Research

the bet

You already compile. Your brain takes partial signal and produces coherence. Other people do the same about you. Those compilations feel true from the inside. They are also skewed by perspective, selective memory, and mood. You rarely understand why you assembled the story the way you did. You almost never see the assembly rules behind how someone else reads you.

Cosmos adds a third compile: an external one across your connectors, with provenance attached. The hope is that an AI compile is less captive to today's mood or last week's argument. It is not pure objectivity. Foundation models carry training data and bias. Long term we want a Cosmos-trained model on anonymized, aggregated trails from people who use the product, so the compiler improves on the population it serves. For now, the baseline we care about is still a step forward: your compile, their compile, and one you can audit.

Each day cosmos compiles signal into a few items on your today screen. Tap any item and you see the receipts. You endorse what is true. You reject what is not. The question: when a story about your life comes back with sources, do you still recognize yourself in it?

what today does

Connectors pull fragments you already produce. Today compiles them into a frog, a few supports, and a count of what can wait. Refresh runs when new signal lands. Corrections are data: what you finished, what you marked wrong, what you said was not important.

Add iMessage alone and you get one compile. Add Gmail and the story may deepen or contradict. Add calendar and a gap you were not naming can surface. Three compiles of one life can finally sit next to each other: what you think is true about you, what someone else thinks is true about you, and what Cosmos compiles from the fragments with receipts shown.

why network readouts

Provenance answers which fragments backed a claim. Network readouts answer how the compile weighted your trail: which cognitive modes dominated (planning, reflection, execution, emotional processing) and whether that pattern fits the story you tell.

Neuroscience named these modes as functional networks in the brain. Cosmos borrows that vocabulary for behavioral signal. It is another inspectable layer when a moment says one thing and the trail says another. Network readouts support the main question. They do not replace it.

what we are learning

The study we want to run on pilot cohorts compares three compiles of the same person:

1. Self-compile. What you believe is true about you right now.
2. Other-compile. What someone who knows you believes is true about you.
3. Cosmos-compile. What the receipt-backed model produces from your connectors.

Then each party ranks which compilation feels most accurate, excluding their own. That blinded comparison is the core experiment. Around it we track: how often you correct Cosmos, what breaks when connectors stack, and whether an endorsed compile changes your self-story afterward.

We are early. Text reply is the correction interface today. Agree/disagree on moments is coming. Cohort exports are available under agreement for collaborators who want to design this with us.

where this goes

Inspectable narrative. Today, questions about how someone is really doing often reduce to conversation: they talk off the top of their head, you guess whether to believe them. If personal narrative becomes inspectable, self-knowledge stops being pure introspection. You get receipts for inner life for the first time: not a diagnosis, but evidence-backed compiles you can check, dispute, and refine.

Your confirmed depth. Most tools already have a perspective on you. They tune it for their goals and never show you the assembly. Cosmos inverts that. You tailor the compile until it matches how you want to be seen. That confirmed model becomes portable: agents, scheduling, music, education, and B2B surfaces that read you as you want to be read, not as one app misread you.

How we build differently. Products that touch human state could ground in provenance instead of vibes. Research on attention and identity could study compiles and corrections, not only surveys. Clinical and educational tools get a path toward evidence-backed self-report without replacing human judgment. The long bet: make the stories we live inside legible enough to argue with, and good enough to share on purpose.

Lab-wide thesis: polarity-lab.com/research · project depth: polarity-lab.com/projects/cosmos

explore

collaborate

Email [email protected] with your question and cohort needs. If you have an invite, sign in and use Thread. That is the study environment.