Methodology

How We Score Ingredient Safety

theformulator.ai scores cosmetic ingredients on six independent hazard axes — carcinogenicity, developmental and reproductive toxicity, sensitisation, systemic toxicity, irritation and endocrine disruption — from regulatory primary data: ECHA CLP classifications, REACH SVHC listings, IARC and NTP monographs, and CIR final safety assessments. Scores exist for 11,000+ ingredients, each axis stored with the source behind it.

The methodology behind the scored ingredient safety profiles powering theformulator.ai

Every formulation generated on theformulator.ai includes per-ingredient safety scores across six hazard axes. This page explains how those scores are calculated, what data sources feed them, and why we built a proprietary system instead of licensing an existing one.

These safety scores power the formulation generator, the formulation audit, the deformulation tool, and the Formulation Partner AI co-pilot.

Why existing safety databases fall short for formulators

The two most widely referenced ingredient safety databases, EWG Skin Deep and SkinSafe, were designed for consumers, not formulators. This creates fundamental problems when R&D teams try to use them as data inputs for formulation decisions.

EWG Skin Deep scores approximately 11,500 ingredients. Its methodology applies a data gap penalty: ingredients with limited safety studies receive elevated hazard scores. For a formulator working with novel or niche raw materials, this means an absence of negative evidence gets treated as evidence of harm. The scoring algorithm is opaque: weights and methodology are not published. Exposure context (whether an ingredient appears in a leave-on serum at 2% or a rinse-off shampoo at 0.5%) is not factored into the score. And the database is consumer-facing, with framing that tends toward alarmism rather than risk-based assessment.

SkinSafe takes a different approach, focusing on allergen and irritant avoidance based on Mayo Clinic data. While clinically grounded for contact dermatitis, it does not cover the full hazard spectrum that regulatory affairs teams need — carcinogenicity, reproductive toxicity, and endocrine disruption are outside its scope.

Neither system provides per-market regulatory status. A formulator developing for both EU and China needs to know that an ingredient permitted in Europe may be prohibited or restricted under NMPA. This cross-market view does not exist in consumer-facing databases. Neither system updates automatically from primary regulatory sources. Neither system adjusts scores based on product type or exposure duration.

We needed a scoring system that:

• Draws from the same primary sources regulatory authorities use
• Treats absence of data honestly — NULL on absence, composite over available axes, minimum 3 of 6 required, “Insufficient data” whenever any axis is missing
• Leaves the intrinsic hazard score unchanged and computes an exposure-adjusted indicator per ingredient at formulation time
• Covers the full hazard spectrum across six independent axes
• Provides per-market regulatory intelligence alongside safety scores
• Updates from primary sources automatically, not via periodic manual review

Six axes of hazard assessment

Each ingredient is scored 0–10 on six independent hazard axes. A composite score is calculated, but all six axis scores are always visible to the formulator, because a single number hides critical information.

Composite scores are banded as Low, Moderate and High. An ingredient with no assessment in our hazard sources is reported as "insufficient data" rather than given a band, and its composite is not shown — absence of evidence is not a low score.

The composite is a weighted sum over the axes that carry data, renormalised so weights sum to 1.0 over those axes. A minimum coverage of 3 out of 6 axes is required before a composite is produced; below that, only a floor override yields a score. Composite floors prevent clean axes from masking a severe classification on carcinogenicity, sensitisation, or irritation. Confidence is derived from coverage: an ingredient assessed on all six axes carries full confidence; any missing axis is labelled “Insufficient data”.

Five principles that guide our scoring

1. 1

   Honest data gaps

   If an ingredient has no studies addressing a particular hazard axis, that axis is NULL — absent, not zero. The composite is computed over the axes that ARE present, with weights renormalised to sum to 1.0 over those axes. A minimum of 3 assessed axes is required before a composite is produced at all; below that, only a floor override (e.g. a confirmed carcinogen) yields a score. Coverage is reported as n/6 alongside every score. Any ingredient with fewer than 6 assessed axes is labelled "Insufficient data" — the score is shown, but the formulator knows it rests on incomplete evidence. This is the single largest methodological difference between our system and EWG Skin Deep, which inflates scores for under-studied ingredients.

2. 2

   Exposure context matters

   The ingredient's intrinsic hazard score does not change with the product: carcinogenicity, sensitisation and the other axes are properties of the substance, not of a formulation. What changes is exposure. At formulation time, for each (ingredient, product type, concentration) the platform computes an exposure-adjusted indicator from the SCCS Notes of Guidance default daily amount and retention factor for that product type, and shows it beside the unchanged intrinsic score with its table reference. This is standard practice in regulatory toxicology, where SCCS and CIR both evaluate safety in the context of exposure, but is absent from consumer-facing databases.

3. 3

   Source hierarchy

   Not all safety data is equal. Peer-reviewed regulatory assessments (ECHA CLP, SCCS opinions, CIR final reports) carry more weight than preliminary findings or supplier-provided data. Our system enforces a clear hierarchy: regulatory authority classifications first, then peer-reviewed clinical data, then curated literature. Scores from higher-tier sources are never downgraded by lower-tier data.

4. 4

   Confidence tiering

   Every safety score carries a confidence level: high (direct ECHA, CIR, or IARC classification), medium (extracted from peer-reviewed literature or CIR group assessments), or low (baseline from COSING registration with no hazard evidence found). Only medium and high confidence scores are displayed to users. Low confidence ingredients show "No hazard data available in our sources" rather than a green score that implies safety has been confirmed.

5. 5

   Conservative on carcinogenicity

   Carcinogenicity is treated asymmetrically. A known carcinogen — classified on the carcinogenicity axis, whatever the source that classification came from — cannot have its composite score pulled below a floor by clean scores on other axes. When multiple sources provide different carcinogenicity classifications, the highest (most conservative) score is retained. This reflects the irreversible nature of carcinogenic harm compared to reversible irritation.

6. 6

   Corpus grounding

   Every formulation is grounded in a scan of the platform's marketed-product corpus. The generation pipeline embeds the brief and retrieves the top K products by cosine similarity, then selects ingredients informed by what the market already uses. The report names 3–5 benchmark products with three scores. Relevance: cosine similarity of the brief embedding to the product embedding, normalised 0–100; formula (1 − d/2) × 100, where d is the Weaviate cosine distance (range 0–2). Coverage: the share of the generated formula's INCI list present in this benchmark product's ingredient list, expressed as a percentage. Novelty: the share of the generated formula's ingredients not found in any of the named benchmark products, expressed as a percentage (shown once per formula). N (the count of products used) and K (the retrieval limit) are live values from the generation that produced the report, not literals.

Where our data comes from

Two of the hazard sources above — IARC's list of classifications and the NTP Report on Carcinogens — are fetched automatically on a daily schedule; each run is recorded with the fetched file's hash and timestamp, so a change at the source is detected the day it lands. ECHA's CLP Annex VI and its SVHC candidate list have no machine-fetchable endpoint anywhere reachable, so they are captured manually instead: the table below shows each source's last capture date and its age, and a capture that has gone stale raises an alert rather than passing silently. CIR assessments are re-run against the platform's own CIR collection rather than fetched from an external site.

Safety scores are only half the picture

An ingredient can score well on all six hazard axes and still be prohibited in your target market. Formaldehyde releasers are a clear example: permitted with concentration limits in some markets, banned outright in others. Safety scoring without market-specific regulatory status is incomplete.

Every ingredient in our system carries per-market regulatory status across 21 markets: EU, Switzerland, UK, US (FDA), China (NMPA IECIC), Japan (MHLW), South Korea (MFDS), India (BIS IS 4707), Canada, Australia (TGA), Brazil (ANVISA), Turkey (TİTCK), Taiwan (TFDA), Thailand, Malaysia (NPRA), Singapore (HSA), Indonesia, Vietnam, and Philippines, plus ASEAN harmonised standards. Status categories include: permitted, restricted (with maximum concentrations, required labelling, and product-type limitations), prohibited, and not listed.

IFRA: the voluntary fragrance-safety plane

A separate check runs alongside the per-market legal status above: whether a declared material, or a Natural Complex Substance carrying it, is restricted or prohibited under IFRA Standards. IFRA — the International Fragrance Association — is an industry body, not a regulator, and an IFRA Standard is not law in any market. We keep that distinction absolute in the platform, not just in this paragraph: every IFRA finding is tagged plane: 'voluntary'and carries its own authority string (“IFRA Standard {name}, 51st Amendment”). It is never merged into a formulation's legal verdict, which is drawn exclusively from the market-law rows described above.

A CAS-match advisory answers one question: does the declared material itself appear on an IFRA Standard. That misses most of what IFRA actually restricts. IFRA Standards reach a Natural Complex Substance — an essential oil, absolute, or extract — through its constituents, and reach certain Schiff bases through the aldehyde they release on hydrolysis, so a restricted material can enter a formula without ever being declared by name. Declaring Balm oil, for instance, carries citral into the mixture invisibly to a CAS-only check. Our screening resolves the declared ingredient to its Natural Complex Substance entry and its constituents, so that citral contribution is counted rather than missed.

IFRA Guidance §7.12 describes a further case: a Natural Complex Substance that is itself the subject of its own IFRA Standard — Ylang ylang is IFRA's own worked example. When that happens, both the constituent-derived limit and the NCS's own limit apply, and the lower of the two binds; scoring only the constituent path under-restricts the ingredient. How much of the library this touches was measured twice. A first pass, linking an NCS to its own Standard only through the Standard's recorded synonyms, found 10 such NCS entities. Re-measuring found a second, wider linkage — an NCS whose own CAS number is itself registered against a Standard — that the synonym-only pass never saw. The union of both linkages is 39 distinct NCS namescarrying a dual path, and the two linkages never double-count: the calculator takes the higher of the summed-constituent path and the NCS's own-Standard path, never their sum.

IFRA's own published materials do not always agree with each other, and where they conflict we say so rather than silently picking one. IFRA's Guidance worked example for Coumarin in Category 4 prints 1.3%, but IFRA's Standards library and IFRA's own overview spreadsheet both independently give 1.5% — two IFRA sources agreeing against IFRA's own prose. We took the library figure and recorded the reconciliation in the standard's own source notes; our worked example for that case computes 7.5% Coumarin in the mixture, 93.75% Hay absolute of the mixture, and 18.75% of the finished product.

When you generate a Certificate of Conformity, the document rendered is IFRA's own template — the December 2019 Certificate of Conformity of Fragrance Mixtures with IFRA Standards — reproduced clause for clause, including IFRA's own wording where it is imperfect. Exactly two lines on the page are ours: a footnote stating that the optional table's concentrations are given in the fragrance mixture, and a closing line naming who prepared the document. Everything else — the attestation, the category table, the disclaimer — is IFRA's language, not ours. The certifying party you name issues the certificate; theformulator.ai computes the figures and renders the template, but does not certify anything itself and does not issue Certificates of Conformity. Standards referenced on the certificate are the 51st Amendment, published July 2023.

The library behind this screening is IFRA's own Standards, category limits, CAS registrations, Natural Complex Substance constituent tables, and Schiff base contributions. Most constituent percentages are read straight from IFRA's published tables; a minority are curated where IFRA's own materials disagree with each other or a raw figure needed interpretation, and every curated value carries a value_curated flag so it is never silently indistinguishable from a value read straight off the page.

What we deliberately exclude

This methodology is not static. As regulatory authorities update classifications, as new clinical data enters the peer-reviewed literature, and as our automated pipelines expand coverage, scores are refined. We publish this methodology because transparency is a prerequisite for trust, and trust is what separates a tool formulators rely on from one they dismiss.

Last updated: September 2026..