The open build and research behind Yu.

Geode Labs is building Yu: a private place a person fills with the pieces of their own thinking, and an AI assistant that helps them use what they have kept. This page is the working record of that build. It changes as the evidence changes, and it says plainly what is still unproven.

Experimental version timeline

Each build answered one question, then exposed the next.

Yu has developed by repeatedly finding the uncertainty that prevents the next important question from being answered.

V4.5Live · Current
Assistant first (current direction)
TRL
6
Working estimate. This stage is in build, not validated.
What we're testing now
If a person has a working assistant from the first minute, does the value land clearly enough that they keep using Yu while it builds underneath them?
What happened
The launch happened, quietly, in September 2026 after being publicly described as coming. The founder's own read afterwards was that the product was roughly 70% there: a founder assessment, not a measured completion score. This was not a story of the launch failing. Feedback was improving. People explicitly said the educational hook was resonating, the personalization idea was beginning to make sense, and the product looked great. What their behavior showed was different: people could admire the product and understand pieces of it without understanding what they were supposed to do with it.
What blocked the test
Blocker
The product was not fully landing
Response
People could see the pieces, self-awareness work, AI education, personalization, the robot, the city, without consistently seeing what the product was or what it did for them today.
Blocker
People did not know where to type
Response
The clearest symptom, almost absurd in hindsight: people repeatedly did not understand where the text or input was supposed to go. The product was asking users to understand an architecture before giving them the most familiar AI interaction model there is.
Current build response
- Assistant first
  A working assistant, Geo, from the first session.
- Yu underneath
  The private context the assistant works from, built from what the person approves.
- 365 progression
  Unlocks depth over time. It does not gate basic usefulness.
- Context Coins
  Spendable rewards for the work a person does.
- Store
  Where Coins are spent.
- University and AI learning
  Learning how to use AI well, alongside the build.
- User control
  The person approves, edits, and can remove what Yu holds.
What we learned / what this established
A product conversation on Sept 21, 2026 connected those pieces into one read: it needs to be chat. The assistant should be the obvious surface, and Yu should build underneath the conversation rather than requiring the user to understand Yu before they can use the assistant. Both V4 and this pivot are big. V4 proved the integrated world could exist; V4.5 changes the entry model around the assistant. If the implementation becomes a substantial chat-first rebuild, it may ultimately land as V5 rather than being treated as a small patch. Parts of V4 may be abandoned in the process, and that is not wasted work: V4 exposed the next constraint.
Always be testing something, but do not let the test freeze the product around the thing you are testing. A test is there to remove uncertainty, not to become a constraint you protect after the evidence changes. Build enough to learn, and keep moving when the critical path moves.
Next question / what this still has to prove
That assistant-first makes the value clearer and keeps people using Yu. It is the current hypothesis, not a validated result.
V4
Integration sprint
TRL
5 → 6
Working estimate.
What we were testing
Can we remove the upstream comprehension, motivation, progression, and integration blockers so the methodology can actually be tested?
The two tracks
- Experience track
  Understand the value → want to start → stay engaged → reach the frameworks.
- Method track
  Strengthen the underlying methodology through research and expert and psychologist review, so what people reach is defensible.
What we learned / what this established
Onboarding, profile/reflection, the 21-step first build, progression, AI 101, the city, weekly delivery, AI handoff, and assistant arrival operated as one end-to-end system rather than separate product pieces. At that point the plan was beta testing the following week, and the product was publicly described as launching.
What V4 assumed
That the assistant should arrive after the 21-step build, as the thing the person earns.
V3Test blocked
First framework MVP
TRL
4 → 5
What we were testing
Does the methodology work when users go through Yu themselves?
What happened
People were not reliably getting far enough through the experience to answer that question properly.
What blocked the test
Blocker
Perceived chore
Response
Gamification, progression, city, rewards, visible accumulation.
Blocker
Value / concepts not landing
Response
AI 101, clearer conceptual scaffolding, stronger explanation of why the work matters.
What we learned / what this established
The methodology test was blocked by an upstream comprehension and motivation problem. That blocker became the new critical path.
V2
Make the idea legible
TRL
~3
TRL barely moved, and this stage still mattered.
What we were testing
Can people understand what Yu is and why it matters?
What existed
The marketing site, positioning work, explanation of the six applied areas, and the language and visual system for explaining Yu.
What we learned / what this established
Explaining the mechanism and value was itself a product problem.
V1
Manual Yu
TRL
2 → 3
What we were testing
Can we actually do cognitive translation?
What existed
Paisley + Danny were effectively the system. They used a framework generator, ran the conversations, interpreted what came back, built their own system instructions, and created individualized scaffolding for other people.
What we learned / what this established
Enough manual proof of mechanism to justify productizing it.

Assessed Sept 9, 2026

Working readiness assessment

A dated internal working assessment from the Yu journey-map evidence review. These are not external certifications and no outside body awarded them.

Technology Readiness Level

Working assessment

TRL 6

Technology demonstrated in a relevant environment.

The assessment placed Yu at TRL 6 because the full pipeline runs end to end against real external users on a production-shaped stack. It deliberately did not award TRL 7, because the system was not yet at or near operational scale in the target environment.

Caution: The missing held-out accuracy test is the biggest technical uncertainty that could change this assessment.

Innovation Readiness Level

Working assessment

IRL 4

Prototype a low-fidelity minimum viable product.

The assessment placed Yu at IRL 4 because user activation and willingness-to-pay evidence were still early while Yu is free to use. That is the current evidence boundary and the next thing the company is designed to learn.

2-level readiness gap

Product build has advanced faster than paid-market evidence. The next experiments are designed to close the evidence gap.

What moves the levels?

TRL: Held-out accuracy and predictive-validity evidence, plus an operational-scale environment.
IRL: Paid use, activation, retention, and willingness-to-pay evidence.

These are evidence thresholds, not promises.

Readiness of the evidence

What we can stand behind, and what we cannot.

Supported

Backed by the current architecture, the shipped build, or established outside research.

Person-specific context can matter to AI personalization.
Yu has a working behavior-informed assessment and translation pipeline.
The product preserves source and provenance, and the person can correct what it holds.

Testing

An open hypothesis with a test designed for it.

Behaviorally derived instructions reduce correction burden compared with generic AI or self-written custom instructions.
Translated directives transfer better across tasks.
Conditional instructions reduce misfires.

Unknown

An important question we cannot answer yet.

Which framework categories add the most incremental value.
How stable different findings are over time.
How well the benefits transfer across model providers.

SUPPORTED means supported by the current architecture, the shipped implementation, or established external research. It does not mean Yu has been scientifically validated. The framework categories are Geode Labs' working applied taxonomy for organising the method, not a validated model of the mind.

Build log

What changed, and when.

Sept 22, 2026
V4.5 assistant-first pivot
The assistant is now the product from the first minute and helps the person build Yu, with 365 progression, Context Coins, the Store, and AI learning adding depth instead of gating baseline usefulness. Geo is the single canonical starter assistant, replacing the nine-model roster in the active product. This is the current hypothesis and it has not been validated. If the implementation becomes a substantial chat-first rebuild, the result may ultimately be V5 rather than a patch.
Sept 21, 2026
Post-launch read and the crystallizing conversation
After the quiet launch, the founder's own read was that the product was roughly 70% there. People explicitly said the education resonated, the personalization was starting to make sense, and it looked great, yet their behavior showed they still did not understand the interaction model, including where their input was supposed to go. A product conversation on Sept 21 connected that evidence into one read: it needs to be chat, with the assistant as the surface and Yu building underneath. V4 was not wasted work; it proved the integrated world could exist and exposed the next constraint.
Sept 9, 2026
V4 integration sprint
Onboarding, profile/reflection, the 21-step first build, progression, AI 101, the city, weekly delivery, AI handoff, and assistant arrival ran as one end-to-end system rather than separate product pieces. At that point beta testing was planned for the following week.
Sept 2026
City simplification
The cognitive city was simplified so every place has one clear job and the map reads at a glance.
Sept 2026
Conditional instruction and misfire design
Instructions gained conditions: they know when they apply and stay quiet when they do not.
Sept 2026
Public methodology and taxonomy research scaffold
The working paper and this build record opened the method up to outside scrutiny.

Methodology

The working paper.

The method is written up so it can be read and argued with, not just described in marketing copy.

Cognitive Translation for Personalized AI

A Taxonomy and Methodology for Modeling Person-Specific Context

Working paper · under active validation

Public PDF coming soon