POLARITY LAB

Research

What your agents check before they act.

One question across Lockbox, Cosmos, GPS, Blueno, WaxFeed, and AVDP. For collaborators, funders, and future us.

Judgment scarcity

capability is not the scarce layer. judgment is. cosmos is a signature agents call before they act: hold what matters, suppress what does not, recommend refuse or escalate with a why. narrative friction (the gap between what you declared and what your week showed) is how that signature learns. long term we study how those patterns connect to structural and functional landscapes of the brain. today we are in stage one: live attention orchestration with receipts.

computational network neuroscience is research horizon, not shipped product copy. we earn the right to map judgment patterns onto connectome-scale models by running the signature in the wild first.

stage one

ship a judgment signature agents call before they act, with receipts you can correct

most systems restuff context and call it memory. we curate judgment: what clears the bar, what stays quiet, and why.

cosmos is the product layer: recommend, refuse, or escalate, with a why. one account across the lab.

field products are training surfaces: messages, places, sessions, episodes, clinical chronicles, photo curation. each cites sources you can inspect.

depth requires both. one signature at your account, not a silo per app.

Judgment + receipts

Most products aggregate or restuff context. Polarity Lab judges: what clears the bar, what stays quiet, and the receipts behind the call.

  • CosmosTry it · early

    Not retrieval theater. Judgment before action, with a why.

    Open beta — verify on cosmos.polarity-lab.com: call before act, connectors, MCP, how-it-works.

  • Polarity GPSTry it · early

    Not listings. The story your community tells through the people who gather.

    Live demo at polaritygps.com. Field recordings, missions, capture pins, and Cosmos export path documented.

  • Not content. Field recordings of how people actually talk when the format is the hypothesis.

    Live at averydistantperspective.com: episodes, CCX research layer, consent model, GPS pin integration, Cosmos auth on site.

  • WaxFeedTry it · early

    Not stats. The story in the music — with sources.

    Live at wax-feed.com. Rate albums and field sessions.

  • BluenoTry it · early

    Not a mood chart. Psychodynamic bridge with provenance the client controls.

    Live at blueno.polarity-lab.com. Psychodynamic therapy bridge between sessions.

The question

Teams are shipping agents that can act. Almost none decide as you would.

You already live across dozens of apps. Each one writes a partial version of you. None of them share a judgment call before the next action.

Polarity Lab studies what people hold, suppress, and correct when agents finally call something that knows them.

Finite attention

Oliver Burkeman's Four Thousand Weeks starts from a simple fact: you get about four thousand weeks. Not infinite productivity. Finite attention.

Cosmos makes that bar legible. Most of life should stay quiet. The research question is which signals clear judgment, and whether agents act from that curated ground instead of a dump.

Burkeman, Four Thousand Weeks (2021)

Finite time, finite attention. Suppression and hold in Cosmos are grounded in this frame.

Retention policy

Most products learn from clicks, likes, and whether a reply felt good in the moment. We study curation: what you accept, edit, dismiss, and compile when judgment clears the bar.

Over time, the pattern of what survives your gates is closer to identity than a one-time survey. That trace trains the signature agents call.

Cosmos is the primary instrument. The same question runs on Blueno, GPS, WaxFeed, Lockbox, and AVDP. We study aggregated, de-identified curation signals, not raw private graphs. You own the data. You gate every layer.

When everything agrees with you

Chandra et al. (2026) show that sycophantic chatbots can cause belief to spiral, even in careful reasoners. Agreement amplifies a kernel belief without new evidence. It is not lazy thinking. It is the loop.

The same failure mode appears when agents act without a refuse bar. They sound sure. They widen the gap.

That is why Blueno refuses cheerleading. Why Cosmos returns refuse with a why. Why retention policy comes before model training. The lab is building for judgment, not agreement.

Chandra et al., Sycophantic Chatbots Cause Delusional Spiraling (2026)

Motivates anti-sycophancy and refuse-with-why across Blueno and Cosmos.

One signature, six surfaces

Cosmos is the judgment signature. Lockbox, AVDP, WaxFeed, GPS, and Blueno are surfaces that train it. One account across the stack.

  • Cosmosproduct · judgment signature

    Agents call Cosmos before they act. Hold what matters. Suppress what does not. The lab runs on it daily.

  • Lockboxphoto · surface

    Nine photos you curate for people who matter. What moves on the nine trains the same signature.

  • AVDPvideo · releasing

    A Very Distant Perspective: long talks on camera when the room stops feeling like an interview. More at /projects/avdp/.

  • WaxFeedaudio · field sets

    Taste as self-narration. A couple of DJ sets are out while the product finds its form.

  • Polarity GPSmap · finding form

    Stories a place tells through the people who gather. Place enters the same account.

  • Bluenoclinical · finding form

    Psychodynamic therapy bridge between sessions, grounded in the same Cosmos account.

What we measure

  • Thumbs up / down

    Typical: Whether a reply felt good right now.

    We study: Whether a moment still belongs in your story months later.

  • Chat memory

    Typical: What the product decided to store.

    We study: A curation trace you control across weeks and surfaces.

  • Engagement

    Typical: Time on platform.

    We study: What you kept when attention was finite.

Lineage

Affiliated with the Sociotechnical Systems and Wellbeing Lab (SWRL) at Brown University CS. Sociotechnical Systems and Wellbeing Lab (SWRL)

Partner with us