no login, no key, every number sourced

Evidence

Bottles: 687. Gates on record: 5. Release 758634b.

Not predicted: tinted (626), powders (1), lip products (45), zinc too low to predict a cast (2), and any render more than 1 L* off its prediction (0).

0.982

blend model fit (R squared, 16 renders)

1.17

dE00 to target on a second selfie (bar 2.0)

0.355

blend strength at colorIntensity 100

API units spent on K1b

K1b blend strength by colorIntensity

blend strength

colorIntensity 0100

Gate 1

Does a film model measured over black and white backgrounds predict on-skin whitening better than the label's zinc percentage? (PLOS ONE 2025 raw data, 13 volunteers, held out by person and by formulation)

FAIL film model 4.79 L* mean error vs label model 1.61. Exploratory only: on the one differently formulated product, a one-parameter film model missed by 0.69 L* vs the label's 2.35. gate1-results.md

Is YouCam's skin tone analysis a usable measurement of the colour in a photo?

FAIL 2 of 3: identical result on a fresh upload (pass), but it moved only 0.15x as much as the measured cheek when the photo was made 10 percent brighter, and it sits 15.1 dE00 from the measured cheek. Reading, not pre-registered: it reports a lighting-normalized tone. Post-hoc daylight check: FAIL, the colour moved 5.81 dE00 against a 1.0 bar. k2-results.md

Can YouCam Makeup VTO render a sunscreen-like film, with coverage as the dose?

FAIL on 4 of 5: L* reversals at coverage 60 in 3 of 6 series, max dL* 12.7 < 15, inverse dE00 2.03 > 2.0, texture 0.556 < 0.80; R^2 0.954 passed. Post hoc reading (not pre-registered): colorIntensity, not coverage, acted as the dose. k1-results.md

K1b

Same question with colorIntensity as the dose and coverage as texture

PASS 5 of 5. k1b-results.md

Gate 2

Does a retail bottle's film, measured on a black and a white card with a phone, predict its whitening on skin better than the label's mineral percentage? (one consenting subject, forearms)

NOTE Not run. On 2026-10-28 the builder decided not to run the in-person session (cards and forearm), so there is no Gate 2 data and nothing on the site rests on it. Bottles are instead predicted from their FDA labels with the label model Gate 1 tested.

1
Read the gates
Each gate has a written pass rule, and every result stays, pass or fail. Gate 1's page notes what was looked at before its rule was written.
docs/gates on GitHub (opens in a new tab)
2
Check the prediction
Refits the label model from the study table and recomputes Gate 1's held-out error.
curl -s https://cast-forecast.vercel.app/api/label-model
3
Ask it as an agent
The same model, as MCP tools an agent can call.
curl -s https://cast-forecast.vercel.app/api/mcp -H 'Content-Type: application/json' -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"predict_cast","arguments":{"bottle_id":"cetaphil-sheer-mineral-face","skin_lightness":55}}}'
llms.txt (opens in a new tab)
4
Check the render mapping
Recomputes the K1b inverse from the published curve.
curl -s https://cast-forecast.vercel.app/api/gates
5
The data
The 13-volunteer study table every prediction is fitted on (CC BY 4.0), and the FDA label behind each of the 687 bottles.
data/ on GitHub (opens in a new tab)
6
Read the code
Lint, types, tests, build and a bundle scan run on every push.
StephenSook/cast-forecast (opens in a new tab)

Read from docs/gates in the repo at release 758634b. K1b used one person and one lighting condition. It tests YouCam's renderer, not any sunscreen.

Evidence

Read the gates

Check the prediction

Ask it as an agent

Check the render mapping

The data

Read the code