrition panel to us directly — we read a transcription of the printed label.
So a barcode scan onto a well-sourced record lands at Disclosed, not Verified, and that is what we call it. A document attests where numbers were published. Only a scanned identifier proves what a thing is. They are different kinds of evidence and collapsing them would be an overstatement, so the weaker of the two always wins.
Limits
What a POWR score is not.
A methodology page that only lists strengths is an advert. These are the real edges of what we can tell you.
We are reading a nutrition panel, not a recipe
The federal menu-labeling rule, 21 CFR 101.11, compels a nutrition panel for standard menu items — calories, fat, sodium, the macros. It does not compel an ingredient list. Packaged food gets that under a different rule; menu items do not. The rule also binds only chains with 20 or more locations, so an independent restaurant is compelled to disclose nothing at all.
So every judgement we make about what is in a restaurant dish — the oil it was fried in, the kind of meat, the flour — is an inference about a recipe nobody published. Only the calorie, macro and sodium rules rest on data anyone was required to disclose.
A score is a snapshot, not a standing verdict
A score is calculated from the ingredients and nutrition data available at the moment the item is analysed, under the methodology version in force that day.
Recipes get reformulated and labels get reprinted. A score you saw last year may not describe what is in the package today.
We do not test products, and we do not originate facts
POWR runs no laboratory. We do not assay products for contaminants, verify that a label is accurate, or investigate how a company operates.
Everything we score is somebody else’s published information — a label, a disclosure, a regulator’s finding, a study. Our contribution is the weighting and the resulting opinion, and we say so.
Some rules are our opinion, and are labelled that way
160 of our 308 rules do not yet carry a citation. Some are waiting on a source we have not finished verifying. Others are genuinely our own editorial position — a threshold we chose, or a signal we treat as meaningful because we think it is.
Where a rule is our judgement rather than an established finding, the rule’s own text says so in the app. We would rather disclose an opinion as an opinion than dress it as a consensus that does not exist.
A score is our opinion about publicly available information on a product. It is not medical advice, not a safety certification, and not a statement about any company's conduct.
Corrections
Think a score is wrong? Tell us.
This is open to anyone, and it is the route we would ask a manufacturer to use. If we have your product's ingredients, nutrition data, or formulation wrong, we want to know — a score built on stale or incorrect data is a defect, and we would rather fix it than defend it.
- 1
Email us and say which item and which reason. The specific driver you disagree with is the most useful thing you can send. If you have the current spec sheet, label or published nutrition data, include it.
- 2
We preserve the record before we touch anything. The score, the reasons behind it, their sources, the methodology version and where the data came from are all kept, so there is something concrete to check the complaint against.
- 3
We re-run the item against our own checks. If any of them fails — wrong data, a rule misapplied, a source that does not support what we said — that is our error. We correct it promptly, and we say what changed.
- 4
If the checks all pass, the score stands, and we will show you the working. A score is an opinion drawn from information you published, with the sources attached. We will always explain it, and we will never characterise a company — only the formulation.
Use the subject line “Score challenge” so we can pick it out quickly. It is the same inbox as everything else, read by the people who build POWR, and we answer within 24 to 48 hours.