Inversiq Research
The infrastructure required when AI takes part in institutional decisions.
Inversiq Research studies what becomes necessary when AI performs part of the work behind consequential institutional decisions: how those decisions are computed, governed, recorded and later reconstructed.
A research programme inside a technical company. We publish Research Notes: working theses, stated so they can be argued with, rather than reviewed findings.
Read the research thesisResearch thesis
AI can do more of the work behind a decision. It does not settle what the decision was.
AI systems can now read the documents, extract the figures, summarise the arguments and draft the analysis behind a consequential institutional decision. That capability is real, and it is improving.
It does not, on its own, answer the questions an institution has to answer afterwards. Those questions are not about model quality. A more capable model does not make them easier to answer; it makes them more frequent, because more of the work that produces an answer happens without a person in the room.
What a model’s output does not answer
- What information was authoritative?
- Which assumptions were accepted, and by whom?
- Which calculation produced the official result?
- Which policy applied?
- Who held the authority to approve?
- What changed during review?
- Can the decision be reconstructed later?
Decision Infrastructure is our name for the software layer these questions have in common: the evidence, computation, rules, review, authority and record behind a decision held in one governed path, rather than distributed across documents, spreadsheets, email and meetings.
That is a category thesis we are arguing for, not an established academic field and not a demonstrated result. It may turn out that these questions are better answered inside existing systems of record, or by governance practice rather than software. This programme exists to work out which.
Research notes
Short notes on questions we are working on.
Each note states a working thesis and the reasoning behind it. None is peer reviewed, and none reports empirical results of our own.
Research Note 001
Auditability and the Decision Record
An activity log records what happened inside software. It does not explain why a decision was authorised. This note sets out what an AI-mediated institutional decision would need to leave behind in order to remain defensible and reconstructable, and proposes a working set of components for such a record.
11 September 2026 · 6 min read · Research Note
Read the noteResearch Note 002
Why Deterministic Computation Matters in AI-Mediated Decisions
Language models are powerful tools for interpretation, but probabilistic model output is a poor foundation for authoritative, reproducible computation. This note argues for a control boundary in which models interpret, propose and explain while versioned deterministic software produces every material figure, and examines where that boundary should sit.
11 September 2026 · 6 min read · Research Note
Read the noteResearch agenda
Four open questions.
This is a research agenda, not a list of results. Two of these questions have produced a position worth defending, published as Research Notes and labelled as working theses. The other two remain open.
Decision Records & Auditability
How should an AI-mediated institutional decision be recorded so that it remains explainable, reviewable and reconstructable?
Research Note 001Deterministic Computation
Which parts of AI-assisted decision-making require deterministic execution rather than probabilistic model output?
Research Note 002Authority & Governed Agents
How should AI agents operate when humans, organisational policy and explicit authority remain accountable for material decisions?
Open question
Decision Infrastructure as System Architecture
Which primitives remain reusable when the domain logic of a decision changes completely, and what would show that they do not?
Open question
Status and claim discipline
How we label what we say.
Research claims and product claims fail the same way: by borrowing certainty from the wrong category. Everything Inversiq Research publishes is one of four things, and we name which.
01
Research question
An open question we are working on. No position is claimed, and the answer is allowed to be inconvenient for the architecture we are building.
02
Working hypothesis
A position we hold and argue for, including Decision Infrastructure itself. Reasoned and published, not demonstrated.
03
Implemented in software
Exists in code and is verified by our own tests. Implemented is not the same as proven, and we do not use the two words interchangeably.
04
Proven in live use
Demonstrated on real decisions, in production, with real institutional data. Nothing carries this label today.
Inversiq Research is a research programme inside a technical company, not an independent institute. Our notes are not peer reviewed, carry no DOI, and are not produced in collaboration with an academic institution. They are working theses, published so that they can be argued with.
Inversiq’s product development provides a practical context in which parts of this agenda may eventually be tested, beginning with institutional Commercial Real Estate workflows. That testing has not happened yet, and until it has, nothing on this page should be read as evidence from live use.
Research is a working process.
These notes set out hypotheses we expect to test, challenge and revise as the architecture meets real institutional workflows. Disagreement is useful to us.
Working on related questions? Get in touch.