OSHUN·
Ingredient science built for textured hair

Measure the fiber.
Predict the product.

Portrait of a Black woman with her textured hair pinned up, seated in a warm red and gold room.
Every ingredient in her routine was modelled on somebody else's hair.
The substrate no model has been fitted on

The industry's standard model cannot answer half of a curl cream — not "answers it badly," cannot answer it at all. The fix is not new physics. It is conditioning the model on the one thing nobody has ever measured: the geometry of textured hair itself. Every piece already exists. Only the data is missing — and it is collectable.

in plain terms

You already know the shade-range problem — foundation lines built by measuring one kind of face, and everyone else gets whatever is left over. This is that exact problem, but for chemistry. The equations that predict how an ingredient behaves were fitted on one kind of skin, and nobody ever asked what the hair was like. OSHUN asks.

534 measured permeabilities huskinDB · CC-BY leave-one-out every loss published
↓ Read the full studyPDF · 30 pages · methods, tables, and the results that failed
The solution, before the problem

Three moves. All three already exist.

There is no research risk in this. Every component below is off-the-shelf and in use somewhere today — the fiber measurements, the geometry, the database. What is missing is that nobody has put them in this order and pointed them at textured hair.

in plain terms

Think of the three steps you would take to fit a garment properly. First you take real measurements off a real body, instead of writing down "size M." Then you keep those measurements attached to the person instead of averaging everybody into one mannequin. Then you only make the promises the measurements support — and you say "I'd be guessing" when they don't. That is the entire product.

Measure the fiber

Ellipticity and twist density off a real head, on an instrument that already exists. Not a self-reported curl type, and never a demographic box — an actual number, per person.

Optical fiber analysis · commodity since the 2000s

Keep the context attached

Store each measurement with the substrate it was measured on, so the model can condition on it instead of averaging it away. This is what a fiber bundle is for, and it is the result the staircase below demonstrates on public data.

GIGI · shipped, running this page

Refuse when you're outside

Every answer carries the boundary it was computed inside. Off the edge of the measured data, the honest output is a refusal with a reason — not a confident number nobody can stand behind.

Applicability domain · regulatory standard practice

Each move is boring on its own. The composition is the product — and the third move is the one that turns a model into something a formulator can actually put weight on.

What this costs you today

Every dead formula is a year and six figures.

Every ingredient-performance model in this industry was fitted on skin and hair that isn't ours — and none of them has a term for texture.

in plain terms

Right now the only way to find out whether a formula works on textured hair is to make it and try it. That is a year of bench work and a substantiation study before anybody learns the answer. It is the equivalent of sewing the whole dress before discovering it doesn't fit — and the fabric is not cheap.

Per SKU · concept to launch
$23.5–98K

12–18 months of formulation, stability and safety work.

Per clinical claim
$25–40K

12–16 weeks each. A full CRO study runs past $500K.

Spend vs. ownership
$2.6B

Black consumers: 11.1% of US beauty spend. Black-owned brands: ~2.5% of revenue.

OSHUN's job is to kill a bad formulation before the $25–40K substantiation study, and before the 12–18 month clock starts. Time is the scarcer resource.

Live · runs in your browser on real data

Check a formula. Watch it refuse.

Build a deck or load the preset curl cream. Predictions come from the published Potts–Guy model and a geometric neighbour model fitted on 534 measured human-skin permeabilities. Nothing is looked up — it computes as you click.

in plain terms

A tape measure that stops at 36 inches doesn't become wrong at 50 — it simply cannot answer. You can hold it up against something longer and read off a number, but you're making that number up.

Most of the ingredients that actually make a curl cream work sit past the end of the tape. Watch how many come back refused. Click any row and it will show you its own arithmetic — where the number came from, or exactly why there isn't one.

Your formula
The headline result

The error is in the context, not the molecule.

Same two descriptors, same model, every step. The only thing that changes is how much context the model may condition on. This is the whole argument for measuring hair type — the one context nobody has ever conditioned on.

in plain terms

Every baker knows this one: the same recipe behaves differently in different kitchens. Altitude, oven, humidity. If you ignore the kitchen and average every bake together, you end up with a recipe that is wrong for everybody and right for nobody.

That is what the industry model does — it averages across every skin type, every condition, every lab. Each step below hands the model a little more information about the kitchen, and nothing else changes. Watch the error fall anyway. Nobody has ever handed it the hair.

What we tested and could not prove

A negative result, published.

We hypothesised that a compound's coherence — how much its behaviour curves as the substrate changes — would predict how reliably we can model it. It is the method our drug-delivery system uses. On this data it does not hold, and here is the test that killed it.

in plain terms

We wanted to know whether an ingredient behaves consistently — whether it's dependable. But the only repeat readings available came from different laboratories using different equipment. That is like trying to judge whether someone's weight fluctuates using scales borrowed from twelve different bathrooms. Most of what you'd see is the scales, not the person.

So we could not answer the question. We are telling you that instead of pretending we did.

The dose-response test failed

If the signal were real it should strengthen as the curvature is measured from more repeat measurements. It weakens, then reverses.

  • Why, and why it still matters: these repeats come from different laboratories, and the founding dataset of this field has documented 1–2 log-unit disagreements between labs on the same compound. So "variation across context" here is measuring lab disagreement, not substrate response. The hypothesis is not refuted — it is untestable on pooled public data.
  • Testing it needs one lab, one protocol, and the substrate varied on purpose. That is a controlled panel, and it does not exist for textured hair.

Everything else on this page survives that: the refusals are hard applicability-domain boundaries, and the staircase is a straightforward conditioning result. We publish the loss because a model that hides its failures is worth nothing to a formulator — and because a competitor's chemist would find this in ten minutes.

Why this is defensible

The data doesn't exist. That's the opportunity.

in plain terms

The recipe is published. The oven is for sale. What nobody has is a kitchen full of the right people — real heads, measured properly, linked to products they actually use.

That is the one part money alone doesn't buy, and it's the part you already have. It is also the part that only gets built once: whoever measures that panel first owns the reference data this whole category gets compared against.

The instrument
GIGI

All 534 measurements live in a GIGI fiber bundle — keys on the base, descriptors and measured permeability on the fiber. Curvature and capacity are computed at insert time and travel with every answer.

The asset
Your panel

Measured ellipticity and twist density — not self-reported type. Type VI fiber runs ellipticity 1.7–1.9 and 12.3 twists per 5 cm, and fractures concentrate at exactly those curvature points. Nobody has conditioned a performance model on that.

Condition on measured geometry, never on demographic category — self-identified race predicts curl class poorly and the groups overlap heavily. Correct science and correct ethics land on the same architecture.

The whole thing, on paper

Read the study. Including the part that failed.

Everything on this page written up properly: the data and its provenance, the baseline model and a unit error we found in its redistributed form, the full conditioning experiment with per-stratum tables, the applicability-domain analysis, the hypothesis that did not survive its own dose-response test, and the panel design that would settle it.

in plain terms

If you want to hand this to a chemist, hand them this. It is written the way a journal would want it — methods you could repeat, tables you could check, and an honest account of the one idea we tried that didn't work.

That last part is the reason to trust the rest of it.

↓ OSHUN — conditioning ingredient-permeability prediction on measured substrate context PDF · Davis Geometric · every figure computed from CC-BY public data