Projects and results

Concrete work, with its current scope.

Explore the experiments, software and mathematical questions within ORBIT. Each item has its own purpose and evidence: a working package, a prototype and an open conjecture are not interchangeable status labels.

AI behaviour

CARF: testing candidate early-warning measurements

Can observations from a multi-turn interaction warn of a developing failure early enough to change its outcome?

Research framework

Conformal Agentic Risk Framework

CARF specifies how candidate trajectory measurements would be tested and used. Controlled synthetic work is internally reported; prospective warning and operational deployment are not established.

Worked experimentCurrent evidence and limitations

Design explorer

Change the conditions in an illustrative experiment and inspect the planned comparisons. It does not run an agent or produce measured results.

Explore the design

Statistics

Statistical conclusions that keep their assumptions visible

Fixed-sample intervals and sequential evidence answer different questions. The sampling design and the chosen method determine the guarantee.

Public software

ExactCIs

Design-aware confidence intervals for 2 × 2 tables. Exact conditional, Mid-P, score and asymptotic constructions are explicitly distinguished.

Methods, examples and source code
Source record on request

eprocesses

Sequential-evidence work within the Inference direction. A sequential validity claim requires its own construction and assumptions; a fixed-sample interval alone does not provide it. A release record is not currently linked here.

Ask for the current evidence record
The wider statistical programme

Mathematics

Erdős Basecamp and problem 364

Exact computation can expose structure. A proof must still establish a precise statement under its stated hypotheses.

Research programme

Erdős Basecamp

Mathematical discovery through representations, counterexample searches and proof engineering.

Programme and worked question

Shared capability

Formal methods check specified claims

A mathematical proof, a statistical algorithm and a release-policy check have different objects and different limits.

Public source; evaluation record on request

tre-airlock

A formal-control demonstrator for declared release policies in a trusted research environment. It is not a number-theory result or evidence that an integrated CARF operational system is deployed. The public repository documents its fixed predicate and implementation boundary; an evaluation record beyond that narrow demonstrator is not linked here.

View the public source repositoryAsk for the evaluation record

Programme

Follow the question or inspect the sources

The three directions connect through a research hypothesis, not a claim that one result establishes the others.