How We Score
The whole method, including the parts that make us look smaller than we could pretend to be. If you are wondering how this differs from asking a chatbot, this page is the answer.
In one sentence
- 1Your list is matched.
Everything you enter is matched against the curated dictionary — 378 supplements, 150 medication and drug names, 108 brand names. Anything that matches nothing is carried through and reported back to you as unrated, by name.
- 2The AI writes a draft.
A language model writes a first-pass analysis: what each item is, what is worth watching, and a starting score out of 10. It has no web access, no browsing tool and no database connection. It sees your list and the curated data we hand it, and nothing else.
- 3The deterministic floors run.
21 floors — the 20 curated interaction rules plus the serotonin-syndrome floor — run over the items you actually entered, not over the model’s prose. Each one that fires forces a minimum risk level, caps the score at a fixed ceiling, and inserts its own written warning. They escalate only: risk can rise, scores can be capped down, warnings can be added. Nothing here runs the other way.
- 4Practical guidance is attached.
8 absorption-spacing rules, an upper-limit tally across 9 nutrients, and 4 pairing notes. This layer never moves the score in either direction.
- 5Coherence passes clean up after the model.
16 of the 20 rules cover risks that spacing cannot mitigate — additive bleeding, serotonergic load, CNS depression, hyperkalemia, enzyme induction. For those, any "space these apart" advice is stripped out, because a separation window would imply there is a safe way to combine something this same report calls dangerous. These passes can only remove a suggestion or raise a severity badge.
No pass in the engine softens a result
There used to be one. It lowered the risk level and raised the score on supplement-only stacks it judged over-rated, which contradicted the promise on this page — so it was deleted rather than tuned. The escalate-only rule is enforced in code on every path: a fresh analysis, a cached result, and the pre-generated stack pages all run the identical passes.
The 20 curated interaction rules
- • blood thinner + supplement
- • potassium + ACE inhibitor/ARB
- • potassium + potassium-sparing diuretic
- • MAOI + stimulant
- • multiple stimulants
- • statin + red yeast rice
- • lithium + interacting drug
- • levothyroxine + minerals
- • antibiotic + minerals
- • CNS depressant combination
- • liver-stress combination
- • diabetes drug + glucose-lowering supplement
- • SSRI/SNRI + bleeding-risk supplement
- • levodopa + iron/vitamin B6
- • transplant immunosuppressant + St. John’s Wort
- • digoxin + licorice/hawthorn/St. John’s Wort
- • warfarin + INR-shifting supplement
- • oral bisphosphonate + minerals
- • hormonal contraceptive + St. John’s Wort
- • nitrate/PDE5 medication + nitric-oxide-booster supplement
Plus the serotonin-syndrome floor, which fires whenever two or more of 12 distinct serotonergic classes appear in one list, whether they are prescriptions or supplements.
We do not query third-party medical websites
- Our curated interaction rule set
20 hand-written rules plus a serotonin-syndrome floor spanning 12 serotonergic drug and supplement classes. Each rule states which two classes trigger it, the minimum risk level it forces, the score ceiling it imposes, and the warning text it inserts. The mechanism and its citations — PMIDs, prescribing information, NIH and NCCIH monographs — sit in the source comment beside every rule.
- Our compound dictionary
378 supplement entries with their alternate spellings, 150 medication and drug names, and 108 brand names mapped to their generic. This is what turns what you typed into something a rule can match.
- NIH Dietary Supplement Label Database (DSLD)
The one live external lookup. Queried over its public API for real product labels and cached for a day. It is where the milligram tally gets the amounts actually printed on your bottles, and where a scanned photo is matched to a real product.
- NIH Office of Dietary Supplements fact sheets
Read by a human and distilled into the 8 absorption-spacing rules, the 9 nutrients we tally against published upper limits, and the 4 pairing notes. We link the fact sheet wherever the number came from one.
Every number here is our coverage, not a census
378, 150, 108, 20, 12, 9, 8 — each of those counts what is in our files today. None of them is a count of everything that has ever been documented about supplement and medication interactions. Nobody has that number, and we are not going to imply we do.
Every result carries a score from 1 to 10, where 10 is safest. One shared function turns that score into a risk label, so the same score can never read as two different labels on two different screens. The bands below are generated from that function when this page is built, so they cannot drift from what the app does.
No curated rule fired, and nothing in the combination raised a flag we hold a rule for.
Something is worth knowing about — an absorption conflict, an upper-limit overlap, a real but non-emergency interaction — or our coverage of these particular items is thin.
A curated rule fired on a well-documented, clinically serious mechanism. Every high result carries the written warning that produced it.
One-way escalation
A label may be escalated above its band and is never softened below it. If a rule states a risk more severe than the score band implies, the more severe label is what you see. The reverse cannot happen: a comfortable-looking score can never talk a warning down. Where a result has no score at all, the label falls back to medium — the conservative default, not the flattering one.
Where we suggest an ingredient rather than warn about one, it carries a categorical evidence grade. Categorical on purpose — never a percentage, a star rating or a confidence number, because those imply a precision the underlying literature does not have. There are 121 graded entries today.
Repeated human trials pointing the same way, with a major public reference or monograph agreeing.
Human trials exist and point the same way, but the set is small, short, or mixed on how large the effect is.
Preliminary or indirect evidence, or a case that rests mainly on a nutrient’s established role rather than on trials of supplementing it.
An active research area with early human data. Shown as weak rather than left out.
A grade cannot ship without its source: the citation link is a required field on every graded entry, so the code will not compile if someone adds a grade with nothing behind it. Weak evidence is shown as weak and left in view rather than quietly dropped, and a grade describes the strength of the evidence only — it is never a promise about what will happen to you.
Silence is not a clearance
Silence is not a clearance: an unrated ingredient means we hold no rule for it, not that nothing happens. The checker names those items back to you rather than letting a quiet result read as a verdict.
Drug-versus-drug checking needs a licensed clinical compendium, which we do not have and do not pretend to have. Your pharmacist has one, will run it against your whole prescription list for free, and is the right person to ask.
The dictionary and rule set are hand-maintained and finite. A compound nobody has added yet gets no rule of its own, however real its interactions are.
No PubMed connection, no clinical-trial feed, no FDA API, no proprietary drug compendium. What we have is a hand-curated file and one public label database.
Rules and dictionary entries are edited by a human on a review cadence, so newly published findings are not reflected the day they appear.
Where a curated rule covers a combination, the rule overrides whatever the model said. Where no rule covers it, the model’s own conservative draft is what you get. That is the honest weak spot of this product, and it is why an unrated item is named rather than quietly scored past.
No labs, no kidney or liver function, no dosing history, and nothing about your circumstances beyond what you tell us. A saved health profile narrows this gap; it does not close it.
If something here is wrong, we would rather hear it than not. Errors are corrected in the source file by hand, and because the deterministic passes run on every serve path — fresh analyses, cached results and the pre-generated stack pages alike — a corrected rule changes what everyone sees without anyone having to re-run their check.
What to include
- • The exact items you entered, and which screen showed the problem
- • What you believe is wrong — a missing warning, a warning that should not be there, a wrong amount or limit
- • A source if you have one. A fact sheet or a PMID is ideal; a clear description is fine too
Where to send it
Use the contact form or email hello@canistackthis.com. Both land in the same inbox — the form posts to it directly.
A report that turns out to be right gets the file changed. A report that turns out to be wrong gets an answer explaining why, not silence.
Readers pay for this, and nobody else does. The only money coming in is from Premium subscriptions and the one-time DNA report. There are no ads, no supplements for sale, no affiliate links, no sponsored placements, and no paid inclusion in any rule, suggestion or score.
Nothing in a score can be bought, because there is nothing to buy from us at the end of a warning.
Warnings are never held back, blurred, delayed or used as a reason to upgrade. Safety content is shown in full on the free tier.
A feature launched free stays free. We can widen the free tier, and we do not narrow it.
CanIStackThis is an informational tool. It is not medical advice, it is not a diagnosis, and it is not a substitute for your doctor or your pharmacist. It does not know your labs, your history, your kidney or liver function, or your doses. Use it to arrive at that conversation with better questions — that is what it is for.
Every score on this site is produced by the method above. For the wider tour of what the engine reads and which tools build on it, see How It Works.