AI for the critical analysis of arguments
judg·ment noun
Middle English jugement, borrowed from Anglo-French, from juger “to judge” + -ment.
Merriam-Webster Dictionary
AI has made writing faster. But convincing prose doesn’t guarantee well-founded conclusions.
Epistemic Engine examines the arguments in a scientific, legal or business document, whether a person or an AI drafted it. It connects claims and evidence, makes assumptions visible and points out possible weaknesses, every step traceable.
It helps professionals, institutions and companies review what they read, improve what they write and ground what they decide.
Generative models have made plausible text cheap to produce. More and more papers, theses and reports are partly drafted by a machine and signed by a person. What hasn’t multiplied is the time of the people with the expert judgment to validate them.
| Domain | What is happening | Source | How Epistemic Engine responds |
|---|---|---|---|
| Science | 47% more articles were indexed in 2022 than in 2016, and expert attention has not grown at the same pace. | Hanson et al., Quantitative Science Studies, 2024 | A structured pre-review, so experts spend their attention where it matters. |
| Scientific publishing | A major publisher retracted more than 11,300 papers from a single portfolio in two years, after an integrity scandal. | The Register, 2024 | The same structured standard for thousands of submissions: what each one supports and where it fails. |
| Automated review | Automated reviewers reward style: having an LLM rewrite a paper raises its score by 0.45 points without better science. | Baumann et al., ICML, 2026 | Style doesn’t count as evidence. Every observation is a structured object, with traceability down to the sentence. |
| University | 45% of Spanish undergraduates use AI to write or correct their assignments. | Fundación CYD, 2025 | It doesn’t try to guess whether an AI wrote it: it examines whether the argument holds, and shows the supervisor between drafts. |
| Law | In three years, courts have had to deal with more than 1,400 cases of AI errors, mostly invented citations. | Scientific American, 2026 | It separates what the brief claims from what its sources allow it to claim, and flags whatever has no identifiable source. |
To apply judgment, you first have to dissect the argument.
Epistemic Engine analyses the text, be it a scientific paper, a legal opinion or an investment report, and breaks it down into a chain of epistemic objects and their relations. Every object is traceable: it is anchored to the exact passage it comes from.
Choose an example. Compare the text with its critical reading, and trace each observation back to its source passage.
Four examples from different disciplines
Scope. The thesis depends on E and S, and neither supports it as written. Repair. Data broken down by group, an explicit normative premise and a reply to the bias objection.
The editorial team decides
Scope. The thesis rests on S, which is exactly what case law disputes. Repair. Argue the disproportion and justify applying the doctrine between businesses, or change the legal basis.
The signing lawyer decides
Scope. The recommendation rests on an unsourced figure and an undeclared assumption. Repair. Cite the source of the 30%, state the share assumption and add a downside scenario.
The board decides
Scope. The promotional claim says more than the trial allows. Repair. Limit it to the subgroup, report the absolute reduction and drop the general recommendation.
The medical-legal committee decides
Illustrative examples written for this page. A real report is longer, and every observation links to its passage and its source. Try it on a document of your own.
Epistemic Engine represents each document’s epistemic chain as a persistent graph that can be queried, compared and aggregated. Across a whole collection, be it a scientific journal, an archive of legal opinions or a committee’s reports, it answers collection-level questions: which concepts run through it, which sources are cited or where authors disagree, with a conversational agent to explore them.
Judgment also rests on a standard, and every field has its own. A paper in theoretical physics can’t be assessed against the same epistemological criteria as one in economics. That is why Epistemic Engine lets you build a rich semantic layer specific to each field or discipline: the profiles. Language models act as inference engines inside this architecture.
The whole platform runs through five questions.
It reconstructs argumentative units: premises, evidence, assumptions, warrant and conclusion, each with its address in the text.
It applies the field’s profile: what counts as evidence and what warrant each kind of claim requires. A legal opinion doesn’t argue like a clinical trial.
It looks for objections by family and records coverage: what was examined, what didn’t apply and what remained undetermined.
A possible objection is not a demonstrated defect, and taking a side on an open question is not an error.
It separates measurement from the paradigm and from policy: the field decides which standard applies; the institution decides what to do with the result.
Criteria belong to the field and the institution
Every evaluation combines three layers. The engine is the common core. The profiles are the semantic layer specific to the field or discipline and to the institution. The profiles guide how the engine behaves.
Why not just ask an LLM?
The difference lies in architecture and approach. Epistemic Engine is not an AI that reviews a paper: generated prose is not the source of truth, and critical attention isn’t left to the model’s implicit judgment. On its own, a general-purpose LLM doesn’t provide verifiable traceability, governed critical coverage or a semantic layer with explicit disciplinary standards and institutional criteria.
The platform was designed for the most demanding needs. Scientific journals and postgraduate programmes are our first area of application, because that is where analysing argumentation is most complex and expert judgment is under most strain: editors, reviewers and supervisors read more and more texts written with AI help. Epistemic Engine helps journals pre-review every submission and postgraduate programmes support every thesis draft, without handing over judgment.
Pre-review of every submission
Before external peer review, a structured reading of every manuscript: what supports the thesis, where it fails and which claim has no backing. The editor decides with more information, and reviewers spend their attention on what is at stake.
Manuscripts don’t train models, and files are stored in the European Union.
Master’s and doctoral theses, between drafts
A progress report for the supervisor, who reviews it and discusses it with the student: what the work claims, what supports it and what to strengthen in the next draft, including when it was written with AI help.
It doesn’t grade: assessment stays with the supervisor and the examining board.
Epistemic Engine can also be applied to other domains where a poorly grounded conclusion costs money or reputation, or creates legal liability.
Editorial integrity at scale
The same structured standard for thousands of submissions: what each one supports, where it fails and which claim has no backing.
Law firms and legal departments
Briefs, opinions and reports: which statute or ruling supports each claim, which leap goes from the case to the rule and which contrary line is left unaddressed.
Boards, committees and professional services
Reports, investment memos and AI-assisted deliverables that someone signs: where the recommendation holds and where it rests on an assumption.
Pharma, finance, energy and government
Promotional, regulatory and policy claims, traced to the evidence that supports them and flagged when they outrun it.
Agents that draft reasoning
A verification layer for what agents produce: structure, provenance and assessments validated by expert judgment.
We start with a calibration trial: each domain is calibrated with its own experts before going into production. The field and institutional profiles change (§ 2); the engine stays the same.
A tool that serves expert judgment has to show where it agrees with it, and say where it doesn’t. Epistemic Engine uses a calibration methodology to measure scientifically the level of agreement between the experts of the institution or company and the engine’s verdicts. Calibration also makes it possible to fine-tune the field and institutional profiles.
Before starting, we fix what is measured, on what scale and what will count as success. During the round, neither those conditions nor the system version change.
Your expertsThey assess the documents without seeing the system’s reading.
Epistemic EngineIt reads the same documents with your field’s and your institution’s profiles.
Among the experts themselves, and between the experts and the system, discounting the agreement that chance alone would produce.
Every disagreement is documented and analysed with your team. If needed, the profiles are adjusted.
A new round, with the adjusted profiles
To find out how far it agrees with your experts, ask for a pilot.
We don’t ask you to trust the system: we propose measuring it on your documents and against your experts’ judgment, before you decide.
Which documents your team reviews, which decisions depend on them and where the time goes. That tells us whether a trial makes sense.
We pick a real workflow and a sample of its documents, whether in a journal, a master’s or a doctoral programme, or in a law firm or a committee. It can be done with documents that have already been reviewed, so measuring carries no risk. We agree in advance on what would count as success.
With fixed conditions and pre-registered metrics: how far they agree, and how long each review takes.
What works moves into supervised production. What doesn’t is documented and fixed.
Where to start
Choose what you’d like to do. Whichever you pick, a person will get back to you.
Received.
One last step