CounterProof Research

Building knowledge with AI

Agentic Epistemics: studying evidence, judgement and lasting learning.

An AI model writes a piece of software and explains why it works. Another agrees. What gives us good reason to rely on the software? If a later test exposes a mistake, how do we make sure the next person or agent learns from it?

CounterProof Research asks: how can people and AI build knowledge and useful systems that remain open to evidence and correction?

Agentic Epistemics is our programme in applied and social epistemology for work with AI agents. It investigates how people, models, tools and institutional practices can support justified decisions and durable learning.

Our first practical field is software engineering: deciding what deserves to be built, examining designs, writing and reviewing code, and learning from failures. Human responsibility runs through that work, including decisions about what evidence is sufficient and when to proceed.

From criticism to construction

CounterProof examines claims and work through adversarial review. PreProof is the constructive side of our research: developing ways to carry lessons from review into the decisions and checks that precede and accompany a build. Sometimes the right decision is to simplify, defer or stop.

A faithful report can describe an unsound design. We therefore distinguish accurate reporting, evidence about how software behaves, and reasons to believe it serves its intended purpose. Better records and more reviewers do not, by themselves, establish better engineering. Whether our proposed methods improve outcomes at an acceptable cost remains a question for testing.

Three research strands within Agentic Epistemics

Agentic Stemmatics

Where did a claim come from, and what does agreement between models tell us?

Stemmatics studies how texts descend from earlier texts. Drawing on this philological tradition, our programme examines transmission and relationships among AI-generated texts and models. It asks when similar answers provide useful corroboration, when they may reflect shared sources or errors, and what evidence would justify an inference of descent.

Working paper available

Agentic Representation

What does an account lead its reader to believe?

A report may cite its sources accurately while leaving out a correction that changes the conclusion. This proposed programme examines how selection, emphasis and the relationships between evidence shape an account, while allowing that several different accounts may be reasonable.

Conceptual preprint · Web v0.4 · Updated 4 November 2026 · Not peer reviewed

Read the paper · Introduction · Deposited v0.3 on Zenodo

Agentic Institutions

How can useful learning survive the next agent, session or model?

This proposed programme studies the rules, libraries, tools and human decisions through which an organisation works with AI. It asks whether those arrangements carry justified corrections into later work, expose inherited mistakes and keep established practice open to challenge.

Conceptual preprint · Web v0.4 · Updated 4 November 2026 · Not peer reviewed

Read the paper · Introduction · Deposited v0.3 on Zenodo

First application: Correction Across Clinical Handovers, a conceptual methods preprint on how a hospital could test whether an AI-assisted correction workflow keeps justified corrections useful across handovers (v0.15, deposited 5 November 2026; the proposed studies have not been conducted).

Companion to that paper, and the programme's second bounded step into clinical work: Agentic Adjudication of Clinical Reports, a conceptual methods preprint asking whether an additional, separately obtained read adds anything worthwhile to the strongest single-model workflow on de-identified, already-signed radiology report text, and how one would know (v0.6, deposited 6 November 2026; no study has been conducted).

The programmes share questions and materials, but each must establish its own value. A finding about textual transmission would not, by itself, show that a report is adequate or that an institution learns well.

Through PreProof, we aim to carry useful findings from this research into software design and construction.

Agentic Settlements and Payments describes a practical outcome towards which this research programme works: a federated settlement layer beneath machine-payment protocols. It is a conceptual preprint that offers the design as an engineering hypothesis, not a plan of record (draft v0.19, deposited 11 November 2026; no system described in it operates and no study proposed in it has been run).

Health, healthy ageing and longevity

We are developing a health research programme focused on healthy ageing and longevity. Its first steps are a source library, critical examination of research claims and tests of whether AI-assisted workflows preserve evidence and corrections. This is research-method development; clinical collaboration and evaluation remain essential future steps.

How we approach the work

This programme builds on existing research in the epistemology of AI. Human and machine authorship do not, by themselves, establish whether a claim is true. The source, the evidence and the conditions of production still matter when deciding how much weight to give it.

We draw on philology and historical source criticism, the philosophy of history and science, and research on AI, distributed cognition and organisational memory. These traditions help us examine evidence and frame questions. Claims that a particular method improves results require further evidence.

Our approach is piecemeal: identify a concrete problem, propose a bounded change, compare it with a simpler alternative, and examine the result and its cost. We distinguish completing a procedure from achieving its purpose. A method must be open to revision or abandonment when the evidence goes against it.

A library can preserve an error as faithfully as a correction. An institution learns when what it discovers changes what it does. Independence and freedom to criticise matter here, including the freedom to challenge our own questions, tests and preferred explanations.

Start reading

Start with Agentic Epistemics: An Introduction: a plain-language guide to knowledge, evidence and understanding, and how our research strands connect.

For a shorter practical companion, read When does an AI answer become knowledge?: an essay on what turns a plausible answer into a claim we have grounds to rely on, and how AI might help knowledge grow.

Introduction papers offer an accessible way into the questions. Research papers develop the arguments, proposals and limits. Our notes examine particular questions and cases; our essays explore the wider ideas behind the work.

Why do we discuss music in research about AI? Why music belongs in our research connects inherited sources, interpretation and the survival of corrections to practical software decisions. Musical scholarship helps us explain and challenge our questions; engineering evidence must establish the value of our answers.

The essay Different writers, shared traditions connects these concerns through oral tradition, writing and institutions that work with AI. Earned, not inherited sets out the importance we attach to authenticity, independence and criticism.


CounterProof Research is the research activity of CounterProof, a commercial adversarial-review practice of Clavestra Capital Limited (Malta, C 113987). We have a financial interest in methods that could benefit that practice. Our papers state their evidence, limitations and review status. The studies proposed in Agentic Representation and Agentic Institutions have not yet been run.