Overview

The proof of concept in one view

A virtual Seint artist. It starts from the palette Seint's own Find My Shade quiz would give, observes the face, applies Seint Artist Theory to adjust products and placements, renders the look on photographs with our own compositor and live with Banuba, and revises the look with the user.

The work is seven modules. Each module can be built and tested on its own. The proof of concept is those modules running as one session, in AR and on photographs.

What "finished" means

Two conditions, both required.

1. Every module has passed its required minimum checks in its own test setup.
2. A Seint artist accepts the full look on at least 95 percent of the frozen photo set. Accept means the artist would give that look to the user as the starting point. A small opacity change is allowed. A change of product, placement, or decision is not an acceptance. Other bars are on the acceptance page, in steps of 5 percent.

A walkthrough with no score report is not acceptance. The measure is defined on the acceptance standard page.

System map

Every piece of work belongs to one activity. Configure happens before a session. Experience is the session. Evaluate measures both. Select a box to see what it is for and which records it shares.

The seven modules

Two modes, one decision

AR and photo use the same analysis, the same theory, the same Look Composer agent, the same Look Editor agent, and the same look JSON. The mode changes only how the face is captured and how the look is shown.

• AR captures four poses: straight, left, right, and chin up, so the neck is in frame.
• Photo requires a straight photograph. Left and right are optional. A neck photograph is optional. The main highlight is matched to the face. When the neck is visible, the report also says whether the neck is darker than the face and how the jaw contour will bridge it.

Overview

Platforms

The user apps capture the face and paint the look. The backend owns the data, the APIs, and both agents. It is the core of the system.

How the pieces connect

iOS

The user app for preferences and the full experience, in AR and photo. Native iOS. The interface toolkit is not chosen yet.

The same app includes test modules for stencils, product rendering, and related checks, so a build can be tried on a phone before the scores are recorded.

Android

The same user app as iOS: preferences, AR, photo, and the same test modules. Native Android. The interface toolkit is not chosen yet.

Web

The same user experience as iOS and Android, without the test modules. Next.js.

Admin

For the Seint team: configure, test, evaluate, and change the platform. Next.js. It is a CMS plus the test platform: catalogue, stencils, theory, retrieval checks, and scores.

Backend

FastAPI. It owns the CMS data, the APIs, and both agents. Capture and live AR rendering (Banuba) stay on the user apps. Photo rendering moves to our own compositor on the API (planned; today the web app's photo test still uses Banuba). Face analysis, composition, retrieval, and edits that need a model are requested through the API.

Data

Postgres. Product use-case search uses pgvector in that same database. Inventory does not add a separate vector database.

Overview

Scope

What the proof of concept includes, and what it leaves out. This is the working boundary until the Seint team changes it.

Included

Face analysis

• AR and photo.
• AR poses: straight, left, right, and up, for the neck.
• Photo: up to three positions. Straight is required. Left and right are optional. A neck photograph is requested and optional.
• Face shape as probabilities over round, heart, oval, oblong, rectangle, square, pear, and diamond. Shape changes contour only when the top shape's confidence is at least 0.55.
• Skin tone by region, including the neck when it is visible.
• Skin texture, and a separate under-eye texture reading, used to choose between an under-eye crescent and corner accents. Lip colour. Lip texture. Eye, hair, and brow colour.
• Issues with a location: redness (mild or vivid), brown pigmentation or melasma, raised blemishes and acne, dark circles, fine lines and texture.
• Dark areas classified as a colour cast (blue, purple, green, muddy) or a shadow from structure, with the darkest point.
• A fixed set of notes the composer can use. The list is on the face analysis page.
• Age. The profile is the source when the user has provided it. Gender is not collected.

Profile

• Age, mood, occasion, season.
• Find My Shade answers: level, face A–D, skin type, coverage, powder finish.
• Concerns the user names: redness, dark circles, pigmentation, texture.
• Coverage per zone, optional. For example, more on a melasma patch, less under the eyes.
• Application order preset: highlight first, contour first, main first, Cheetah dots, or one brush. It changes the guide only.
• Free-text preference, shown back to the user as structured points before a look is built.
• Known skin conditions, allergies, and ingredient limits.
• Seint palette: the products the user already owns.
• A saved base profile that a session can override without deleting.

Inventory

• A digital record of current Highlight (24), Contour (14), Lip + Cheek (40), Illuminator (7, all cream), Balm, Bronzer (5), and Cream Corrector (12, the Demi line) shades.
• For the features step, the products seintofficial.com sold on 29 September 2026: Eyeshadow (56), Eyeliner (one shade, black), Pencil (13), Brow Wax, Archangel Brow Gel, Lipstick (3), Lip Tint (5), Lip Gelato (9), Lip Oil (11), Lip Whip, Lip Balm, Blackberry Lip Flush, and Pressed Pigment (20). Seint sells no concealer and no mascara. Highlight is Seint's concealing shade, and Cream Correctors treat colour casts.
• Level-set membership: which No. 4, No. 8, and No. 12 collections and which of the 48 quiz faces include each shade.
• Each product categorised by the element roles it can fill.
• A written use case on every product.
• A vector index of those use cases, stored in Postgres with pgvector.
• Situation retrieval that can be tested against artist-labelled answers. Retrieval uses filters and shade scoring first, then the pgvector index.

Elements, theory, composer, renderer

• A fixed list of elements. Each element has a short theory paragraph passed into the composer, and a set of stencils. Each stencil has a one- or two-sentence usage line, a blend style (gradient, feather, overflow, directional drag, or stipple and buff), and a blend recipe that moves the placed product the way the artist's brush would.
• Stencils carry depth ceilings per zone, for example under-eye brightening at most 0.60, and a safe direction when landmarks are uncertain.
• A management screen for theory text, stencil usage, and stencil weights. The screen cannot add an element.
• Artist Theory v2 as the starting document, editable in the application as markdown, without a code deploy.
• The Look Composer agent. It starts from the brand base: the Find My Shade palette and the No. 4, No. 8, and No. 12 sets for the user's level. An LLM agent then fine-tunes that base in four steps, in dependency order: base complexion (highlights and contour), correction, optional colour and glow, and features. Code does the maths and the rule checks. jev (TypeSafe) answers the atomic checks at each gate. A vision model checks the render. The user always confirms the face report (G1) and the final look (G4). The complexion (G2) and the optional colour (G3) are shown as editable summaries, and ask the user only when a check is not confident. The flow is on the composer page.
• The Look Editor agent, used in both modes. It either changes one setting, redoes one element, or rebuilds the plan. On-screen toggles and sliders are not this agent.
• The look is a JSON object. Our own compositor renders it on photographs; Banuba applies it in live AR only.

Not included, unless the Seint team adds it

• Eye, brow, and lip products in live AR when the face reading for that area is not confident enough (below 0.70, a starting value). They are then drawn on the photo only, with a short note on the final-look screen. The features step always decides them. Tools and setting products appear in the guide as text. The guide names the powder and spray the quiz would give: matte to Vanilla Dust, radiant to Feather Blotting Dust; oily or combination to Fixate, dry or balanced to Plush Prep Mist. Plush Prep Mist was sold out on 29 September 2026, so availability is checked when the guide is written.
• Drawing Cream Correctors in AR, if the first build cannot draw them honestly. Cream Correctors (the Demi line) are in the catalogue. In photo mode they render as a subtle colour shift, because they are drawn at 0.10 opacity (Seint artists describe them as 10 percent opaque). In AR they may be named in the guide without being drawn. A deeper highlight is not a substitute for a corrector: it treats a darker spot, not a colour cast.
• A purchase checkout. The result lists products to keep and products to buy.
• Accounts beyond a saved profile on the device or session.
• Continuous analysis on every camera frame. The session captures, then analyses.

Build order

Photographs come before AR. A still image can be scored again. A live scan cannot.

1. Agree the records modules pass to each other, and freeze the face set.
2. Pass the required minimum checks for each module on its own.
3. Run the photo session from start to finish with no hand-edited JSON. Record the score, even if it is under the photo bar.
4. Raise the required scores to their bars. Optional rows are recorded and do not block this step.
5. Add AR capture, live render, and the live-scan bar.
6. Pass the edit checks, and confirm the guide and the product list match the look the user saw.

Overview

Find My Shade prior

Seint's Find My Shade quiz is Seint's own codified artist. The composer starts from the palette the quiz gives, and leaves it only when the face or the brief gives a reason, which the plan states.

What the quiz decides

Six answers, 6,912 paths. The level and face answer choose the palette. The other answers choose tools and finishing products.

How the composer uses it: the brand base

The quiz palette and the level-based No. 4, No. 8, and No. 12 palettes are the base logic of every look. They keep the result consistent with what Seint itself recommends. Face analysis only fine-tunes the choice, and every change carries a reason.

• Step S3 of the composer builds the brand base with no model: the quiz palette for the confirmed level and face, plus the No. 4, No. 8, and No. 12 sets for that level. On 29 September 2026 every level's No. 4 set sat inside its No. 8 set, and No. 8 inside No. 12. The sources only add products; they never disagree.
• The agent picks from the brand base first. A product outside it must carry a deviation: what it replaces, why, the rule it follows, and the finding it answers. A plan with an unexplained product fails the schema.
• The main highlight is chosen from the quiz highlights first. The search widens to all 24 only when none is close to the face, and the wider pick is recorded as a deviation.
• If the user has not taken the quiz, level and face are estimated from the photograph and confirmed by the user at the first gate. Skin type, coverage, finish, and age are asked, in the quiz's wording, because a photograph cannot tell them reliably.
• If the capture cannot be trusted for colour, the quiz palette is the look. This reproduces Seint's own recommendation instead of guessing.

Acceptance

Overview

Acceptance standard

Bars follow how much a miss would hurt, and how hard the check is. A required row must pass for the proof of concept. An optional row is scored and shown, and a miss does not fail it. The set is frozen by the Seint team and the development team before scoring starts.

Accept and revise

Accept. The artist would give this look to the user as the Seint starting point. A small opacity change is allowed.

Revise. The artist would change a product, a placement, or a decision.

The frozen set

• At least 40 faces.
• At least a quarter of faces at Monk 7 to 10, and at least one face in every Monk band.
• Cool, warm, neutral, and olive undertones.
• At least five faces with a colour cast under the eye, five with a shadow from structure, and five with textured under-eyes.
• At least five faces over 50, and five with pigmentation or melasma.
• The issue types listed under face analysis, and some faces with none of those issues.
• Photo mode is scored on the full set.
• AR is scored on at least 15 live scans taken from that set. Live tracking fails for different reasons, and 40 live scans are not required for this proof of concept.

Checks

Required must pass. Optional is scored, and a miss does not fail the proof of concept. Percentage bars move in steps of 5.

How a row is read

Required. The row must pass. A module may be connected to the others when its required minimum rows pass.

Optional. The row is scored and shown. A miss does not fail the proof of concept.

Minimum. The module behaves correctly on the stated cases.

A percentage. Measured on the shared set after the modules are connected. Bars move in steps of 5. A straightforward check sits at 95 or 100 percent. A hard check that still decides the look sits at 90 or 95 percent. A hard check that can be a little less precise sits at 80 or 85 percent. The proof of concept is not complete on a minimum pass alone.

Each module page lists its own rows, with the same colours.

Activity

Configure

Work that can be prepared before anyone opens a camera. A session uses only what these four editors have published.

Four editors

What stays fixed

The element list does not change during the proof of concept. Products, level sets, stencil weights, depth ceilings, usage lines, theory text, and the base profile do change, through these editors, without a code deploy.

Activity

Experience: AR

A live session. The user can change today's brief, then the system captures four poses, reports what it saw, composes a look, and paints it on the live face.

Wait targets are requirements. They can be changed by agreement. They cannot be left undefined.

Steps

1. Brief

   User sees. The saved profile, and controls for today's occasion, mood, coverage (overall, and optionally per zone), application order, and a note.

   System. Overrides are stored on the session. The base profile is unchanged.

   The user sets the pace

2. Capture

   User sees. Direction for four poses: straight, left, right, chin up. Hair off the forehead. Neck in frame. No flash.

   System. Each pose is kept or rejected by the quality gate, with a reason.

   During capture

3. Analysis

   User sees. Short statements as observations are ready. Not an empty waiting state.

   System. Face analysis on the kept frames.

   10 seconds or less

4. Face report

   User sees. Skin level, undertone, face shape with its confidence, issues on a map (each dark area marked as a colour cast or a shadow), under-eye texture, and anything uncertain. The user can correct the report, for example by marking freckles to keep.

   System. Corrections are written onto the session before composition. Any Find My Shade answer still missing (skin type, coverage, finish, age, and level or face when colour was unreliable) is asked here, in the quiz's wording. This is gate G1.

   The user sets the pace

5. Composition

   User sees. Two short, editable summaries while the look is built: the complexion (highlights, contour, and any correction), then the optional colour and glow. Each has "Looks right, continue" as the default. They wait for the user only when a check is not confident; otherwise the look carries on.

   System. The composer steps, with gates G2 and G3. The render check uses a snapshot of the current frame.

   First card in 3 seconds or less. Full look in 25 seconds or less of system time, not counting time on the cards.

6. Live look

   User sees. The face, live, with the look applied. A before-and-after control.

   System. Banuba tracks the face and paints the blended look from its baked textures. Eyes, brows, and lips appear live only where that area is tracked with confidence; otherwise they are shown on a photo, with a note.

   The user sets the pace

7. On-screen edit

   User sees. Element on or off, opacity, zone weight, and one step lighter or deeper.

   System. The render updates. No model is called.

   Immediate

8. Chat edit

   User sees. A text field. The change matches the request: one setting, one element, or a new plan. Undo is available.

   System. The request is classified. The classes are defined on the composer page.

   Direct: immediate. One element: 12 seconds or less. New plan: 25 seconds or less.

9. Result to keep

   User sees. The application steps in the user's chosen order, the matching official collection (No. 4, No. 8, or No. 12) where one fits, and a list of products already owned and products to buy.

   System. Saved with the theory version used to build the look.

   The user sets the pace

Activity

Experience: photo

The same decisions, the same report, and the same editors as AR. Capture and display are the only differences.

Steps

1. Brief

   User sees. The saved profile, with the same overrides as AR.

   System. Overrides stay on the session.

   The user sets the pace

2. Photographs

   User sees. Straight is required. Left and right are optional. A neck photograph is offered and optional. A rejected photograph shows the reason.

   System. The quality gate runs on each file. The depth match always uses the face. If the neck is not visible, the later report says the face-to-neck check was not possible.

   The user sets the pace

3. Analysis

   User sees. The same progress text as AR.

   System. Analysis of the photographs the user kept.

   10 seconds or less

4. Face report

   User sees. The same report as AR, including corrections.

   System. Corrections are written onto the session before composition. Any Find My Shade answer still missing (skin type, coverage, finish, age, and level or face when colour was unreliable) is asked here, in the quiz's wording. This is gate G1.

   The user sets the pace

5. Composition

   User sees. The same two cards as AR.

   System. The render check uses the rendered straight photograph.

   First card in 3 seconds or less. Full look in 25 seconds or less of system time, not counting time on the cards.

6. Look on the photographs

   User sees. The look on each uploaded photograph, with angle switching and before-and-after.

   System. The same JSON, rendered once per photograph.

   The user sets the pace

7. On-screen edit

   User sees. The same controls as AR, redrawn on the stills.

   System. No model is called.

   Immediate

8. Chat edit

   User sees. The same editor as AR.

   System. The same three classes.

   Same targets as AR

9. Result to keep

   User sees. The same guide and product list.

   System. The same saved plan.

   The user sets the pace

Activity

Evaluate

The activity that makes each module testable on its own, and makes the acceptance bars meaningful.

What it contains

• A test setup for each module, with fixed inputs and a visible result.
• The frozen face set, including artist reference notes.
• A review in which the artist marks Accept or Revise and gives a reason. The artist does not see which looks were made by the system and which were references, where a blind review is possible.
• A blind comparison of three looks per face: the system's look, the artist's reference, and the quiz-only look. The system should beat quiz-only and come close to the artist.
• An artist-to-artist agreement check, which sets the ceiling for every bar.
• A score report for the run. The report is the acceptance record.

Rule for a change

A change to theory, retrieval, stencils, or render values is kept only if the score stays at or above the bar. If a supporting check falls, the change is reverted. The theory page states the same rule for a published theory edit.

Main difficulty

Without artist reference judgments, the percent cannot be calculated. The set also has to include deeper skin and poor lighting. A set of only clear, evenly lit, light faces will overstate the result. Deeper skin needs its own quota, Monk 7 to 10, because image colour is least reliable there and a brightener can turn ashy.

Module 1. Experience

Face analysis

Turn a guided capture into a structured description of the face, neck, and skin. Each claim carries a confidence. This module does not choose a product.

What it produces

• Skin level from I to XII, as a distribution with the most likely level, measured from face skin with shine excluded.
• Undertone: cool, warm, neutral, olive, or undetermined, with a confidence. Undertone is read relative to other skin at the same level. Undetermined is a valid result.
• Face and neck: whether the neck is darker or lighter than the face, when the neck is visible. This is a check, not the match target.
• Face shape as probabilities over round, heart, oval, oblong, rectangle, square, pear, and diamond, read from the landmarks. Shape changes contour sub-zones only when the top shape is at least 0.55 confident. A round or heart face gets more contour at the temples.
• Tone for forehead, temples, both cheeks, under-eye, nose, chin, jaw, and neck when visible.
• Skin behaviour: oily (the share of shine in the T-zone), dry, combination, mature, textured, or unknown. A separate under-eye texture reading.
• Lip colour and a short lip texture note. Eye, hair, and brow colour.
• Findings. Each finding has type, region, position, size, severity, confidence, and whether it is worth covering. Types: redness (mild or vivid), brown pigmentation or melasma, raised blemish or acne, dark circles, fine lines and texture. Dark areas are also classified as a colour cast, with its hue (blue, purple, green, muddy), or a shadow from structure, and carry their darkest point.
• For each finding, a predicted result after each correction step, so the composer can stop early.
• Age markers: jowl, hooded eyes, under-eye hollow. Cheekbone position, with a confidence.
• Notes from a fixed list only: where light and shadow already are, natural flush kept separate from redness, freckles or marks to preserve, hair and brow colour lean, fine lines, and asymmetry to leave alone.
• Age. If the profile already has it, the profile wins. A guess from the photograph is stored only as a low-confidence hint. Gender is not collected or estimated.
• The pose or photograph each value came from.

What is required, and what is optional

The proof of concept needs only what Find My Shade collects. Everything else fine-tunes the look, and its absence never blocks composition. If something essential is missing, the user is asked before anything is composed.

Each measured value has a confidence threshold. Below it, the value is treated as unknown: a required value becomes a question, and an optional one falls back to the base default. An optional question is asked only when the answer would change this look, for example eye shape only when the eyes will be made up.

What an artist reads, and the signal for it

Colour cast or shadow

A dark area is either a colour cast (blue, purple, green, or muddy), which a complementary corrector can filter, or a shadow from the bone structure, which only light can lift. The analysis measures the hue shift and lightness drop against the skin around it, checks whether the area sits in a hollow, and records the darkest point, where an artist starts. A shadow never gets a corrector.

Capture quality

Before analysis, the capture is checked (gate G0) for light, white balance, blur, flash, and makeup already on the face. A failed capture is retaken or stored as unreliable. It is not analysed as a normal result. The gate also records how far the colour can be trusted. When it cannot, the composer uses the quiz answers alone, which reproduces Seint's own recommendation, instead of guessing.

In AR, the four poses are straight, left, right, and chin up. A failed pose is retaken or recorded as missing, with the reason. In photo, analysis does not start without a usable straight photograph.

Difficulties

1. Phone cameras change colour with the light, white balance, flash, and beauty processing.
2. Artists ask to see the neck, but the main highlight is matched to the face. The neck only decides how the jaw contour bridges the two. Many photographs crop the neck, so the report must say when the check was not possible.
3. A shadow on one cheek can look like a dark spot. Hair at the hairline can look like forehead spots. An eye in shadow is not a dark circle. A dark under-eye can be a colour cast or a shadow from the bone structure, and they need different products.
4. Redness, a pimple, and natural flush look similar and lead to different products. They must be separated.
5. Unstructured notes cannot be used reliably. The note list is fixed for that reason.
6. Age estimated from a face is unreliable. The profile is the source of truth.
7. AR frames are less even than a still photograph the user selected.
8. Colour from a phone is least reliable on deeper skin, where the set needs its own quota.

Acceptance

Test setup and records

Test setup. Load a photograph or a recorded AR session. Show regions, findings, and confidence. Score a folder of labelled faces with one command.

Reads a capture set. Writes a face analysis, used by the report, the composer, and the renderer.

Module 2. Configure and experience

Profile

A saved description of the user, plus what the user says in this session. The session wins. The system still produces a look when the user provides very little.

Fields

• Saved before a session: age, season, skin conditions, allergies and ingredient limits, and the Seint palette of products already owned.
• Set or overridden in the session: mood, occasion, coverage, and free text.
• Stored when provided: undertone quiz answers and how the undertone was judged (veins, jewellery, draping), vibe, finish, comfort with a made-up look, a keep list, time and skill, and the setting the look has to last in.
• Find My Shade answers, stored exactly, so any of the 6,912 quiz paths can be kept without loss.
• Concerns the user names. They start the correction ladder and are confirmed or rejected by the face analysis.
• Under-eye texture (smooth, some lines, crepey), optional.
• Coverage overall, and optionally per zone. Under-eye values are capped by the depth ceilings.
• Application order preset: highlight first, contour first, main first, Cheetah dots, or one brush. It changes the guide, not the look.
• Intensity: Natural, Everyday, Full, or Glam, asked at the first confirmation with Everyday pre-selected. The same face gets different stencils and opacities.
• Features to enhance or soften, and colour likes and dislikes, such as a lip family or no shimmer.
• Time and skill map to collection No. 4, No. 8, or No. 12.
• Free text is converted to a short list of constraints and shown to the user before composition. The user can remove a constraint.
• Allergies are hard filters. A blocked product cannot appear in a look.
• The palette is a preference. The look stays inside it when it can. If it steps outside, the plan states the reason.
• Source order, highest first: this session, the saved profile, a hint from analysis, the system default. This holds for identity and preferences. For values the camera also measures (level, undertone, skin type), the stated and measured values are both kept and reconciled by the theory's conflict rules, and a large disagreement is shown to the user. Every field stores which source filled it.
• If the user provides nothing, the default intensity is Everyday, with the coverage the quiz answer implies, light to medium unless stated. The plan states that the system chose them.
• A consent flag for storing photographs and the analysis.

Difficulties

1. Many users will skip fields. The look must still be valid, and it must say that defaults were used.
2. Free text can contradict a selected option.
3. The owned palette may not contain a shade that matches the face. Staying inside it without a check produces a poor look.
4. Allergy filters need ingredient data. Early product records may not have it.
5. Skin conditions are health information. Storage needs a retention rule, not only a field.
6. Stated and measured level often disagree. Users tend to choose a lighter level than they are.

Acceptance

Test setup and records

Test setup. A form with a live view of the merged session and the source of each field. Fixtures: complete, empty, contradictory, and text only.

Writes a client context, used by face analysis for quiz hints, by inventory for filters, and by the composer.

Module 3. Configure

Inventory

The digital record of in-scope Seint products: what each product is, when it is used, and how it should look on skin.

Contents of a product record

• Identifier, name, official descriptor (the depth and undertone label, for example "MEDIUM - OLIVE UNDERTONE"), element roles, finish, level-set membership (which No. 4, No. 8, and No. 12 collections and which quiz faces include it), colour from the official photograph and from the site swatch, use-case text, ingredient notes where known, availability or retirement with a family substitute, an eye-safe flag, and render values.
• The proof-of-concept catalogue holds every product line that an element in the stencil library can use: 232 records, listed by element in the table below. Cream Correctors resolve for the correction element only. Balm is the illuminator at Level XI faces C–D and Level XII faces A–D.
• One Lip + Cheek product can be retrieved for the cheek and for the lip.
• Use cases are written as situations: who the shade suits, which depth, which undertone, and which problem. They are stored in Postgres and searched with pgvector. There is no separate vector database.
• Retrieval order: hard filters (element, availability, allergy, palette preference), then shade scoring on depth, undertone, and pigment, then the vector match, then a final ranking. The test view shows each stage.
• A shade graph links lighter, deeper, warmer, and cooler neighbours, ordered by photographed lightness, not by the depth label. Seint's labels and levels are not strictly in lightness order, so "one shade deeper" is a step on this measured ladder, not a new search. Contour neighbours also carry a chroma check: a contour must be less colourful than the main highlight.
• Render values are tuned per skin-depth band, and differ by role. A single hex value is not sufficient.

Products for every element

Every element in the stencil library, the Seint products it can use, and how many stencils it has. Shade counts are the records in the proof-of-concept catalogue, taken from seintofficial.com on 29 September 2026.

The catalogue also holds Saved Salve (skin care). It is not an element and is not drawn. The stencil library has 87 stencils in total. The elements page lists them.

Not Seint products, or not on sale. Every element maps to a current Seint product line; none needs another brand. These names are flagged so they are never offered as Seint products:

• "Blush" is not a Seint product name. The powder blush switch uses Pressed Pigment.
• Seint sells no concealer and no mascara. Highlight and Cream Corrector fill the concealer role. Mascara is not modelled.
• Plush Prep Mist was sold out on 29 September 2026. The guide falls back to Fixate or Setting Spray, or leaves spray out and says so.
• "Lip Lush" is not a product. It was a misreading of Blackberry Lip Flush.
• Retired shades (for example Petal, Scarlet, Carousel, and Tiger Lily) are not offered. A shade from the same family is suggested instead.
• Brushes, sponges, and the spoolie are tools. They appear in the guide as text only.

Difficulties

1. The commercial catalogue changes. The boundary on the scope page is what keeps this finite.
2. Colour in the tin is not colour on skin, and it changes again across skin levels.
3. Vector search alone will confuse neighbouring shades. Shade choice is primarily a distance on depth and undertone.
4. Retrieval can be called accurate only against situations an artist has labelled.
5. Missing ingredient data weakens the allergy filter.
6. Tuning render values is slow visual work. It needs an ordered queue or it will consume the schedule.
7. Colour today comes from photographs and swatches, not instrument measurement. Production needs measured colour and opacity for each pan.

Acceptance

Test setup and records

Test setup. A query view that shows every retrieval stage, and a preview of the swatch beside the render on two skin depths.

Writes a candidate set per element for the composer. The renderer reads the render values.

Module 4. Configure

Elements and stencils

An element is a Seint role. A stencil is a placement that role is allowed to use. The element list is fixed. Stencil text and weights are editable.

The elements

Each paragraph below is stored on the element and is given to the composer when that element is decided.

Placement map

These are real renders on one test face (f010, from the Face Research Lab London Set, CC BY 4.0), made with the proof-of-concept stencil library. Each stencil is rebuilt from the face's landmarks, so the same definition works on any face. The stencil tables below list the full library.

The renders below show each element on its own. They use marker colours so that the stencils can be told apart; these are not the product colours. Select a render to open it at full size.

• Dashed outline: the placed seed, where the product is first put down.
• Coloured fill: where the product ends up after blending. It is strongest where the product is densest.
• Thin solid line: the outer edge of the blend, where coverage falls to a tenth of its strongest point.

The full look, with the eyes, brows, and lips, is on the blending page.

Minimum stencil library

Stencils are tied to facial landmarks, with separate weights for sub-zones, so one definition is used in AR and on photographs. Contour exposes weights for hollow, temple, jaw, and neck. Stencils carry their gradient (the cheek contour is strongest at the ear) rather than a flat fill.

The full proof-of-concept library has 87 stencils across the complexion, cheek, corrector, eye, brow, lip, and powder roles. Each is defined from landmarks in eye-distance units, so it can be rebuilt on any face. Each has variants for depth, face shape, age, texture, asymmetry, and findings (dark circles, redness, pigmentation, hooded, monolid, or deep-set eyes). Each also has a default opacity range, the stencils it must not overlap, the face-analysis inputs that select it, and a blend recipe. The composer sets an opacity for each stencil it uses. The opacity is the placed deposit: how much of the skin the product covers where it is put down, before blending.

Blend styles

Each stencil names how its product is blended, so the render can mock the brush. The style changes the edge and the spread of the deposit, not where the stencil sits.

Each stencil also carries a full blend recipe: the tool, the direction and distance the product is moved, the number of passes, and the edges it must stop at. The blending page explains it, element by element.

Face shape and depth ceilings

Face shape selects and weights contour sub-zones. A round face adds temple and jaw. A diamond face adds the top of the forehead and the tip of the chin. A pear face uses the jaw only. The library itself is the same for every shape. Shape is applied only when the analysis is at least 0.55 confident, and an artist or user can always override it, because Seint artists warn against a one-size-fits-all diagram.

A ceiling caps the product placed in that zone. Blending then spreads it, so the finished coverage is lower.

Difficulties

1. A stencil that is correct on a front photograph will drift in AR when the head turns, unless it is tied to landmarks.
2. The jaw and neck often fall outside the face mesh Banuba paints. Neck continuity is a separate rendering problem. If the neck is not painted, the report must not say that it was.
3. Face shape changes which contour sub-zones are used and how heavy they are. It is the weakest signal in the analysis, so it is applied last and only when confident.
4. A localized stencil has to meet a small finding. A mask the size of the cheek is not localized.
5. Cream Correctors are about 10 percent opaque. Drawn too strongly they look like paint. Drawn honestly they are barely visible. The guide must still name them.
6. Under-eye and cheek contour are where placement most often fails: a triangle that is too bright or too low, or a contour that sits too low and drags the face down.

Acceptance

Test setup and records

Test setup. The manager, with the mask drawn on reference faces: eight face shapes, a front pose and a turned pose, one lighter skin and one deeper skin, and one textured under-eye.

Writes element and stencil records, used by the composer for what is allowed and by the renderer for what is drawn.

Module 5. Configure

Artist theory

The written Seint method. Artist Theory v2 is the starting text: one cream layer, the IIID roles, the placement map, the order of decisions, and how an artist chooses.

What the module holds

• The v2 document, labelled with a version, as the text the composer consults.
• An editor in the application. Publishing creates a new version. No code deploy is required.
• Two uses of the same document. The prose is given to the generative steps. A rule list taken from the same document is checked by the rules gate.
• Every look stores the theory version it was built from.
• A draft can be replayed on a fixed set of faces before it is published, so the Seint team can see which looks changed.
• A rule list in which every rule carries a principle, a strength (hard, strong, soft), and its source, with an evidence label: sourced (Seint or a named artist), inferred (our design), or secondary (from video notes). The composer may state a sourced rule as Seint teaching, and must present an inferred rule as the system's choice.
• A conflict log. Where Seint sources disagree, for example cheek contour on or under the bone, or bronzer before or after blush, the default and the alternative are both recorded.

Foundations

Elements

Rules checked on day one

1. Bronzer is never used as contour.
2. The main highlight matches the face. It is chosen before contour, and before everything else.
3. Contour is less colourful than the main highlight, and reads as shadow.
4. No lighter shade on a dark spot. A darker spot gets a deeper correcting highlight.
5. A corrector is used only for a colour cast that remains after the highlight steps, never on a shadow from structure. At most three correction zones.
6. Under-eye depth stays within 0.60 for brightening and 0.45 for the correcting highlight. The Cream Corrector stays at about 0.10.
7. No shimmer or illuminator on fine lines or textured zones.
8. Illuminator follows depth from the level set, and Balm replaces it at Level XI faces C–D and Level XII faces A–D.
9. Lip and cheek stay in one colour family.
10. Cheek contour stops at the outer eye corner and never moves toward the mouth.
11. Intensity stays within each stencil's allowed range, never above 1.
12. Missing evidence is not recorded as a skip.
13. The plan uses only catalogue identifiers.
14. No geometric morphing.

Acceptance

Test setup and records

Test setup. Markdown editor, version list, rule list, and replay with a plan diff.

Writes a theory version: prose sections and a rule list.

Module 6. Experience

Look composer

The Look Composer agent starts from the brand base (the quiz palette and the No. 4, No. 8, and No. 12 sets), fine-tunes it from the face analysis in dependency order, and records a reason for every change. For each element it decides the products, the stencils, and an opacity for each stencil: how much product is placed, before blending. This mirrors how an artist delivers a cart and a face map. The Look Editor agent then changes the look by the smallest action that satisfies the user's request.

Look Composer agent

This graph is the agent's loop. Solid edges are the forward order. Dashed edges are the retries. Code does every lookup, number, and rule check. An LLM agent makes the look decisions in four steps. jev answers the atomic questions at the gates. A vision model reads the photo and checks the render. Select a step to read its job.

The generative model sits behind one interface, so its identifier (the site's gpt 6 luna) is a configuration value and can change without changing the sequence.

Steps in dependency order

Decision order follows what each choice depends on. Layer order is the order in which the product goes on the face and is drawn. They are two separate lists, stored as separate fields, and both are on the theory page. The steps below group the decisions: S4 chooses the main, the brightening highlight, and the contour; S5 then runs the correction ladder against the main. Contour and correction do not depend on each other.

Why an agent, not fixed code: the theory has 77 rules and 37 thresholds, several of which conflict by design, across about 30 face attributes and 87 stencils. A fixed decision tree cannot state its reasons in the user's terms or handle the combined cases. The agent chooses only among options that code has already computed and filtered to the catalogue, and it never does the maths.

Why not fewer steps: correction needs the chosen main highlight, cheek colour needs the contour line, and lips need the cheek. Why not more: one step per element would add prompts and waiting without a new dependency, because eyeshadow and liner share one eye look, and bronzer, blush, and illuminator share the cheek area.

Gates

Each gate runs code checks first, then jev questions, then shows the user one card with "Looks right, continue" as the default. The gates sit after the decisions that later steps depend on. G1 and G4 always ask the user. G2 and G3 carry on by themselves only when all three hold: every code check passes, no required attribute is missing, and jev scores the gate at 0.85 or higher (the lowest confidence among its answers). Otherwise they ask. Either way the user always sees an editable summary, and an edit re-runs the step that owns it.

About jev. jev is TypeSafe's judgement model (currently jev-1.13.0). It takes a text or JSON state and a set of typed questions, and returns one typed answer per question: yes or no (noul), choice, or score. Choice and score answers carry a confidence. jev does not read images and does not write text. So the vision model checks the render, the agent writes every reason, and jev checks the reasons. TypeSafe has no separate gate object. A gate here is code checks, one jev request with one question per item, and thresholds in our code: below 0.6 confidence the issue is shown or the step is retried, 0.6 to 0.85 shows "please check", above 0.85 passes. These starting values follow TypeSafe's confidence-routing example and are tuned on the frozen set.

Prompts built per step

There is no single large prompt. Each agent step gets a template, plus fragments included only when they apply: only the elements, products, and stencils that exist in the catalogue and were selected for this user. A step never carries logic about something it cannot choose. Each fragment names the theory rules and book sections it states, so a changed rule points to the fragments that must change.

What the plan contains

• The look JSON: element, decision, products, stencils with a placed opacity and a blend recipe for each, zone weights, and the complexion role and coverage maps.
• A reason for every element, including skip. Every reason cites the theory rule it follows.
• For each finding, the ladder step chosen and why it stopped there.
• Where the plan left the brand base, and why, with the rule and the finding.
• Two to four alternatives.
• Application steps in the theory's order.
• A product list split into owned and to buy.
• The theory version, and the result of every gate.

Correction ladder

Artists treat a spot with the smallest step that works. The composer does the same for every finding, and records where it stopped. This is the concealer role: Seint sells no separate concealer, and the Highlight product is described as its concealing shade.

A shadow from structure leaves the ladder after step 1 and gets brightening instead. Under the eyes, every step stays within the depth ceilings. A lighter shade never goes on a dark spot.

Look Editor agent

One conversational agent handles chat in both AR and photo. It classifies the message, then takes one of three exits. On-screen toggles and sliders stay outside this agent and do not call a model. Looks are versioned. A chat change can be undone.

Difficulties

1. If the user sees nothing until the end, the wait is unclear. The time targets on the experience pages are the limit.
2. A model will invent shade names unless an unknown identifier cannot be saved.
3. The vision check will sometimes pass a poor render and sometimes fail a good one. The retry limit and the visible failure are required.
4. Classifying "less contour" as a new plan is a defect.
5. Model names and prices will change. They cannot be written through the feature code.
6. Elements decided in separate steps can conflict. The overlap checks at each gate and the harmony question at G4 exist to catch that.
7. Too many confirmation cards tire the user. There are four, each with a default. G2 and G3 pass on their own when every check is confident, and still show an editable summary.
8. jev reads literally and should not do arithmetic. Code turns distances and counts into named buckets before jev sees them.
9. Over-correction is the common artist complaint. The ladder and the depth ceilings exist so the system does not reach for a corrector or heavy coverage when the main highlight is enough.

Acceptance

Test setup and records

Test setup. Run the graph on a frozen face analysis and profile. Inspect each step, the cost, and the time. Replay the edit set.

Reads face analysis, client context, the brand base (quiz prior and level sets), candidate sets, element and stencil records, prompt fragments, and the theory version. Writes a look plan and the gate results.

Module 7. Experience

Look renderer

Apply the look JSON to a face. On photographs our own compositor draws and blends the stencils; in live AR Banuba tracks the face and paints the baked textures. This module does not choose a product or a placement.

Behaviour

• The same JSON produces the same products and the same regions in AR and on photographs.
• The cream complexion (main, brightening, and correcting highlight, contour, Lip + Cheek, bronzer, and illuminator) is one cream layer. Each product is placed, blended in layer order, and the colours are mixed once. The application order in the guide therefore cannot change the result, and blending keeps each role's gradient.
• Layers are drawn in the layer order on the theory page, lowest first: Cream Corrector (a colour shift under the highlight), correcting highlight, main and brightening highlight, contour, Lip + Cheek, bronzer over the blush, illuminator, setting powder, powder blush (Pressed Pigment, if chosen), eyes (eyeshadow and liner), brows, lips, and setting spray.
• Each stencil is blended with its own recipe (tool, direction, distance, passes, and stops), so it looks blended by the right brush. The opacity in the plan is the product placed, not the strength after blending. See the blending page.
• Eyes, brows, and lips are drawn in AR and on the photo when the face analysis tracks that area with a confidence of at least 0.70 (a starting value, not yet tested). Otherwise they are drawn on the photo only, and the user is told why.
• No geometric morphing. Any skin smoothing is minimal and disclosed.
• Without a new compose, the user can toggle an element, change opacity, change a zone weight, and move one step on the shade graph.
• Before and after is available in both modes. On photographs, each uploaded angle can show the look.
• The renderer can produce a still snapshot for the vision gate.
• A localized layer is drawn at the finding's position, not as a wash across the cheek.

How each role is drawn

Rendering path

• Photographs use our own compositor, which draws the stencils exactly. It is also the reference for evaluation.
• Live AR uses Banuba stencil textures once that path is validated. Until then it uses Banuba's built-in regions, and the screen says placement is approximated.
• The two are never mixed for the same element.
• Banuba accepts one texture per element, so all contour stencils are merged into one texture.
• For AR, the blended look is baked into textures on the face mesh, one per element and one per layer group. They are the same blend as the photo, and they are baked again only when the look changes.

Difficulties

1. A correct product id still looks wrong if the render values are a flat tint. Matching Seint cream is calibration work.
2. AR and photo will diverge if each mode keeps its own masks. Both must read the stencil record.
3. Neck contour may have no surface to paint if the mesh ends at the jaw. The system must not claim the neck was painted when it was not.
4. A spot mask will slide between frames unless it stays attached to landmarks.
5. Rebuilding the whole effect on every slider move will drop frames on a mid-range phone.
6. A brown contour multiplied onto skin drifts orange and over-darkens deep skin. Contour must be checked for hue as well as depth.
7. Banuba's texture path is deprecated and its face layout is undocumented.

Acceptance

Test setup and records

Test setup. A JSON view with the photograph and a recorded AR clip side by side. Saved plans can be rendered again and compared when render values or stencils change.

Reads a look plan, stencil geometry, and render values. Writes the live render and a snapshot.

Module 7. Experience

Blending

How the renderer turns placed product into a blended look. The stencil says where the product is put down. The blend says how an artist then moves it. This is model BL-2, approved on 29 September 2026 and proven on one test face (f010).

The model in plain language

1. Seed deposit. Each stencil is put down first, as an artist places a stick or a dense brush. Its opacity is the placed deposit: how much of the skin the product covers where it lands. At the default Everyday intensity, cheek contour is placed at 0.90.
2. Direction-aware spread. Each stencil then moves the way its brush would: along the swipe, and toward a point on the face. Cheek contour is pulled up and back toward the top of the ear. Lip + Cheek is pulled toward the temple. Jaw contour is pulled down the neck, and hairline contour into the hair. Each stencil names its tool, how far it spreads, how much further along the swipe than across it, how far it drifts, and how many passes it gets. More passes soften the edges without spreading further.
3. Nothing is added or lost. Blending only moves product, so a wider blend is also a softer one. The strength after blending follows from that. Cheek contour placed at 0.90 ends at about 0.40 at its strongest point, spread over four to five times the placed area.
4. Hard stops. Some edges are walls that product never crosses: the eyes and lash lines, the lip line, the nostrils, the ears, and the background. Other edges fade the product out, so no line is drawn: the nose and the area around the mouth for cheek colour, the brow hairs, the outer edge of the brightening zone, the contour for illuminator, and the hair.
5. Passes in layer order. The layers are blended in the order they are applied: Cream Corrector, then the correcting highlight, then the cream complexion (main and brightening highlight, contour, and Lip + Cheek, all placed first and then blended), bronzer over the blush, illuminator, setting powder, powder blush (if chosen), eyes, brows, and lips. Setting spray goes on last and is not blended. A later cream layer slightly softens the wet cream under its brush, as a real brush does.
6. Pigment mixing. Where cream products overlap, their colours mix as pigments do, rather than being laid one on top of another. Skin texture shows through: about half of the skin detail under the complexion and lips, 90 percent under the eyes, and all of it under the brows. The main highlight is matched to the measured skin, so the complexion reads as the person's skin, evened.
7. Baking for AR. The blended result is baked into textures on a standard face mesh: one per element, and one per layer group (cream, eyes, brows, lips). Banuba then draws them on the tracked face in every frame. Nothing is recomputed per frame. The textures are baked again only when the look changes.

How each element is blended

Sourced means the direction or end point comes from Seint or a named Seint artist. Secondary means it comes from notes on a tutorial video. Inferred means it is our own construction. Every distance and number is inferred: each was tuned by eye on f010 and checked against the face's landmarks and the zones in the theory.

What changed from BL-1

In BL-1 an opacity was the strength after blending. Contour at 0.30 therefore put very little product down, and the blended look on f010 was almost invisible. BL-2 makes the opacity the amount placed, which is how an artist thinks about it, and blends further, to where artists say each product ends up. All 23 reach checks now pass on f010 (BL-1 passed 22).

Area after blending is the blended area divided by the placed area. The figures come from the element renders on f010.

Limits

• Tuned by eye on one fair face with muted shades (Nude and Astoria). The sculpt is soft, not editorial. Artists have not reviewed it yet, and it has not been run on the other test faces.
• Contour and Lip + Cheek are placed close to full coverage (0.90 to 0.97), so the stronger intensities differ mainly in how much product there is. That shows after blending as a wider, deeper shadow or flush.
• The reach checks measure where product arrives, not how it looks. The visual review is the other half of the check.
• Product blended into the hair or down the neck falls outside the face mesh, so AR shows only part of it (30 percent of the contour and 44 percent of the bronzer on f010).
• Photo colour is not calibrated, and the pigment mixing uses a single constant per product.
• Brows and lips have no tracking confidence of their own in the face report yet.

Integration

How the modules connect

Modules meet only through the records below. A module test is complete when it can run on fixtures of its input and produce its output. Integration is complete when the demonstration runs with nobody pasting JSON between steps.

Session flow

Records

Integration

Demonstration

The script that must run without hand-edited JSON. Passing this script is integration. It is not, by itself, the acceptance bars.

Prepare once

1. Publish Artist Theory v2.
2. Load the in-scope catalogue, use cases, and render values, and the Find My Shade map.
3. Confirm the elements and the stencil library.
4. Save a base profile with a palette and one allergy.

Then, with one person

1. AR: override the occasion, complete four poses, read the report, correct one item, wait for the look, and see it live.
2. Turn contour off and on. Change highlight opacity.
3. Ask for a softer jaw. Ask for an evening version. Undo once.
4. Photo: straight, plus one side, and no neck. The report must state that the face-to-neck check was not possible. The look renders on both photographs.
5. Leave with a guide and a product list that respects the allergy and prefers the palette.
6. Repeat the photo run on a face with a brown spot. The plan stops at the main highlight or the correcting highlight, not at a corrector.
7. Repeat on a face with a textured under-eye. The plan uses corner accents, not a crescent.

Integration

Responsibilities

Who supplies the material, who builds the system, and who signs the result.

Division of work

Integration

Open decisions

What has been decided, and what still needs an answer before the face set is frozen. Until an open item is decided, the scope page states the working assumption.

Decided on 29 September 2026

Achin Kumar approved these decisions.

1. Gates G2 and G3. They pass on their own only when every code check passes, no required attribute is missing, and jev scores the gate at 0.85 or higher. Otherwise the user is asked. An editable summary is always shown. G1 and G4 always ask.
2. Two orders. Decision order (which choice depends on which) and layer order (how the product goes on and is drawn, 13 layers) are two separate lists, stored as separate fields. Both are on the theory page.
3. Powder blush. The cheek product is a switch between Lip + Cheek (the default) and Pressed Pigment, the powder blush. The composer suggests Pressed Pigment first only with a stated reason, and shows that reason to the user. Seint sells no product named "blush".
4. Eyes, brows, and lips in AR. They are drawn in live AR only when the face reading for that area is confident enough (at least 0.70). Otherwise they are drawn on the photo only, with a short note on the final-look screen. The 0.70 is our starting value (inferred), not a Seint figure.
5. Neck. Face only, the Seint default. The main highlight is matched to the face, not the neck. When the neck is visible, the report says whether it is darker or lighter than the face, as a check, not the match target.
6. Application presets. The guide offers five: highlight first, contour first, main first, Cheetah dots, and one brush. A preset changes the guide text only, never the layer order or the render.
7. Cream Corrector (Demi) at 0.10 opacity, from 0.08 to 0.15 by intensity. Kelly Snider: "these Cream Correctors are unique with only 10% opacity". Heather Gilligan: "Seint Demi under eye have just 10% opacity". Seint: "apply a light touch of color directly to the places on your skin that need filtering". More coverage on a spot comes from the correcting highlight, not from more Demi.
8. Opacity is the placed deposit. The opacity in a look plan is how much product goes down where it is placed. Blending then spreads it without adding or removing any, so the finished strength is lower. The blending page explains the model (BL-2).

To decide

1. Which artists will rate, and whether they can review 40 looks plus 15 live scans.
2. Whether Cream Correctors are drawn in AR in this proof of concept, or named in the guide only. They are drawn in photo mode.
3. The reference phone for the frame-rate check.
4. How long photographs and analysis may be stored, and who can delete them.
5. Whether a look may recommend a product the user does not own, or must stay inside the palette.
6. Whether the product list links to purchase in this proof of concept, or only names the products.
7. Who is allowed to publish a theory version.
8. Whether the wait targets of 10 seconds for analysis and 25 seconds for a look are acceptable.
9. Whether bronzer is on by default, as in the official level sets, or opt-in.

Working assumptions

• The catalogue covers every element in the stencil library: Highlight, Contour, Lip + Cheek, Pressed Pigment, Bronzer, Illuminator and Balm, Cream Corrector, Setting Powder, and setting spray for the complexion, and Eyeshadow, Eyeliner, Pencil, brow products, and lip products for the features. All are products Seint sold on 29 September 2026. The list is on the inventory page.
• The brand base (quiz palette and level sets) is the starting point of every look. Departures carry a reason.
• Tools and setting products appear in the guide as text only.
• Cream Correctors (Demi) are catalogue data. The correction ladder decides when they are used.
• The palette prefers owned shades and may step outside with a written reason.
• Photographs and analysis are stored only with consent, for a period the Seint team sets.
• The neck remains part of the contour stencil even if the first AR build can paint only to the jaw. The report must say so when the neck was not painted.