Research

We publish little. Only what survives.

A line opens when a problem has appeared in three independent settings. It closes when we have an answer that holds outside the case that produced it — or when we establish that it does not.

01 / Position

What we do and do not research.

Scope

We do not research models. Laboratories with budgets of another order do that, and do it better. We research the space between the model and the company: which architecture survives contact with a real process, what breaks as volume grows, where automation stops paying.

There is very little useful literature on that ground because almost everything published comes from whoever sells the tool. Our advantage is access to systems in production inside real companies, with their dirty data and their exceptions.

Everything is published anonymised, and negative conclusions are published too. An experiment showing that something does not work saves more money than three success stories.

02 / Lines

Currently open.

  1. L-01
    Active

    Adaptive orchestration of multi-agent systems

    How to choose the execution topology automatically from measurable properties of the request instead of leaving it to the operator. Cost, latency and quality compared across modes.

  2. L-02
    Active

    Verifying generative output in regulated domains

    Architectures where the model proposes and a deterministic engine validates against explicit rules. The open question is not how to validate, but how much of the decision can be delegated before traceability becomes theatre.

  3. L-03
    Active

    Domain knowledge locked in documentation

    Building knowledge graphs from real document corpora, and how much structure can be inferred without an expert validating it.

  4. L-04
    Active

    Prediction over fragmented business data

    Reconciling sources that were never designed to be joined, and which predictions survive an imperfect join.

  5. L-05
    Open

    The limit of the agentic organisation

    How much of a company can operate without continuous human intervention, and which functions stop making sense when automated even though they technically can be. Deliberately uncomfortable: we do not assume the answer is "everything".

03 / Publications

Nothing published yet.

Status

The first reports are in preparation and will appear when the material has enough production history for the conclusions not to be provisional. We would rather publish late than publish a hypothesis dressed as a result. When they exist, they will be here in the open.

In preparation
  • Execution modes in multi-agent systems: cost, latency and quality compared
  • Anonymising real cases for demonstration without losing technical fidelity
  • What breaks when a document assistant moves from pilot to full corpus