Every fund measures factor exposure. Nothing measures causal exposure.

Sozu maps what sits upstream of your positions, grades each causal link, and shows what the market has not priced.

Causal engine Illustrative
How causal exposure is computed Instruments feed a graded causal graph, which propagates only to positions with a traced path. EVENT MARKETS OSINT & TELEGRAM FLIGHT & VESSEL TRACKING REGULATORY FILINGS OPTIONS & ON-CHAIN FLOW GRADE AND PROPAGATE PRICED · OBSERVED · ASSERTED PRICED OBSERVED NO PATH Semiconductor assembler 1 HOP · 4 MIN LEAD · HELD 79% European industrial 3 HOPS · 3 WK LEAD · HELD 61% US regional bank NO TRACED PATH · NOT SCORED Your book
Instruments feed the graph. Every edge is graded by what can verify it, and only paths that reach something you hold are surfaced. The third position has no traced path, and that is reported rather than filled in.
Who it is for

Pick the seat you sit in.

Built for investment teams. The engine is the same across the four seats below. What lands on your screen, and what the chain ends in, is not.

Finance

    Typical shape
    What it does

    Trace the path, grade every edge, check the evidence, run it forward.

    An edge is one causal connection: this moves, therefore that moves. Most systems in this space will produce a confident chain of them about anything you ask. Here the grade decides whether an edge counts at all.

    01

    The graph

    Every node with a traced path to what you hold, out to the hop where the trail goes cold.

    02

    Edge grading

    Priced, observed, or asserted. Edges nothing can verify score zero rather than being dressed up.

    PRICED OBSERVED ASSERTED · 0
    03

    Evidence ledger

    A dated record of every source behind a claim, so later conclusions trace back to exactly what was known and when.

    04

    Conditional queries

    Run a root forward before it resolves. Which of your names it reaches, and how long you would have.

    Products

    Four ways to use the same causal record.

    Read what changed each morning, interrogate a chain directly, connect through the API, or inspect the published method.

    Delivery

    Morning read

    A daily read of what changed upstream of your book and why, organized by the instruments you actually hold — not by internal graph identifiers.

    DailyPer seat
    Delivery

    Ask the chain

    Interrogate any specific causal chain directly — how it formed, what confirmed it, what would invalidate it — rather than reading a static output.

    ConversationalPer seat
    For technical teams

    API and webhooks

    For teams that want to pull the same record into their own systems directly, rather than through the workspace.

    API and webhookConfigurable
    Research

    Published method

    Inspect the definitions, grading rules, forward record, and failures behind the product.

    Open researchPublic

    What it is not

    Not a black box. Sozu states a probability on every root, and states where that number comes from — a debiased market price, or the graph's own evidence-driven estimate where no market exists. What's still being built out in the open is the calibration record proving the estimate adds value beyond the market alone. Where a chain is short and already fully priced, it says there is no lead time to sell rather than manufacturing one.

    The proof

    A claim only matters if it can be checked later.

    Sozu makes causal claims before outcomes resolve, tests them on unseen periods, and publishes the resulting record.

    The claim

    A named cause, a consequence, and a time window. Specific enough to be supported, challenged, or proven wrong.

    The test

    Learn from an earlier period, then test on a later period the system has not seen. The bar is whether adding Sozu improves on the market alone.

    The record

    Claims and test dates are published before resolution. Wins, misses, and corrections remain visible rather than being rewritten afterwards.

    The problem, stated properly

    What actually goes wrong on a finance desk.

    Not a data problem. The work is split across terminals, research, spreadsheets, memos, and chat, leaving no maintained record of the chain.

    Funds · portfolio managers
    Six drivers, forty names.

    You know what moves each position. That is 240 things to watch, most outside the instrument you hold and several in someone else’s coverage. You watch the loudest eight and hope the ranking was right.

    Market makers · traders
    The flow arrives before the reason does.

    Something is being lifted and nothing on the tape explains it. By the time the chain is legible you have already been adversely selected, and the post-hoc explanation arrives after you have widened.

    Private equity · private credit
    You hold it for five years.

    The chain that impairs the asset starts somewhere nobody is monitoring, moves over quarters rather than days, and reaches a covenant before it reaches anyone’s screen. Diligence mapped it once, at entry.

    Banks · research
    The note is only worth writing before the revision.

    Coverage is deep and narrow by design. The chain that moves your names runs through policy and industrial territory owned by other desks, and there is no meeting where it gets handed across.

    The graph

    Worked graphs, as they actually run.

    Hard truths on the left, then what has happened, then the branch points where somebody chooses, then what each choice does, then the combinations, then the outcomes that resolve. Click any node to trace its chain and read the full reasoning.

    Detail
    Click a node

    Every node carries its own reasoning. Outcomes carry either a live market price or a stated model read, and the two are never shown as the same thing.

    Hard truth Confirmed event Option at a branch Direct effect Combined path Outcome
    Narrative engine

    Find it early. Know whether it is real. Trace what it reaches.

    A story starts somewhere: a regional forum, a trade publication, a cluster of accounts that all posted within the same hour. The engine surfaces it while it is small, classifies whether it grew or was built, and follows it through the graph to whatever it touches.

    First seen

    Select a narrative

    Organic or built
    Where it is running
    Which channels
    What it reaches

    Conditional queries

    The question correlation cannot answer.

    Not what is likely to happen. What happens to your book if it does. Only a traced structure can tell you which positions sit downstream of a root that has not resolved yet, how long you would have, and how much of that path is scored rather than assumed.

    If this resolvesIllustrative

    Select a scenario

    PositionPathDirectionLeadScored

    The engine

    Three views turn information into a decision.

    First establish the facts. Then trace what they can affect. Finally compare that consequence with what the market already reflects.

    What do we know?

    Facts before interpretation

    A dated record of what happened, where it came from, and when it became knowable. Later conclusions can always be traced back to the information available at the time.

    What can it change?

    Consequences, not correlation

    A map from an event to the positions it can reach, the reason each connection should hold, and the time available before the consequence arrives.

    What is the market missing?

    The gap that can matter

    A comparison between the traced consequence and current market expectations. If the market already reflects the path, there is nothing new to act on.

    Your week

    What actually lands on your screen.

    Not a feed of everything that moved. Only the hops that reached something you hold, with the path and the grade attached. Most of what the engine sees never reaches you.

    Week of 14 September · illustrative0 surfaced · 0 suppressed
    Mon
    Licence docket entry2 names · 2 d lead · priced
    Freight fixture driftwatchlist · observed
    Tue
    Options anomaly, unconfirmedbelow promote threshold · expired
    Wed
    Memory contract print1 name · 3 hr lead · priced
    Edge demoted to observedregister update
    Thu
    Fri
    Rate path repriced1 name · 90 sec · no edge
    Weekly failure log3 edges published
    Thursday is empty on purpose. A system that finds something every day is not filtering, and the suppressed count is published next to the surfaced one for the same reason.
    Objections

    The questions we get in the first ten minutes.

    Is this just an LLM generating a plausible causal story?

    An LLM will produce a confident chain about anything you ask it, and nothing in the output tells you which edges are real. The difference is the grade. An edge only carries confidence if something downstream can verify it, and every edge carries its firing history. Where nothing can verify a path, the readout says asserted and it contributes zero. A system that never says asserted is the one to worry about.

    Does this replace our risk system?

    No. Factor exposure and causal exposure are different objects answering different questions. Your risk system decomposes what you own along axes that already exist. This decomposes what sits upstream of it. Most teams run them next to each other.

    Do our positions and our own edges stay private?

    Yes. The substrate graph is shared. Anything built on top of it, including your positions, your own edges and your own scoring, stays isolated and never enters the shared graph. Every propagation is written with the time it occurred and the information it saw, and prior states are never overwritten, so any past state can be reconstructed for audit.

    Book a call

    Bring a position. We’ll trace what sits upstream.