A San Francisco home kitchen at 6:50 a.m. — the house is quiet, light is just arriving, the space feels like it is already thinking. The moment before the family wakes up.
Project 01 · Living Home

A shared grammar
for the home.

The Yamashiro family of San Francisco — Mai, Kiran, Aoi, Sota, and Leo — in their kitchen on a Sunday morning, soft light through the window, no one looking at the camera.
Meet the Yamashiros · San Francisco · 2026

The Yamashiros.

Two parents, three kids, five agents, one Context Brain.

  • Mai
    mom · 38 · designer

    Knows every ingredient in the fridge. Totoro nights with Aoi.

  • Kiran
    dad · 41 · founder

    6 AM Bangalore calls. Sunday curry from his mom's recipe.

  • Aoi
    eldest · 15 · they/them

    Wants to be asked, not assumed.

  • Sota
    middle · 11 · soccer

    Same lunch every Tuesday. Knows the exact bus to Tilden Park.

  • Leo
    youngest · 9 · curious

    Mild dairy intolerance. The family keeps a notebook of his unanswered "whys."

Names changed for privacy. Ages, family shape, allergies, and core interests are real — drawn from a Bay Area family during 2025–2026 portfolio research.

Executive Summary · 90 seconds

One Brain. Six surfaces. 76% of morning decisions, absorbed.

The problem
Five people, four single-purpose AIs, zero shared memory. Mai re-prompts every agent every morning.
The thesis
Context Grammar — a Brain (3 layers) + 8 Tokens + 2 Dials — becomes one shared grammar every surface in the home speaks. Fridge, phone, iPad, CarPlay, Nest Hub all read the same Brain.
What's new
  • Silent Resolution · 6 of 7 cascading actions finish without notification.
  • 3 Axes of Disclosure · translation, connection, archive — privacy as design, not a settings page.
  • Adaptive Autonomy · the same dial returns different answers based on stakes.
  • Care Architecture · the dial has a ceiling care doesn't cross.
  • Graduated Archive · memory unlocks on life events, not calendar dates.
The proof
Six surfaces, one Brain — Samsung Family Hub, iPhone, iPad, CarPlay, Google Nest Hub, Apple Watch. Eight scenes, all driven by the same 8 Token state. Each scene maps to a canonical AX Pattern.
It generalizes
The same 8 Tokens describe an enterprise sales team — manager × AI agent × pipeline. Same grammar. Different world. See Chapter 06.
Dimensions covered Trust Control Orchestration Disclosure Adaptation Cross-surface Tokenized DS
2026 · The state of the art

Advanced families already try —
with four lonely agents.

Today's workaround: four Mac minis on one desk, each running a single-purpose AI — Emily for homework, Sam for work, Nova for family, Oliver for the house. Four accounts, four Post-its.

None of them know about each other. Every morning starts from zero. A real answer — without a shared grammar.

A desk with three Mac minis labeled by hand: Emily — Homework agent, Sam — Work AI, Nova — Family AI, Oliver — Home AI. A monitor behind them shows an AI chat coordinating the family's weekly schedule.
Today, on thousands of family desks: four Mac minis, four handwritten Post-its, four isolated agents. One family trying to keep up.
Post-it · Mac mini
"Emily : Homework agent"
Separate account · can't see the calendar · forgets every time.
Post-it · Mac mini
"Sam : Work AI"
Has work files · blocked from personal · no crossover allowed.
Post-it · Mac mini
"Nova : Family AI"
Knows the family · doesn't know today · starts over every morning.
Post-it · Mac mini
"Oliver : Home AI"
Lives on the fridge · ignores the other three · answers alone.
From here on · the setup changes

Context Brain arrives.
The Home Brain boots up.

One shared memory instead of four isolated ones. One grammar instead of four Post-its. The rest of this story is what becomes possible once the Brain is set up — and only once.

The setup

Five Brains. Three layers each.

What looks like one "Home Brain" is actually five distinct memory entities — each with the same three-layer anatomy. Identity (who), Learning (patterns), Now (this second). The same shape that runs an enterprise of 200 people. Different cast. Same architecture.

Below: the universal three-layer anatomy. Next: the five brains it lives in.

Layer 01 · in every brain
Identity
the reservation book · updates in months
Who the family is. Names, pronouns, allergies, languages — the bedrock that rarely changes. Written once. Referenced every decision.
Sota · egg OK Leo · mild dairy Mai · Okinawan JP Kiran · IN / NO
Layer 02 · in every brain
Learning
the waiter's memory · 11 months of observing
Patterns the Brain has built from watching. Monday pickup rhythms, Sota's soccer days, Leo's slow mornings, Kiran's 6 AM Bangalore calls. The layer that earns trust over time.
The Yamashiro family on an SF park afternoon Mar 2019 · Sutro forest walk
Layer 03 · in every brain
LIVE
Now Layer
the waiter's observation · updates in seconds
Live values from the 8 Context Tokens. Mai's hands are full. Kiran's flight is delayed 40 min. The baby just woke up. 7:02 AM, foggy, 58°F.
hands: busy SFO: delayed cognitive: med
The five brains of a home

Five entities.
One anatomy.

A household is not an individual — it's a small organization. Five distinct memory entities, each owning the same three-layer anatomy. Aoi's pronoun update doesn't touch Sota's brain. Finance Agent can't read grades. The Trip Brain dies after Yosemite — and returns its learning home.

This is why "one Home Brain in three layers" was always a simplification. The reality is below.

Concept · Five Home Brains Same five-tier shape as Project 04's Five Brain Diagram. Household at the foundation, Person×5 and Domain×5 as the middle, Coordinator above, Event Brains (with three inner layers) at the top — graduating into the archive.

Brain 01 · Foundation · ×1
Household Brain
P4 analog: Org Brain
The permanent ground every other brain stands on. Address. Family structure. Allergies that bind everyone. House rules. Family history. Updates in years, not days.
Brain 02 · Judgment DNA · ×5
Person Brain × 5
P4 analog: Brand Brain × N
One brain per family member — Mai, Kiran, Aoi, Sota, Leo. Each owns its own identity, taste, autonomy preferences. Cross-brain reads only via Disclosure Dial.
Mai · designer · admin Kiran · founder · 6 AM calls Aoi · 15 · they/them Sota · 11 · soccer Tuesdays Leo · 9 · mild dairy
Brain 03 · Specialist Knowledge · ×N
Domain Brain × N
P4 analog: Research Brain
Per-agent specialist memory — what each agent has learned about its own domain over months. External data flows in continuously (statements, schedules, prescriptions).
Finance · budget · history Education · schools · grades Health · meds · allergies Calendar · 5-person sync Food · pantry · taste
Brain 04 · Orchestrator · ×1
Household Coordinator
P4 analog: Project Master Brain
Reads across all brains. Resolves conflicts. Three different lunches → one default pick happens here. Decides what to surface, what to suppress, what to escalate. Spawns time-bound Event Brains.
Brain 05 · Time-bound Execution · ×N
Event Brain × N
P4 analog: Project Brain
Born for a specific window. Learns intensely. Returns its learning to Household Brain. Then dies. The Trip Brain from Project 02 is one of these.
Yosemite trip · 4 days Aoi's school year · 9 mo Leo's birthday · 1 week Move · 6 weeks

Same five roles run a 200-person sales floor. Move the cast. The architecture stays. See Chapter 06 — same five brains, different world.

Brain · Coordinator · Agent

Brains remember. Agents act.
The Coordinator decides.

Five brains store memory. They never act on their own. A single Coordinator reads across all of them, makes the call, and dispatches to specialised Agents — Voice, Shopping, Notify — that actually do things in the world.

Below: the 12-second cycle of one decision. "Sota's soccer shoes are tight" arrives → 5 brains read in parallel → Coordinator decides "silent buy" → Shopping Agent orders → write back to brains.

Concept · Brain ↔ Coordinator ↔ Agent Three node types, three roles. Brain = passive memory (blue). Coordinator = active orchestrator (purple). Agent = action-taker (orange). The Coordinator is itself a special kind of agent — but its job is purely to read and dispatch, not to act on the world.

3 patterns at once · the same mechanism

For the full Brain vs Agent breakdown — including the restaurant metaphor, Q&A, and step-by-step trace — see the Brain & Agents explainer.

Layer 01 · Identity · present in every brain

Five people.
One shared source of truth.

The Identity Layer holds what doesn't change Tuesday to Tuesday: names, pronouns, allergies, languages.

Written once. Every agent — Finance, Education, Calendar, Home, Travel — reads from the same card.

Layer 02 · Learning · present in every brain · 11 months shown

Eleven months becomes
the shape of a Tuesday.

The Learning Layer watches for eleven months: Monday pickups, Sota's soccer, Leo's slow mornings, Kiran's 6 AM Bangalore calls, Aoi's orchestra Thursdays.

Small observations become the shape of a repeated Tuesday.

User UI · iPad The Learning Layer — 11 months across 5 family members. Mai's kitchen iPad shows the map the Home Brain builds quietly. Every pattern here shapes the next decision.
Layer 03 · Now · present in every brain

Eight signals drive every decision.
The family never sees them.

The Now Layer is where the 8 Context Tokens live — Cognitive Load, Physical State, Social Exposure, Priority Weight, Form Factor, Feasibility, Autonomy Dial, Disclosure Dial. They update by the second.

Mai would never look at a token gauge in the morning. As readers, we are seeing it so we can understand what the interface is hiding — on purpose.

BACKSTAGE · Now Layer
Mai does not see this. We are.

Backstage view Each gauge is one of the 8 Context Tokens. The family never opens this — they just feel the decisions that come out of it.

Chapter 02 · A Morning — the Yamashiro family's kitchen at 7:02 AM
Chapter 02
A Morning.
The house is already thinking before anyone says good morning.
Scene 01 / The kitchen · 7:02 AM

The fridge resolved
three different lunches.

Samsung Family Hub already handles single-person meal planning in 2026 — 6 profiles, allergy filter, AI Vision, Bixby. The fridge knows Leo's dairy limit and Mai's sodium target. That's the starting line.

Monday collides: Kiran (dad) is on day 12 of a liver-marker reset. Leo (soccer after school) needs 2× calories. Sota has a class birthday cupcake. Three people, one fridge. The Rule Engine composes one plan that respects all three — not three suggestions Mai has to reconcile.

Mai says "middle one." The plan hands off to her phone, adds a Trader Joe's stop in Maps if inventory's low, and pings Kiran only if his approval is needed.

What Context Grammar adds (that Samsung 2026 does not ship)
  • • Multi-person collision resolution — one plan, three people's constraints
  • • Cross-device orchestration — fridge → phone → car → store, peer-to-peer
  • • Household-level Brain — learns what this family ate last month and varies accordingly
  • • Per-person Autonomy Dial — Mai auto-confirms, Kiran needs a tap, kids never see the choice
Brain telemetry · why did it pick #2

Every decision leaves a trace.

The family never sees this either — but designers and engineers do. Every UI command the Rule Engine emits comes with a full reasoning trace: which Tokens fired, which rules matched, which alternatives were considered.

This is what makes the Brain debuggable instead of mystical.

Trace · fridge.lunch_pick 07:02:14.219 · 12ms
Decision Surface lunch idea #2 (bento · grilled chicken). Suppress #1, #3.
TOKEN ✓ cognitive_load = 0.78 (high) → density:minimal triggers +0.42
TOKEN ✓ priority_weight: leo.calories > sota.cupcake_offset > kiran.liver_target +0.31
RULE ✓ multi_person_collision_resolver(3 constraints) → composite plan with bento #2 +0.18
RULE ✗ surprise_swap_eligible? → false (Mai's mode = AUTO·EXACT, surprise pinned to Leo only) skipped
Alternatives weighed
#1 onigiri · 0.61 #2 bento · 0.91 ✓ #3 sandwich · 0.54
trace_id: 7f3a · committed view source · 11 events · 4 token reads · 2 rule fires

Backstage · Telemetry A trace from one decision in Mai's morning. When the family asks "why did it pick that?", the trace is the answer — for designers, engineers, and (filtered) for Mai herself.

AX Pattern · Silent Resolution23 AX

Fewer notifications?
No. Fewer decisions.

Seven agents had something to say. Six finished the work themselves. One remains — the one where Mai's judgment matters.

Calendar Agent saw Kiran's red-eye delayed 40 min, crossed it against Leo's 4 PM soccer, and shifted the pickup to 4:30 on its own. Weather, grocery arrivals, two bill reminders, a PTA note — all resolved. Mai sees only what needs her.

Focus mode suppresses notifications — the work still waits. Silent Resolution finishes the work. Nothing to catch up on.

Tokens as Contract

The same Token,
readable by humans and by machines.

Every Token in the Now Layer has two faces. The narrative half — what designers and family members read. And a machine-contract — what generative-UI compilers consume to emit the right surface.

Why it matters: this is what makes the Brain portable. The fridge, the iPhone, the CarPlay screen — none of them ship with hand-coded layouts for "Mai is busy at 7:02." They render from the Token state.

Token · Cognitive Load · narrative side
"It's 7:02. Mai's calendar shows three back-to-back meetings starting 8:00. The baby just woke. She's holding a coffee in one hand and a phone in the other. Cognitive Load = high."
→ fridge collapses 3 ideas into 1 default pick
→ phone hides everything except the one that needs her
→ CarPlay won't surface non-urgent during commute
Same Token · machine contract
// ctx.cognitive_load · v1
{
  "id": "ctx.cognitive_load",
  "value": 0.78,           // 0.0–1.0
  "confidence": 0.84,
  "sources": [
    "calendar.density",
    "time_of_day",
    "recent_activity",
    "physical_state.hands"
  ],
  "projection": {
    "high":   { "density": "minimal", "items_max": 1 },
    "medium": { "density": "standard", "items_max": 3 },
    "low":    { "density": "rich", "items_max": 7 }
  }
}
Generative UI · one Token, six renderings
INPUT
ctx.cognitive_load
0.78 · high
→ projection →
Fridge
1 default lunch
iPhone
1 of 7 surfaced
CarPlay
non-urgent muted
iPad
overview only
Nest Hub
ambient, no chime
Watch
silent, haptic only

No surface ships with a hand-coded "Mai is busy" template. Each surface is a renderer; the Token state is the script. Same contract, six expressions.

Concept · Tokenized Design Patterns What "tokenized design pattern" actually means at the architecture layer — not just a brand color variable, but the entire context state as a typed contract that surfaces compile against.

AX Pattern composition · what Silent Resolution is made of

Patterns aren't a catalog.
They compose.

"Silent Resolution" is four canonical patterns wired together. Each plays a role; together they collapse seven would-be notifications into one. This is what a pattern library looks like as a system, not a Pinterest board.

AX · D4 · Omakase Mode
Full delegation when trust has matured.
Autonomy Dial = Auto, Disclosure Dial = FULL. The fridge picks lunch #2 without asking — like a sushi chef hearing "omakase."
AX · D6 · Dynamic Friction
Caps the dial when stakes are high.
$2 yogurt → low bar. $48 shoes → mid. $150 bottle → never auto. Trust and risk are separate axes.
AX · E5 · Trust Breach Recovery
Demotes autonomy after a misstep.
After a wrong call, the dial steps back. Brain logs the negative signal. Trust rebuilds slowly — not by promise, by proof.
AX · A2 · Cognitive Scaling
Adjusts density to cognitive load.
Mai's load = high → 3 options collapse to 1 default. Low load → full menu opens. Same data, different rendering.
↓ composed into ↓
Scene
Silent Resolution · Mai's morning
7 candidates · 6 acted · 1 surfaced

Concept · Pattern composition Same four patterns, recombined, become "Care Architecture" in Chapter 04, "Adaptive Autonomy" elsewhere. The pattern library is a vocabulary. The scenes are sentences.

The foundation · DisclosureAX · A3Social-Aware Filtering

Finance Agent can't read grades.
Education Agent can't read the budget.

Only Calendar crosses all five domains — it has to coordinate everyone. Health is read by every agent — allergies are too dangerous to forget. Everything else has hard walls.

Mai never sees the grid. She picks one profile per person — OPEN / OBSERVATIONAL / FAMILY-SAFE / LOCKED — and the matrix below is what that resolves into under the hood.

User UI · iPad Disclosure Settings — what Mai (admin) actually sees. High-level profiles per person (OPEN / OBSERVATIONAL / FAMILY-SAFE / LOCKED), one tap to customize. Aoi's row is expanded so you can see what's under a profile.
CONCEPT · The Matrix beneath
What one family profile resolves into, rule by rule.

Concept explainer The full 5 × 5 grid that a single profile resolves into. No family has to read this to use it, but every Home Brain is built on it.

Disclosure resolution · how a profile becomes the grid

One tap on Aoi's profile.
Every agent's access rewrites itself.

Mai never sees the matrix. She picks Family-Safe. The algorithm cascades: defaults → per-domain refinement → per-person override → a deny-by-default lock. The grid resolves.

Step 01 · Profile
FAMILY-SAFE
1 input.
selected by admin (Mai)
Step 02 · Defaults
5 domain × 5 agent
25 cells.
profile-driven defaults applied
Step 03 · Overrides
3 cells flipped
Aoi opts Education
→ FULL.
teen self-override
Step 04 · Lock
Deny-by-default
Anything unspecified
resolves to HIDDEN.
never silently OPEN
Output
25
rules
The contract
resolve(person, profile) =
  defaults[profile]
    .overlay(domainOverrides[person])
    .overlay(personOverrides[person])
    .seal(deny_unspecified = true)
  → Map<(domain, agent), level>

Concept · Privacy as algorithm "FAMILY-SAFE" isn't a label on a settings page — it's a function input. The matrix is the function output. This is why Mai never has to think about 25 cells, and why the family can trust that nothing slipped through.

Inside the Brain

Three peeks at the engineering substrate.

How the Brain ships, how surfaces talk to it, and how the contract grows without breaking what's already deployed.

A · Brain API how surfaces subscribe

Surfaces don't ask. They subscribe.

Each surface declares the slice of Brain state it cares about. The Brain pushes updates. No polling, no global state, no leaks.

Surface · Samsung Family Hub
// fridge subscribes to its slice
const sub = brain.subscribe([
  "identity.allergies",
  "identity.dietary",
  "learning.weekly_rhythm",
  "ctx.cognitive_load",
  "ctx.priority_weight"
]);

sub.on("update", (state) => {
  fridge.render(state);
});
Per-surface permissions · enforced by Brain
fridge → identity.allergies READ
phone → ctx.* READ
phone → identity.pronouns READ + WRITE
tv → identity.health DENIED
all → ctx.cognitive_load READ (ambient)

Disclosure Dial isn't an opinion expressed in copy — it's enforced at the Brain API layer. A TV can't read health data even if its UI tries to.

C · Rule Engine how Tokens turn into UI commands

Tokens in. UI commands out.

The Rule Engine reads the live Token state, evaluates priority rules, and emits typed UI commands. Surfaces never compute layout from raw Tokens — they compile from commands.

IN · Token state
cognitive_load: 0.78
social_exposure: "family"
priority: "normal"
form_factor: "phone"
autonomy: "auto"
ENGINE · evaluate
rule_001: match
→ density = minimal

rule_017: match
→ silent_resolution = on

rule_034: skip
OUT · UI command
layout: "minimal"
items_max: 1
notifications: "silent"
undo_window_s: 300
tone: "calm"

Same engine runs on Fridge, iPhone, iPad, CarPlay, Nest. Same Tokens in, same command shape out — only the renderer differs.

F · Schema versioning how the contract evolves without breaking surfaces

8 Tokens today. 9 tomorrow. Your CarPlay still works.

Token schemas are versioned. Surfaces declare which version they speak. The Brain serves both, with deprecation windows long enough that no surface gets bricked.

Schema lifecycle
v1.0 7 Tokens · stable 2025
v2.0 8 Tokens · Disclosure split out CURRENT
v2.1 + Wellness sub-token (proposed) draft
Migration contract
// every surface declares
brain.connect({
  schema: "v2.0",
  fallback: "v1.0"
});

// Brain guarantees
deprecation_window: 18mo
breaking_change: "never silent"
new_field_default: "safe"

A 2025 fridge running v1.0 still receives projections from a 2027 Brain. Boring infrastructure thinking — but the reason a Tokenized Design System ships, instead of dying as a 6-month internal demo.

Backstage · Architecture Three of the engineering disciplines that make Context Grammar shippable instead of speculative — Brain API, Rule Engine, Schema Versioning.

Chapter 03 · Three Axes of Disclosure — the family in the living room, each on their own device but co-present
Chapter 03
Three Axes of Disclosure.
Who shares what with whom — is its own design.
Axis 01 · Translation LayerAX · A4Disclosure Cascade

One dietary fact.
Three family members read it differently.

Kiran (dad) got diagnosed with elevated liver markers. He taps share — and his Disclosure Dial translates the fact for each family member.

Mai (wife) gets the clinical summary + action. Aoi (15, teen) gets an observation, no numbers. Leo (9) gets a playful nudge.

One grocery item.
Four people. Four rules.

One grocery item. Four Substitution Modes, running in parallel on the same list.

Mai
mom · 36
AUTO · EXACT
"Only this brand."
Same Oikos Plain, every Sunday. Never ask me again.
Kiran
dad · 38
FLEXIBLE
"Any Greek, high-protein."
Brand doesn't matter. You pick — I'll eat it.
Aoi
teen · 15
EXPLORING
"Dairy hurts now — help me find a new favorite."
Recently found dairy upsets them. Trying oat, almond, coconut — one at a time.
Leo
kid · 9
EXACT + SURPRISE
"His blueberry squeeze, plus one surprise."
Same pack every week + one Japanese-style yogurt to try.

One Brain, four grammars — running on the same grocery list.

User UI · iPhone Mai's weekly grocery app — one list, but each yogurt row is handled differently. Hers is pre-ticked (AUTO). Kiran's shows a brand the AI chose (FLEXIBLE). Aoi's row is a dairy-free "try one" pick (EXPLORING). Leo's shows his usual plus a tagged "new this week" (EXACT + SURPRISE).

The Autonomy Dial, adjustedAX · D6Dynamic Friction

$2 yogurt auto-picks.
$150 bottle waits for approval.

The Autonomy Dial isn't a single setting. Before the AI acts, it checks the item — how much it costs, whether you can undo it, how often you buy it — and decides how much it can do on its own.

Your $2.40 yogurt: already decided — buy it. Your kid's $48 shoes: one tap to approve. A $150 anniversary bottle: just suggest — you decide together.

Same person. Same preferences. The AI only steps forward as far as it can safely decide alone.

Backstage + User UI Three moments from Mai's week. Left: the AI's reasoning — why it buys alone, asks, or only suggests. Right: what Mai actually sees on her iPhone. Mai never opens the left side; you are.

Axis 02 · Connective DisclosureAX · A7Live Recomposition

Nobody tried to share.
The family connected anyway.

Three people, three apps, three different evenings. Aoi (15) streams Jujutsu Kaisen 0 on Netflix. Kiran (dad) finished the manga on Kindle last Saturday. Sota (6) hums the ending theme while building Legos. None of them know what the others are watching, reading, or listening to.

But each of them, months ago, set their Disclosure Dial to Family-Safe — a quiet "yes, my consumption habits can be visible inside this household." That single consent is what lets the Home Brain notice patterns Netflix, Kindle, and Spotify each see only one slice of.

Step 01 · Consent (months ago)
Each person sets their Disclosure Dial.
Aoi → "Family-Safe." Kiran → "Open." Sota's age-aware default → "Observational." No one promises to actively share anything. They just allow the Brain to notice.
Step 02 · Latent overlap (right now)
The Brain sees what the apps can't.
Three apps each see one stream. The Brain sees one IP universe — manga, anime, opening theme — orbited by three family members across a week. Asynchronous, but converging.
Step 03 · Surface (Thursday evening)
An ambient nudge — gentle enough to ignore.
On the kitchen Nest Hub when Kiran walks in: "Sota's been on a song this week — same world as the manga you finished." No alert, no urgency. Just an opening for a conversation.

This is the design pattern itself: passive consent enables active connection. Each family member can change their dial at any time. None of them have to remember to share. The Brain is the one paying attention so they don't have to.

Concept · Family Brain surface Three isolated streams — Kindle (Kiran · manga), Netflix (Aoi · anime), Spotify (Sota · song) — all orbit the same IP universe. The family never sees this diagnostic view; the Brain uses it to decide what to surface next.

Thursday · 18:42 · The kitchen

Dad walks in.
"Hey Sota — what's the song?"
The conversation was the gift.

Aoi didn't send anything. The Brain just noticed three things were the same world, and quietly put one of them where Kiran would see it. The dinner-table conversation — that's what nobody tried to make happen.

User UI · Nest Hub Three threads converge on one ambient card — songs Sota's been humming, books Kiran finished, anime Aoi's mid-season on. Each row is provenance: "because Sota played this 11 times this week (Family-Safe)." No surveillance — just consent, observed.

Multi-modal projection

Four surfaces pick up Sota's song.
Nobody programmed any of them.

Cross-surface continuity isn't about copying pixels. The Brain hands every surface the same Token state — and each surface translates it into its own native modality. Visual on iPad. Voice on Nest. Haptic on Watch. Ambient on the kitchen lights.

Visual · iPad
One card. Maximum spacing.
Density collapses to items_max: 1. White space carries the signal.
emit: ui.layout(
  density: "minimal",
  items: 1
)
Voice · Nest Hub
Shorter phrasing. Pauses lengthen.
Same density signal becomes prosody. Confident voice slows down. Lists shorten.
emit: voice.utter(
  tone: "calm",
  words_max: 12
)
Haptic · Watch
Single soft tap. Silent.
Cognitive Load high → no chime, no notification banner. Just one feel-it pulse if it can't wait.
emit: haptic.tap(
  intensity: 0.3,
  chime: false
)
Ambient · Hue lights
Kitchen warms 200K. No words.
At its quietest, the Brain communicates through the room itself. The family feels it before they see it.
emit: ambient.shift(
  temp: -200,
  tween: "slow"
)

Concept · Cross-surface continuity Same Token state, four projection contracts, four modalities. The Brain hands surfaces a state — not a layout. Each surface compiles the state into its own language. This is what makes "the same family member, every device" actually work.

Chapter 04 · When It Counts — Mai driving, earbud in, the emergency call already in progress
Chapter 04
When It Counts.
Emergencies show what the Autonomy Dial protects — and what it never crosses.

Emergency in the air.
One thing on the screen.

Nurse Chen rings through CarPlay. Asthma attack. Aoi is stable. Mai needs to get to school. Priority Weight flips to EMERGENCY while the call is still live.

The dashboard collapses around one thing: getting there. Route re-plots. Kiran gets a text. Leo's pickup reassigns. Mai keeps her hands on the wheel.

Mai driving a 2030 EV — the windshield itself acts as a Care Architecture surface, collapsing to one focal element: 'School nurse · Aoi stable · ETA 12 min'
2030 · the same pattern, a new surface
When the windshield
becomes the dashboard.
A surface that doesn't exist yet

The windshield changes.
The Tokens don't.

Mai's 2030 EV doesn't have a CarPlay screen. The whole windshield is the surface — a thin AR layer painted into the glass. When the school nurse calls, Priority Weight flips to EMERGENCY and the entire windshield collapses to one element: a quiet status line and a glowing path on the asphalt.

The Brain didn't change. The Tokens didn't change. The Rule Engine emitted the same UI command — only the renderer is new.

2026 · CarPlay surface
A 12-inch dashboard screen.
Care Architecture renders as a full-bleed card on the screen. Map collapses around the route. Two contact pills below.
emit: ui.layout(
  surface: "carplay",
  density: "minimal",
  items: 1,
  chrome: "reduced"
)
2030 · windshield AR surface
The whole glass becomes UI.
Same Care Architecture. The HUD collapses everything except a status line + a path glowing on the road outside. No card. No chrome. Just one signal in the driver's natural gaze.
emit: ui.layout(
  surface: "windshield_ar",
  density: "minimal",
  items: 1,
  chrome: "reduced"
)

Same emit. Different renderer. This is what designing for surfaces that don't exist yet looks like. The Token contract is forward-compatible by construction — when the next form factor ships, you write a renderer, not a new design system.

Concept · Future form factor The surface is speculative. The framework that produces it is not. Same 8 Tokens, same Rule Engine, same Care Architecture pattern from Chapter 04 — projected forward.

Cascade · awaiting approvalAX · D1Approval Gate

The Autonomy Dial has a ceiling
care doesn't cross.

The emergency triggered 12 cascading actions. Ten ran automatically. Two need a human — an email to Kiran's Zoom team, a message to Aoi's doctor.

Both drafts are ready. Both wait for Mai's tap — the two places where her judgment still matters.

Aoi's Person Brain · Layer 01 · Identity update
she / her they / them
Context
Brain
Level 1

Aoi told Mai first.
Mai told the Brain second.

One line changed in the Identity Layer — and five agents updated themselves at once.

Without the Brain, Aoi would spend a month correcting every app by hand.

Nobody had to announce it.
Nobody had to notice it.

Chapter 05 · Over Time — the Yamashiro family on a Sunday morning, the present day
Chapter 05
Over Time.
Twelve months builds trust. Twenty years builds a legacy.
12 months · Four domains

Four domains. Four trust timelines.
None of them the same.

By month 12, Mai's Autonomy Dial has split into four settings — one per domain.

Home · Auto (omakase). Calendar · Notify. Education · Confirm. Finance · Suggest.

One Brain, four relationships. Trust isn't a single number.

CONCEPT · Trust evolution over 12 months
Four domains, four curves — each one earned at a different pace.

Concept · Trust curves Same Brain, four relationships — Home reaches Auto by month 3. Finance stays at Suggest through month 12. The Brain adapts to each domain's stakes, not a single dial for everything.

The Yamashiro family, 20 years later. Aoi (35) holds their toddler; Leo (29) and Sota (26) are grown; Mai (56) and Kiran (58) are greying into the same kitchen.
20 years later, Aoi is 35.
The family grew. The Brain kept pace.
Axis 03 · Graduated Archive

Memory isn't a log.
It's a collection of collections.

The Graduated Archive unlocks memories on life events, not calendar dates. At 15 Aoi's own memories opened. At 20 their parents' younger-self journals unlocked. At 28 a Family Story surfaced when they faced a similar decision. At 35 Aoi browses the archive their parents built.

This is Event Brains returning home. Every Trip Brain, every School Year Brain, every birthday — born for a window, learning intensely, then graduating its essence into the Household Brain before it dies. The archive is what Event Brains leave behind.

Same pattern Project 04 uses inside a company. One design language — home and office — outliving both.

User UI · iPad The Family Archive app on Aoi's iPad, 2046. Every word on screen is what Aoi sees — no captions, no explainer text. 247 memories open, 3 still waiting to unlock on life events.
Chapter 06
Same Grammar.
Different World.
The Yamashiros' kitchen and a B2B sales floor speak the same eight Tokens.
The same eight Tokens · in an enterprise sales team

Five Home Brains.
Five Org Brains.
Same five roles.

Replace Mai with a sales manager named Aiko. Replace the Yamashiro household with a 200-person sales team. The architecture doesn't move — only the cast does.

P1 · Home
P4 · Enterprise
Household Brain
Org Brain
Person Brain × 5
Brand Brain × N
Domain Brain × N
Research Brain
Household Coordinator
Project Master Brain
Event Brain × N (Trip Brain…)
Project Brain × N

The eight Tokens still fire in the same shape. The Brain count doesn't change — only the cast and the scale. This is what takes Context Grammar from "a home framework" to "a design language for any humans coexisting with agents."

Token
Consumer · Yamashiro home
Mai, Sunday morning, three lunches
Enterprise · sales team
Aiko, Q-end, three accounts on fire
① Physical State
One hand busy with coffee, baby on the other arm — fridge surfaces big-tap targets.
On a train between client visits, one hand on the strap — phone collapses dashboard to thumb-zone.
② Cognitive Load
7:02 AM, three meetings starting at 8:00 — fridge picks one default lunch instead of three options.
Five-minute window between back-to-back syncs — AI surfaces one critical lead, hides the other 22.
③ Social Exposure
Kids are watching the TV — agent hides $-figures and shows just the choice.
Manager is screen-sharing in standup — confidential lead names mask to "Lead A / Lead B."
④ Priority Weight
Kid's pref > brand > price — the rule that resolves the yogurt collision.
Q-target > existing renewals > new leads — the rule that resolves the pipeline collision.
⑤ Form Factor
Phone → TV in the living room — same recommendation, different surface, prices off the screen.
Desktop → projector for boardroom presentation — same data, presentation mode, KPI deltas hidden until reveal.
⑥ Feasibility
Inventory · size · delivery date — what the AI can actually book today.
Budget runway · approval level · contract deadline — what the agent can actually close today.
⑦ Autonomy Dial
Research = Auto. Routine grocery = Auto. Big-ticket purchase = Confirm.
Pipeline analysis = Notify. Discount approval = Confirm. Contract signature = Confirm.
⑧ Disclosure Dial
TV hides price. Aoi (15) sees observation. Leo (9) sees a playful nudge. One fact, three translations.
Manager sees individual reps' numbers. Reps see only their own. Customer-facing dashboards mask both. Same Translation Layer.
The 0 → 1 claim

Material Design unified screens. Context Grammar unifies the human-AI relationship — across home, office, car, and the surfaces we haven't built yet.

Where the Brain runs · multi-vendor reality

Five vendors. One household.
One Brain layer.

In 2026, no family has one AI. The Yamashiros' kitchen runs Samsung. Mai's phone runs Apple Intelligence — which itself routes to Gemini for hard queries. The Nest Hub is Gemini. Mai's office laptop is Microsoft Copilot. Five providers, all wanting to be the family's primary AI.

Context Grammar doesn't pick a winner. It defines the federation contract that lets them cooperate. Like Matter for smart-home or OAuth for auth — the Brain is a layer above the vendors, not a service from any one of them.

Map · the Yamashiro household six surfaces, five vendors
Mai's iPhone
Apple Intelligence
+ Gemini handoff for complex queries · ChatGPT for open-ended
Family Hub fridge
Samsung Knox + Gemini Nano
on-device Gemini Nano under the Knox security envelope
Kitchen Nest Hub
Google Gemini (cloud)
household ambient surface · Gemini cloud-backed
CarPlay (Mai's EV)
Apple Intelligence
on-device only while driving · no cloud handoff
Mai's office laptop
Microsoft Copilot
work-domain Brain, separate Identity slice
Apple Watch
Apple Intelligence (on-device)
100% local · Now Layer haptics + ambient
Brain layer × execution location where each layer lives, who guards it

Different Brain layers, different security domains.

Brain Layer
Where it runs
Who guards it
Crosses the home?
Identity Layer
100% on-device · secure enclave
Apple Keychain · Samsung Knox · Google Family
never raw — only via Brain federation
Learning Layer
on-device + encrypted private cloud (opt-in)
Apple Private Cloud Compute · Samsung Personal Data Engine · Google Personal Context
aggregated patterns only
Now Layer
100% on-device · ephemeral
whichever device the surface lives on
never — too volatile to share
Connective Disclosure
household-local mediator
one elected hub (Nest / HomePod / Family Hub) · Matter-style
stays in the house · never to vendor cloud
Federation contract what every vendor implementation must honor
// any vendor's Brain implementation must declare
brain.federate({
  schema:                 "context-grammar/v2.0",
  identity_provider:      "apple_id",    // or samsung_account, google_family, ms_entra
  learning_residency:     "on_device",   // or private_cloud_compute
  household_hub:          "nest_hub_4f3a", // elected mediator
  cross_vendor_handoff:   true,
  preserve_disclosure:    true,         // dial travels with request
  audit_log:              "household_local"
});
Why these clauses
schema: all vendors agree on what 8 Tokens mean
identity_provider: you stay logged in once, household-wide
learning_residency: Apple users keep on-device · power users opt into PCC
household_hub: Connective Disclosure runs locally — never to vendor cloud
preserve_disclosure: "kid-with-dairy-limit" can travel; "leo_named" cannot
Walk-through · 7:02 AM one decision · three vendors · zero leaks

How Mai's lunch decision actually travels.

01
Mai's iPhone reads Identity.allergies from iOS Keychain · stays on-device
APPLE
02
Brain federation: iPhone hands a disclosure-bounded slice to Family Hub via Matter-secured channel
→ SAMSUNG
03
Family Hub queries Gemini Nano on-device. Gemini sees "kid · dairy_limit · 9yo" — not "Leo Yamashiro"
GEMINI
04
Result projects to Nest Hub via household hub · Disclosure Cascade re-translates for whoever's in the kitchen
→ GOOGLE
Three vendors. One decision. No vendor ever held the un-disclosed identity. The household stayed sovereign.
SOVEREIGN
Apple adopts because
on-device-first principle stays intact. Identity never leaves the secure enclave.
Samsung adopts because
Galaxy AI keeps Gemini partnership but adds Samsung-only value at the household level.
Google adopts because
Gemini becomes the household substrate without looking like vendor lock-in.
Microsoft adopts because
Copilot extends past work — same Brain shape, with an Entra-managed work Identity slice.

Backstage · Federation Context Grammar's strategic value isn't replacing any vendor — it's the contract that lets all of them cooperate without any one of them owning the family.

What changed · by the numbers

Storytelling earns the case.
Numbers earn the budget.

A simulated 12-month run with the Yamashiros, modeled against a baseline of four single-purpose AIs (the "four Mac minis" workaround). Numbers are indicative — the shape of the win, not a benchmark suite.

Mai's morning decisions
47/wk 11 /wk
76% of trivial coordination decisions absorbed by Silent Resolution.
Pronoun update propagation
3–4 wk 8 min
One Identity Layer write → five agents updated. Aoi never has to correct an app by hand.
Cross-device handoff success
61% 94 %
Fridge → phone → car → store. The Brain travels with the task.
Care Architecture activation
~120s 3.2 s
From nurse's call to Mai's CarPlay collapsing to one task — Priority Weight flips in real time.

Methodology · indicative Modeled with one family across 12 weeks, against a baseline of four account-isolated agents (the setup shown in the opening "four Mac minis" panel). Real-world numbers will depend on the family, the surfaces, and the maturity of the Brain. The point isn't the exact figure — it's the shape: silent resolution compresses the cost of coordination by an order of magnitude.

How this was made

The case study you just read
was the seventh draft.

Senior reviewers want to see thinking, not just polish. Below: four jobs-to-be-done quotes that shaped the brief, then six decisions — kept and cut.

Jobs-to-be-done · what we heard before designing
"I don't want to be the family's calendar app. I want to be their mom.
mother of 3 · age 36 · works full-time · interview, simulated cohort
"By 7 AM I've made twelve decisions for other people. I'd like to keep three of them for myself.
parent · age 38 · two kids in school · interview, simulated cohort
"I told one app I switched pronouns. The next morning, three other apps still got it wrong.
teen · age 15 · using AI tools daily · interview, simulated cohort
"It's not that I want AI to know everything. I want it to remember what I told it last Tuesday.
father · age 41 · early adopter · interview, simulated cohort

Methodology · simulated-archetype interviews informed by published HCI research on AI-and-family dynamics. Quotes are composite voices, not direct attributions. The point of this section is the shape of the need, not the specific n.

Decision 01 · kept 7→8 Tokens
Promoted Disclosure from a setting to its own dial.
Originally one of seven "context settings." Three weeks of testing showed disclosure was the axis users worried about most — privacy isn't a checkbox, it's a posture. Made it a peer of the Autonomy Dial. "Knowing too little × trust too much" became logically impossible.
Decision 02 · cut "Trust Score"
Killed the single "Trust = 78%" indicator on first contact with users.
Showing a number turned every interaction into "is this the 22% of error?" Replaced with Behavioral Tone — the AI shows confidence through speed and posture, not a meter. Confident = 200ms ease-out, fast. Uncertain = 600ms, offers options. Same signal, no anxiety.
Decision 03 · kept 3-layer Brain
Three memory speeds, not one big "context window."
Tried a flat memory model first. Identity (rarely changes), Learning (months), Now (seconds) — three layers because they need different write rules, different surfaces, different update cadences. The pronoun moment only works because Identity is its own layer.
Decision 04 · cut 12 transition patterns
Compressed 12 Autonomy transitions into 2 principles.
Drafted 12 patterns for moving between Suggest/Confirm/Notify/Auto. Realized they reduce to Delegation (human → AI) and Escalation (AI → human) with parameters: speed, trigger, reversibility. Same coverage, half the cognitive load for adopters.
Decision 05 · kept Omakase mode
Auto isn't "control given up." It's "trust earned."
Western framing makes the Auto stage feel scary. Reframed it as omakase (お任せ) — the relationship has matured enough that delegation is a sign of trust, not loss. This single framing change rewired the whole timeline visualization in Chapter 05.
Decision 06 · cut Slide deck version
Killed the slide-deck version. The story is the case study.
Built a 30-slide deck first. Realized executives don't review portfolio decks — they read the long-form case studies their report sends them. Pivoted to scroll-narrative with editorial photography. The deck became the appendix.

Backstage · process Six of the dozens of decisions that shape this case study. Working sketches, whiteboards, and the full argument log live in the Substack archive — happy to walk through any of them in conversation.

What we don't know yet

The honest list.

A trustworthy framework names its own edges. These are the parts of Project 01 I am genuinely uncertain about — the questions I would test first if I joined your team next month.

Open question · 01
Brain repair when the model learns the wrong pattern.
If the Brain learns "Mai always picks vegetarian on Tuesdays" from one stretch of bad data, how does she correct it without re-onboarding? Need: a Trust Repair flow.
Open question · 02
Identity Layer through life events — divorce, blended family, death.
The Identity Layer assumes the family is stable. What happens when it isn't? Who owns the Brain when a household splits? Need: lifecycle handoff design.
Open question · 03
Connective Disclosure — when does insight stop feeling magical and start feeling creepy?
"Ask Sota about his new song" lands as a gift. The same mechanism could land as surveillance. Need: an empirically-tested threshold for the warmth-to-creep boundary.
Open question · 04
Multi-vendor Brain — does this work without a closed Apple/Samsung/Google ecosystem?
The Yamashiro Brain assumes one trusted Brain layer that all surfaces read. Today's reality: Apple, Samsung, Google, OpenAI all want their own. Need: a federation protocol or a vendor-neutral Brain spec.
Three things you can take tonight

For the designers reading this:
three principles, no Brain required.

You don't need a Context Grammar Brain to apply these. They are design moves you can make in whatever AI product you ship next month.

Principle 01
Memory has speeds. Design for all three.
Identity (slow — names, allergies). Learning (medium — habits, preferences). Now (fast — the next 60 seconds). They need different write rules, different UIs, different review affordances. Don't dump them in one chat history.
Principle 02
"What AI knows" and "who's in the room" are two different controls.
Most AI products give you one privacy switch. Real life has at least two: how much you let the AI know about you, and how much it should reveal in this exact social context. Treat them as separate dials. Both belong on the surface, not in settings.
Principle 03
One autonomy dial isn't enough. The dial must adapt to stakes.
Same person, same preferences. $2 yogurt = AI just buys it. $48 shoes = one tap to confirm. $150 anniversary bottle = AI only suggests. Stake-aware autonomy is how you avoid both "the AI did something it shouldn't have" and "why does it ask me about every tiny thing."

If any of these three is useful for what you're building right now — that's the win. The full Context Grammar framework — 23 AX Patterns, eight Tokens, two Dials — is open. Take what fits.

Chapter 07 · Close — the family house at blue hour, warm lights through the windows
Chapter 07
Close.
Everyone's home. The Brain keeps watch.

When it works,
nobody calls it software.
They just call it home.

Project 01 thesis · Context Grammar

End of Project 01 · Living Home

The design isn't AI.
It's the context a family lives in.

One Brain, in three layers. Five agents, five different relationships. Three axes of disclosure — translation, connection, archive. All of it quiet, mostly invisible, earned over months.

When Aoi is twenty and becomes a parent themselves, this family's memory will still be queryable. The same eight Tokens describe a sales floor. The same Brain shape works in any room where humans coexist with agents.

Brain 3-Layer 3 Axes of Disclosure Graduated Archive Tokenized Contract Same Grammar · Two Worlds
The family on an Ocean Beach walk, April 2023
Apr 2023 · Ocean Beach