How Distil works.

Public evidence-review protocol.

Most supplement companies do not publish their methodology because hiding the reasoning is how they sell more. Distil's commercial position is the opposite of that. We earn when you are informed, not confused. So this page exists.

What follows is the actual rulebook: the compound database, the grading system, the exclusions that are off the table by rule, the personalisation pipeline, the independent safety check, and the limitations of what an evidence-graded report can and cannot do. All of it checkable.

The compound database.

Distil's database holds 104 compounds we evaluate for inclusion in personalised stacks, plus a set we explicitly track as "considered but not recommended" so the interactions checker can still flag them (kava, St John's Wort, 5-HTP, high-dose biotin, niacin, vitamin K2 supplements, yohimbe, echinacea, astragalus, ginger, SAMe, L-tryptophan, red yeast rice, cinnamon, lithium orotate). Each links to a short page explaining why it is left out and what to consider instead. There are thousands of supplements on the market. Most do not have enough peer-reviewed clinical evidence behind them to justify a recommendation. The 104 are the ones that do.

Every compound is graded A, B, or C against the published research, by the rules below. Anything that does not reach Grade C is not in the database. Anything that drops below Grade C in a future review is removed. The database is a curated set, not a complete index.

The database is not fixed.

Evidence moves, so the database moves with it. We review continuously, not once a year: every quarter as a full sweep, and immediately whenever a major trial or meta-analysis lands. When a compound accumulates enough good human evidence to clear the Grade C bar, randomised controlled trials or meta-analyses in people rather than promising lab or animal work, it goes into the database at the next review rather than waiting. When the evidence for something weakens, it moves down a grade or out. A compound is only ever added if it clears both a benefit bar and a safety bar: meaningful benefit, and a safety profile that holds up. Low or mixed evidence, or potential harm for little benefit, means it stays out until the evidence is there. This page reflects the database as it stands today, and the figure above will change as the research does.

What the grades mean.

Grade A. Multiple high-quality randomised controlled trials, or one or more meta-analyses, demonstrating the claimed benefit in human populations relevant to who might take it. The effect size is meaningful, the evidence has been replicated, and the finding is consistent across independent research groups. Example: Magnesium for sleep onset, Omega-3 for cardiovascular risk markers, Vitamin D for deficiency-corrected bone health.

Grade B. Solid clinical evidence, typically several human RCTs with consistent direction but smaller sample sizes, or one strong RCT supported by clear mechanistic understanding. Less fully replicated than Grade A but well-supported. Example: L-Theanine for stress-related arousal, Saffron for mild-to-moderate low mood, Ashwagandha (KSM-66) for sleep quality and subjective stress.

Grade C. Emerging evidence. A small number of human RCTs, or strong mechanistic and animal-model evidence with limited human replication. Reported as emerging in every report it appears in. Always flagged so the reader can decide whether the evidence threshold is enough for them. Example: NMN/NR for NAD+ raising in over-45s, Fisetin for senolytic activity, Urolithin A for mitochondrial function.

Grade D. Never recommended. Used in the database internally to mark compounds we have considered and rejected, so the same decision does not get re-litigated each review. Compounds in this category have either failed to replicate, shown null results in well-powered trials, or carry a safety profile that makes the benefit-to-risk ratio untenable. Grade D compounds are listed in the database with the reason, but they never appear in a recommendation.

Mixed grades use the lower badge.

A compound with Grade A evidence for one outcome and Grade B evidence for another is treated as Grade B in any report where the Grade B claim is the relevant one. The reader sees the lower of the two grades, never the higher. This is a deliberate under-claim rather than an over-claim.

Hard exclusions, off the table by rule.

Some compounds are excluded for specific clients regardless of how well they would otherwise score. The exclusions are absolute, not "use with caution":

• 5-HTP on SSRI, SNRI, or MAOI medication. Serotonin syndrome risk. Hard exclusion, no exceptions.
• NMN or NR with any cancer history. NAD+ raising compounds may support cancer cell proliferation. Hard exclusion.
• Iron without confirmed ferritin in the deficient range. Iron supplementation in a non-deficient person carries real oxidative-stress risk. We do not recommend iron unless a blood result confirms the need.
• High-dose biotin (over 5,000 mcg). Always flagged for interference with thyroid function tests and troponin assays used in cardiac diagnostics.
• St John's Wort. Grade D. Never recommended due to CYP3A4 induction interfering with oral contraceptives, SSRIs, warfarin, statins, and immunosuppressants.
• Cancer treatment in progress. No supplement stack is produced. The intake form gates this answer before any payment is taken.
• Pregnancy or breastfeeding. Same. Defer to NHS antenatal prescribing and your midwife or health visitor.

There are more. The full list lives in the database itself. The principle is consistent: the report does not produce output we believe is unsafe for the specific client, even if the client wants it. The answer is sometimes "not for you, not yet, here is what to test first".

How a compound earns its place.

A compound enters the database only after a structured evidence review. The protocol:

One. A literature search across PubMed: meta-analyses and systematic reviews first, then primary RCTs in the relevant outcome and population. Mechanistic-only evidence does not qualify a compound for inclusion.

Two. Each citation supporting the entry must have a verified PubMed identifier (PMID). Citations written from general knowledge alone are not acceptable. The PMID is the auditable trace.

Three. A grade is assigned by the rules above, against the evidence. Grade is determined by what the published research actually supports, not by what the compound is marketed for.

Four. Population modifiers are documented. A compound that works for older adults may not work for younger ones. Effective doses for women may differ from doses for men. The database holds these qualifiers explicitly so the personalisation step can apply them.

Five. Drug interactions are checked against the master table and added to the entry. Any interaction that would make the compound dangerous on a common medication is written into the exclusion logic.

Six. A "Last PubMed review" date is stamped on the entry.

The database is recalibrated every quarter. The cadence is enforced by an automated forcing function: on the first day of January, April, July, and October, a GitHub workflow opens a tracked issue requiring the calibration to be completed before the issue closes. The full log of every calibration appears at the bottom of this page.

The personalisation pipeline.

A report is not a search of the database. It is a personalised stack built for one specific person, against their full health profile. The pipeline runs in five stages:

Compound scoring. Every compound in the database is scored against the client's profile: their goals, their conditions, their medications, their diet, their sleep, their bloodwork, their lifestyle, the supplements they already take. Most compounds do not score high enough to be included. Selectivity is the product.

Clinical safety cull. A second pass removes any compound that survived the first pass but fails clinical justification on closer inspection. This includes medication interactions, contraindicated conditions, and risks specific to the client's age or population.

Dose-locking. For every compound that survives both passes, the personalised dose is fixed, against the client's specific conditions and medications. Anticoagulants cap doses of certain compounds. CKD changes thresholds. Pregnancy (where the report is generated, not gated) constrains forms. Dose-locking is the safety-critical step.

Personalised writing. The report itself is written section by section, against the locked manifest. The same compound recommendation appears identically in the introduction schedule, the daily timing table, the compound card, the interactions summary, and the sleep stack where relevant. Cross-section drift is engineered out by architecture, not editorial review.

Independent safety review. A second model, separate from the one that wrote the report, reviews the output against a full safety checklist: dose caps, exclusion rules, medication flags, language standards, missing GP review notices, banned-claim language. The check is fail-closed: any critical failure holds the report for manual review and prevents delivery.

The independent safety check.

The safety check is the most important step. It runs as a separate model call after the report is generated. It does not write the report; it audits it. If anything fails, the report does not deliver. It is held until reviewed.

A failure can mean a dose that exceeds a safety cap. A compound that should have been excluded for a stated condition. A claim that crosses from "may support" into "will". A missing GP review notice on a compound that requires one. An interaction that was not flagged.

When a report holds, an alert reaches Sebastian directly. The report is reviewed, rebuilt if necessary, and only delivered when it passes. This has happened. Reports that were technically deliverable but visibly imperfect have been rebuilt rather than shipped.

The QC stage is the reason Distil reports can be trusted. It is the difference between an LLM-generated supplement guide and a clinical-grade artefact.

What Distil deliberately does not do.

The boundaries are as important as the methodology:

We do not sell supplements. We do not stock supplements. We do not have an associated brand of supplements.

We do not take affiliate commission from any retailer.

We do not recommend specific brands. The reports tell you what the product needs to contain (the form, the dose, the purity certifications, the things to avoid). You buy it from whoever you trust.

We do not run a subscription. The report is a one-time purchase. If your situation changes, you can buy a refresh. We do not auto-renew, and we do not email you reminders to come back.

We do not include affiliate-bait compounds purely because they sell well. The selection is dictated by the evidence score, not by what would generate the highest cart value.

Naming what we do not do is part of the brand. Most supplement companies cannot honestly make these statements. Saying it directly is the point.

What this is, and is not.

Distil reports are personalised, evidence-graded information. They are not medical advice.

A Distil report does not replace a GP, a registered nutritionist, or a specialist. It does not diagnose or treat any condition. Where the report identifies that a condition or medication interacts with a compound, the report flags it and recommends discussion with your GP. Where bloodwork is the right next step before supplementation, the report says that.

Compounds are graded against published clinical research. That research has limitations. Sample sizes vary. Effect sizes vary. Replicability varies. The grade reflects what the evidence supports, but no Grade A claim is "this will work for you specifically". It is "the research suggests this tends to support this outcome in populations relevant to your profile".

Honest framing is part of the brand. Anyone who promises more than this is selling something else.

Reassessment, and what comes next.

The report you receive is built against the profile you submit. Profiles change. Conditions emerge. Medications start and stop. Children happen. Goals shift. The report includes a 12-week reassessment framework so you know what to track and when to revisit.

Returning customers within six months pay £49 for a refreshed report (the same as the founder pricing rate); after six months it is the standard rate. There is no auto-renewal and no reminder email. You decide when, or whether, to come back.

Two changes are coming:

A named clinical reviewer. An Association-for-Nutrition registered nutritionist will join as the named clinical reviewer of every Distil report. Their qualifications, registration number, and review responsibilities will be on this page when the relationship is signed.

Continued public log. Every quarterly calibration is documented in the calibration log below. The cadence is enforced automatically: a GitHub workflow on the first of each calibration month opens an issue that stays open until the calibration is complete. If a quarter passes without a new entry, it is publicly visible here.

Calibration log.

Every quarterly calibration of the compound database is logged here. The most recent calibration is at the top.

What this tool is, and is not.

The Distil interactions checker is a free reference for adults who take supplements alongside prescription or over-the-counter medications. You enter your supplements and your medications. We tell you which pairs interact, how seriously, and what the published action is. We list our reasoning per pair and link the primary citation.

It is not a medical device. It is not personalised. It does not store your inputs. It does not replace clinical advice.

The tool exists because the alternative reference resources have gaps. The British National Formulary, the UK clinical gold standard, restricts its scope to conventional medicines and only mentions supplements where they affect conventional therapy. The mass-market checkers cover supplements broadly but are advertising-supported, do not publish their inclusion rules, and tend to return a soft "no interactions found" message that does not distinguish "we checked and there is no interaction" from "we have not checked this combination."

We publish our rules. We separate those two cases. The sections that follow explain how.

How we built the medication list.

We check supplement interactions against 296 medications. The list comes from two sources.

The bulk: the most-dispensed UK drugs. Most of the list comes from the most-dispensed drugs in the NHS Business Services Authority Prescription Cost Analysis for England 2024 to 2025. The top of that list is atorvastatin (73 million prescriptions in a year), amlodipine, the proton pump inhibitors, levothyroxine, the antihypertensives, and SSRIs like sertraline. These are the drugs UK patients are most likely to be taking when they ask about supplement interactions.

The remainder: high-stakes specialist additions. Some drugs with narrow therapeutic windows or critical interaction profiles fall outside the volume cut. Sirolimus and everolimus (mTOR inhibitors used post-transplant). The older MAOIs phenelzine and isocarboxazid. Rasagiline and selegiline (MAO-B inhibitors for Parkinson's). Tacrolimus and ciclosporin (calcineurin inhibitors, post-transplant). A real patient asking about a supplement interaction on any of these is in serious clinical territory, so we add them in despite the low volume rank. We have also extended the list over time to close class gaps a pharmacist would test first: the common antibiotics, the ADHD and mental-health drugs, and the GLP-1 agonists.

The substitution table below lists the most clinically important medications by category, including both top-volume drugs (warfarin, lithium, phenytoin, methotrexate, tamoxifen, the immunosuppressants) and the low-volume specialist additions. Every choice is published with its reasoning.

The complete medication list with each entry's volume rank or substitution reasoning is at data/medications.yml in our open repository. The list is regenerated quarterly when the next NHS BSA Prescription Cost Analysis release is published.

How we decide which interactions go in the database.

Most interaction checkers do not publish the rule they use for deciding whether a pair belongs. The British National Formulary lists about three thousand interactions but explains very little about why each one is there, a transparency gap an editorial in the British Journal of Clinical Pharmacology noted twenty years ago and which is largely unchanged today.

We publish our rule. It has two layers.

Layer 1. Does this pair belong in the database at all?

A supplement-medication pair belongs in the database if it meets at least one of five inclusion criteria:

1. Pharmacokinetic interaction with documented clinical impact. A published mechanism such as CYP induction or inhibition, transporter modulation, or protein-binding displacement, plus a measured clinical endpoint such as drug-level shift or AUC change of twenty per cent or more. Example: St John's Wort with warfarin via CYP3A4 and CYP2C9 induction.
2. Pharmacodynamic interaction with documented clinical impact. A published mechanism such as receptor co-agonism, additive physiology, or antagonism, plus a measured clinical endpoint such as serotonin syndrome event rate, INR shift, blood-pressure change, or hypoglycaemia event. Example: 5-HTP with SSRIs via additive serotonergic action.
3. Established absorption or chelation interaction. A physico-chemical mechanism such as cation-binding, complexation, or lipid-barrier effects, plus a clinical operational rule such as timing separation in hours or a formulation switch. Example: calcium with levothyroxine, four-hour separation per BNF.
4. BNF-listed pair. Any pair where the BNF Appendix 1 entry carries a "potentially hazardous" bullet is automatically included. BNF is the primary citation; we add our own mechanism and management note alongside.
5. Active-condition safety flag where mechanism is established and endpoint is serious. Examples include NMN with cancer history, and high-dose biotin with thyroid and troponin immunoassays. These flow through the optional context flag system, not the supplement-medication pair channel.

A pair is excluded if the only evidence is from cell culture, animal models, double-extrapolation chains across population and ingredient and mechanism and effect, or a single case report with no published mechanism to explain it.

Layer 2. What evidence quality is required for each call?

Every interaction-pair row in our database carries an evidence_grade field. The grade is set per pair using two published assessment tools:

• The Drug Interaction Probability Scale (Horn, Hansten, and Chan 2007), purpose-built for evaluating drug-interaction case reports. It adds DDI-specific questions that the original Naranjo nomogram (designed for single-drug adverse events) does not cover.
• The WHO-UMC causality scale for individual reports (certain, probable, possible, unlikely, conditional, unassessable).

The severity tier and the evidence grade are coupled. A Red severity tier always carries an evidence grade of established or well-documented. We do not put high-stakes calls on thin evidence.

When a pair sits at the edge of what evidence supports (a single strong case report, a plausible mechanism with sparse human data, a disagreement with a canonical source), we flag it for clinical reviewer escalation. Until a named clinical reviewer is appointed (an open work item against our first-revenue threshold), flagged entries are included with a visible "awaiting clinical reviewer sign-off" footer rather than silently demoted or omitted.

How a pair is checked before it goes live.

An inclusion rule and an evidence grade decide whether a pair belongs and how strong the call can be. Neither one checks that the citation behind it actually says what we claim. That is a separate job, and it is where most of the work is, so we made it a pipeline every pair passes before it is published.

One. The pair is researched and drafted against the primary literature, with the mechanism, the clinical endpoint, and the candidate citations attached.

Two. A second, independent pass re-verifies every claim the grading rests on against PubMed: the direction of the effect, the dose, the numbers, and each citation. It works with no sight of the first pass's reasoning, and its job is to disagree. It regularly catches a real paper cited for a claim it does not actually make, an effect stated in the wrong direction, or a number that has drifted. Those are corrected, or the pair is held, before anything ships.

Three. Every citation's identifier is machine-checked against the National Library of Medicine to confirm the paper exists and is the one named.

Four. A standing probe re-runs the textbook interactions a pharmacist would test first: the chelation pairs, the serotonergic stacks, the CYP3A4 cases, the bleeding-risk additions. It fails the build if any of them ever comes back as anything other than the documented severity. It currently passes with zero misses.

We do this because the dangerous failure for a tool like this is not a missing entry, since the user is told when something is missing. The dangerous failure is a confident wrong answer. The pipeline exists to catch our own mistakes before a reader sees one.

Where coverage stands now.

We have completed the assessment for every supplement in the checker against the full 296-medicine list. 924 of those combinations carry an individually researched, cited entry, every pair where the literature documents something, graded red, amber or green. For the remaining combinations, the supplement has been assessed against that drug and no documented interaction was found, so the tool returns green worded as "no documented interaction recorded". That is not the same as an explicitly studied "no interaction", and we word it differently, but it is emphatically not the silence other checkers return for things they have not looked at. Grey, "not in our database", is reserved for a supplement or drug we do not yet cover at all, and we always say so rather than defaulting to green.

How we grade interaction severity.

Interaction checkers tend to overuse severity warnings. When the same pair gets a Major rating in one database, Moderate in another, and is silently ignored by a third, the warning loses its meaning. One published comparison found that across the four most-used commercial interaction databases, only eleven per cent of flagged pairs were detected by all four. Another study, conducted in an intensive care unit, found that database severity ratings agreed with experienced clinician judgment on three drug pairs out of more than four hundred.

We treat that as the integrity bar to clear, not the standard to aspire to.

Distil uses a three-tier rubric calibrated against four published frameworks: the Hansten and Horn Operational Classification of Drug Interactions (ORCA), Stockley's two-axis action-and-severity model, the Agency for Healthcare Research and Quality 2016 high-priority criteria, and the Tilton 2013 consensus on clinical decision support tiering.

Red Severe

Clinically reportable harm has been documented, or a published mechanism predicts a serious endpoint at standard doses. Combining a Red pair without direct GP or specialist supervision is not advised.

Examples in our database include 5-HTP with SSRIs (serotonin syndrome), St John's Wort with warfarin or with transplant immunosuppressants (tacrolimus, ciclosporin) via CYP3A4 induction collapsing therapeutic levels, and iron supplementation without confirmed ferritin status (overload risk in unrecognised haemochromatosis).

Red maps to ORCA Class 1 (Contraindicated) and Class 2 (Provisionally Contraindicated), to Stockley's "avoid" action, to the AHRQ "high seriousness" tier, and to Lexicomp's X risk rating.

Amber Caution

A real clinical impact at typical doses, but operationally manageable. Most Amber pairs reduce to a specific action: separate the doses by a number of hours, hold within a stated dose ceiling, or monitor a specific lab value.

Calcium and levothyroxine separated by four hours is a typical Amber. Curcumin with warfarin, INR monitored, is another. Berberine with metformin, additive glucose-lowering monitored at initiation, is another.

Amber maps to ORCA Class 3 (Conditional), to Stockley's "monitor or adjust" action, to the AHRQ "medium seriousness" tier, and to Lexicomp's C (monitor) and D (consider modification) ratings.

Green No known interaction (explicitly assessed)

The pair has been assessed against our inclusion rubric and no clinically meaningful interaction has been identified at typical doses. Green covers two cases we word differently on the result page: a pair we have explicitly studied and found clear, and a pair where a fully-assessed supplement has no documented interaction with that drug. Both are an answer. Neither is silence.

Green is not the same as silence. Most interaction checkers default to silence for anything they have not checked, which is indistinguishable from "we looked and it's fine." We separate those two cases. The next section explains how that separation renders on the result page.

Green maps to ORCA Class 4 (Minimal Risk) and Class 5 (No Interaction), to Stockley's "no action needed" action, and to Lexicomp's A (no known) and B (no action) ratings.

Grey Not in database

If you enter a supplement or medication outside our coverage, we say so. We do not return Green by default. The result page shows a clear message above any per-pair findings, asking you to tell us, and the request enters a queue we review in our quarterly update.

Of the eight interaction checkers we audited (Drugs.com, WebMD, Examine, Memorial Sloan Kettering About Herbs, NHS BNF, Natural Medicines, OpenFDA, and DrugBank), none surface "assessed → no interaction" as a distinct user-facing state different from coverage silence on a public consumer UI. We do.

How we render results.

Interaction-checker user interfaces tend to fail in one of two directions. They either return a flat list of pair-by-pair findings, weighted equally regardless of stakes, or they hide a lot behind disclaimers and the user does the prioritisation work themselves. Both produce the same downstream outcome: a clinically significant finding gets missed because it was surrounded by visual noise.

The clinical literature on alert fatigue is consistent on this point. A 2006 systematic review found drug-safety alerts in hospital systems were overridden in 49 to 96 per cent of cases. A 2020 national evaluation of 1,599 hospitals found that some hospitals achieved higher safety scores by over-alerting, at the cost of clinician burnout and worse downstream outcomes. A 2005 study found that selective interruptive design, surfacing only the high-stakes alerts with prominent visual treatment, reduced override rates by a factor of nine.

We took the principle and translated it for a consumer surface.

The result page renders in three bands.

Header band. The scope statement at the top names how many pairs we checked and when the database was last updated.

High-stakes band. Red and Amber severity findings render in full above the fold, in tier order, each with the specific action it implies. Action language follows the NHS BNF convention ("avoid", "monitor or adjust", "inform your GP"), not informal severity adjectives.

Assessed-no-interaction band. Green findings render below, collapsed behind an explicit counter you can expand. The counter is named so you can verify that nothing was silently dropped, for example "Show all no-interaction pairs".

Per-pair detail is on demand. The default render shows the severity tier, the paired names, and a plain-English one-liner. Mechanism, citations, and dose-sensitivity notes sit behind a "Read more" expand on each finding. The principle is to keep the baseline density low enough that the high-stakes findings are visible without scrolling, and to put the depth where users go when they actually want it.

If you entered something we do not have data on, the "Not in database" message renders at the top of results, visually distinct from any severity finding, with a feedback button that adds the gap to our queue for the next quarterly update.

How we compare to other interaction checkers.

Before we built this tool, we audited eight existing reference resources. Each does something well. Each has gaps. We made our design decisions against that landscape.

Drugs.com and WebMD cover the largest stack of medications and supplements, are free, and use a familiar three-tier severity scheme. They are advertising-supported on the results page, do not publish their inclusion methodology, and do not show per-pair citations. We do not place programmatic advertising on a health-decision surface.

Examine.com has the closest voice to ours and the strongest independence model (no advertising, no industry sponsorship). Their Safety Checker for drug-supplement interactions sits behind the Examine+ subscription paywall. We chose to make our equivalent free.

Memorial Sloan Kettering About Herbs is the best free evidence-led supplement reference and uses inline PMID citations on every monograph. It is monograph-format, not a pair-checker, and covers 290 herbs rather than drug-anchored stacks. We adopted the per-pair PMID anchoring; we built the pair-checker UX they do not have.

NHS BNF is the UK clinical gold standard, free, and uses a five-tier action scheme that is more granular than ours. Its scope is explicitly restricted to conventional medicines; supplements are mentioned only where they affect conventional therapy. We fill the gap on supplement-medication coverage and we cite the BNF as a primary anchor whenever a pair is BNF-listed.

Natural Medicines (formerly Natural Standard) is fully paywalled. We acknowledge it as a clinical reference; we publish our methodology where they require subscription.

OpenFDA and DrugBank are developer-facing APIs. DrugBank's clinical API covers 1.3 million drug-drug interactions with a severity-and-evidence-separation architecture that informed our own. Both are drug-anchored; supplement coverage is secondary.

Compliance and regulatory posture.

Medical device classification.

The Distil interactions checker is a rule-based reference tool. It looks up pre-curated pair data in a static database and returns published clinical findings with citations. There is no novel calculation, no machine learning, no personalised risk scoring, and no device-specific calibration. Every call we make is independently verifiable against the cited primary literature on this page.

Under the Medicines and Healthcare products Regulatory Agency's Software as a Medical Device guidance, low-functionality reference tools (the cited example is BMJ Best Practice) are not regulated as medical devices in their current form. The Distil interactions checker sits in that category.

Advertising standards.

The tool is an information surface, not product marketing. It reports documented interactions between products users already have or are considering. It does not promote a specific product, claim to treat or cure any condition, or recommend a supplement as a treatment. The brand-voice rules we apply to all customer-facing copy (probabilistic language, no medicinal claims) extend to the tool.

Data protection.

A pair-check query is anonymous. You type in supplement names, medication names, and optional context flags. We do not capture an identifier, build a profile, or create a persistent record of you as a user. No personal data is processed through the pair-check path.

The missing-item feedback form is the only surface where a small amount of personal data is captured (an email address and a free-text item description). It is opt-in, used only to acknowledge your submission and to notify you when the missing item is added in the next quarterly update, and you can opt out at any time. The lawful basis is consent under UK GDPR Article 6(1)(a).

UK GDPR Article 22 gives you the right not to be subject to decisions based solely on automated processing where those decisions have legal or similarly significant effects. The interactions checker does not engage Article 22, for three independent reasons. The pair-query flow captures no personal data for decision-making purposes. The tool returns information about documented interactions, not a decision about you. And where you act on the output, your GP or pharmacist is the gating reviewer for any clinical change.

NHS Digital Technology Assessment Criteria.

The DTAC framework applies when a digital health technology is procured or used by the NHS. The interactions checker is a consumer-facing free tool, not procured by the NHS. DTAC is not applicable at our current scale. If we ever pursue NHS partnership or NHS app library inclusion, DTAC assessment becomes required and we will publish that submission.

How we update the database.

The database is regenerated quarterly. Every quarter we publish:

• An updated data/interactions.yml reflecting any new evidence, withdrawn calls, or escalated reviews.
• An updated data/medications.yml against the most-recent NHS BSA Prescription Cost Analysis release.
• A calibration report listing every change since the prior quarter, with reasoning per change.

The calibration log lives in our open repository alongside the data files. Every change since the database first shipped is traceable.

If you cannot find a supplement or medication in our coverage, tell us. The feedback button on any result page adds the missing item to a queue we review at the next quarterly update.

A note on what this section does not cover.

This part of the methodology page describes the interactions checker as it ships at launch. The compounds database that powers our supplement reports (the product described in part one above) has its own methodology, calibration log, and evidence-grading framework. The two methodologies share principles (probabilistic language, transparent inclusion rubric, quarterly cadence, published calibration reports) but the evidence base differs because supplement clinical literature is structurally thinner than drug-drug literature, a point Williamson's review of Stockley's Herbal Medicines makes explicitly.

When you click into a Distil supplement compound from a pair finding, you arrive at the relevant entry in our compounds database. The same evidence-grading rules apply on both surfaces.