Research papers
Three programmes on evidence, accounts and the institutions that work with AI.
Our available work includes the Agentic Stemmatics working paper and its companion research note, Collation and the Microscope. Agentic Representation and Agentic Institutions are forthcoming. The descriptions of these forthcoming papers announce the questions being developed; they do not report completed studies.
Agentic Stemmatics
Collation, Provenance, and Conservation for an Emerging Machine Textual Tradition
Available · Working paper · Not peer-reviewed
How should we assess agreement between AI models when their sources and errors may be related? This paper draws on philology to investigate textual transmission and model relationships. It distinguishes shared error from evidence of descent, proposes tests and records the limits and failures of its own approach.
Read the paper →
Read the introduction · Zenodo record
The website may carry a later revision than the deposited version. The paper’s front matter records both.
Collation and the Microscope
How agentic stemmatics and interpretability could check each other
Available · Research note · Methodological proposal
Can comparisons of model outputs and measurements of internal features help test each other? This companion to Agentic Stemmatics proposes ways to investigate that question and sets out conditions a valid test must meet. It reports no experimental results.
Read the research note →
Zenodo record
The website carries version 1.1; the deposited text is version 1.0.
Agentic Representation
Evidence, Selection and Warrant in AI-Assisted Reports
Forthcoming · Draft in development
A report can contain accurate quotations and still encourage a conclusion its sources do not support. This paper will examine how evidence is selected and connected, how corrections change the standing of earlier claims, and how to distinguish misleading omission from reasonable compression.
The proposed evaluation compares an explicit record of those relationships with competent general review and a simpler checklist. Whether the additional record improves understanding or correction remains an open question.
Agentic Institutions
Knowledge, memory and correction across successive agents
Forthcoming · Research in preparation
What allows an organisation to learn when its AI participants change from one session to the next? This paper will investigate how rules, records, retrieval and human judgment affect the survival of useful learning. It will distinguish preserving a justified correction from merely repeating an inherited claim.
The programme draws on distributed cognition and organisational memory. Its proposed comparisons ask whether particular institutional arrangements improve correction beyond simpler ways of retaining and checking information.
For the wider purpose connecting these programmes, read the research introduction. For an accessible entry into the available work, see our introduction papers.