Research

Open questions,
built into a system.

Nexefiy is infrastructure for studying and operating persistent, governed, evolving intelligence. These are the questions it is designed around — how it currently answers them, and what remains open.

R01
Persistent objectives
What is the right unit of autonomous work, if not the task or the agent?
Current approach
A versioned objective genome that outlives the agents and strategies serving it, with success conditions judged against evidence.
Open problem
How objectives should be decomposed, merged and retired as organizations change.
See it in the system : Persistent objectives
R02
Grounded world models
How should intelligence represent the part of reality an objective depends on?
Current approach
Typed entities and relationships; facts with provenance, confidence and freshness; contradictions kept, not overwritten.
Open problem
Keeping a world model current and trustworthy when sources disagree or go stale.
See it in the system : Grounded world models
R03
Epistemic honesty
How can a system say exactly how it knows what it claims?
Current approach
Observed, estimated, simulated and predicted values as distinct types; confidence capped below certainty; an explicit statement when evidence is insufficient.
Open problem
Calibrating confidence over time, and communicating uncertainty so people act on it well.
See it in the system : Epistemic honesty
R04
Governed autonomy
How much authority should autonomous intelligence hold, and who decides?
Current approach
Layered boundaries — Constitution, policies, permissions, budgets, approvals — enforced before action, with human-only decisions enforced in the database.
Open problem
Granting authority in proportion to demonstrated reliability, and withdrawing it safely.
See it in the system : Governed autonomy
R05
Intelligence compilation
Given an objective, what intelligence should exist to pursue it?
Current approach
A deterministic compiler that decides the architecture — including when the right answer is a single entity, a person, or a refusal.
Open problem
Learning better architectural decisions from outcomes without losing reproducibility.
See it in the system : Intelligence compilation
R06
Controlled evolution
How can a system improve itself without drifting outside its boundaries?
Current approach
Versioned strategies and entities, compared on the same evaluation, promoted through governance, reversible by lineage.
Open problem
Detecting when an improvement on the measured objective is a regression on what was not measured.
See it in the system : Controlled evolution
R07
Structured disagreement
What should happen when intelligent entities disagree?
Current approach
Challenge and verification as protocol acts; disagreements recorded with claim, evidence and resolution; escalation to people.
Open problem
Using preserved dissent to improve later decisions rather than merely documenting it.
See it in the system : Structured disagreement
R08
Unknown unknowns
How can a system notice what it is not measuring?
Current approach
Systematic questions about untested assumptions, missing information, ignored dependencies and changed conditions.
Open problem
Prioritising which unknowns deserve investigation before they matter.
See it in the system : Unknown unknowns

Method

How we hold ourselves to account

The flagship demonstration is a reproducible, end-to-end run of one objective through the whole loop — every step recorded, including the options that were rejected.

See the flagship demonstration

Claims follow evidence. We do not describe capabilities the system does not have.
Simulations are experiments, not forecasts. Results are labelled with how they are known.
Negative results are results. Failures are recorded and shared with the system that produced them.
Reproducibility first: decision logic is deterministic so a decision can be re-derived and audited.

Research

Open questions,
built into a system.

R01
Persistent objectives
What is the right unit of autonomous work, if not the task or the agent?
Current approach
A versioned objective genome that outlives the agents and strategies serving it, with success conditions judged against evidence.
Open problem
How objectives should be decomposed, merged and retired as organizations change.
See it in the system : Persistent objectives
R02
Grounded world models
How should intelligence represent the part of reality an objective depends on?
Current approach
Typed entities and relationships; facts with provenance, confidence and freshness; contradictions kept, not overwritten.
Open problem
Keeping a world model current and trustworthy when sources disagree or go stale.
See it in the system : Grounded world models
R03
Epistemic honesty
How can a system say exactly how it knows what it claims?
Current approach
Observed, estimated, simulated and predicted values as distinct types; confidence capped below certainty; an explicit statement when evidence is insufficient.
Open problem
Calibrating confidence over time, and communicating uncertainty so people act on it well.
See it in the system : Epistemic honesty
R04
Governed autonomy
How much authority should autonomous intelligence hold, and who decides?
Current approach
Layered boundaries — Constitution, policies, permissions, budgets, approvals — enforced before action, with human-only decisions enforced in the database.
Open problem
Granting authority in proportion to demonstrated reliability, and withdrawing it safely.
See it in the system : Governed autonomy
R05
Intelligence compilation
Given an objective, what intelligence should exist to pursue it?
Current approach
A deterministic compiler that decides the architecture — including when the right answer is a single entity, a person, or a refusal.
Open problem
Learning better architectural decisions from outcomes without losing reproducibility.
See it in the system : Intelligence compilation
R06
Controlled evolution
How can a system improve itself without drifting outside its boundaries?
Current approach
Versioned strategies and entities, compared on the same evaluation, promoted through governance, reversible by lineage.
Open problem
Detecting when an improvement on the measured objective is a regression on what was not measured.
See it in the system : Controlled evolution
R07
Structured disagreement
What should happen when intelligent entities disagree?
Current approach
Challenge and verification as protocol acts; disagreements recorded with claim, evidence and resolution; escalation to people.
Open problem
Using preserved dissent to improve later decisions rather than merely documenting it.
See it in the system : Structured disagreement
R08
Unknown unknowns
How can a system notice what it is not measuring?
Current approach
Systematic questions about untested assumptions, missing information, ignored dependencies and changed conditions.
Open problem
Prioritising which unknowns deserve investigation before they matter.
See it in the system : Unknown unknowns

Method

How we hold ourselves to account

The flagship demonstration is a reproducible, end-to-end run of one objective through the whole loop — every step recorded, including the options that were rejected.

See the flagship demonstration

Claims follow evidence. We do not describe capabilities the system does not have.
Simulations are experiments, not forecasts. Results are labelled with how they are known.
Negative results are results. Failures are recorded and shared with the system that produced them.
Reproducibility first: decision logic is deterministic so a decision can be re-derived and audited.

Open questions,built into a system.

Persistent objectives

Grounded world models

Epistemic honesty

Governed autonomy

Intelligence compilation

Controlled evolution

Structured disagreement

Unknown unknowns

How we hold ourselves to account

Open questions,built into a system.

Persistent objectives

Grounded world models

Epistemic honesty

Governed autonomy

Intelligence compilation

Controlled evolution

Structured disagreement

Unknown unknowns

How we hold ourselves to account

Open questions,
built into a system.

Open questions,
built into a system.