Skip to content
OxARCA

Archos

An epistemically grounded LLM system for archival exploration

Digitisation has produced archives no individual scholar can read. Archos helps historians work through them without giving up source criticism — every claim traceable to a specific document, every ambiguity left standing.

How the system works Get involved

What one research question produced

primary sources synthesised
261
sentence-level citations
1,688
words of source-grounded prose
85,789
computation cost
$12.70

From a single research question the system assembled 261 primary sources into 5 sections and 27 subsections. Whole-run compute: approximately $12.70.

Our workflow

The Archos pipeline. Steps, left to right: a historian writes the research statement; the system retrieves archival evidence, applies an epistemic constitution and produces source-grounded synthesis; a historian verifies the result. The first and last steps are marked as domain expert involvement, the three middle steps as scholarly AI.

Tap the diagram to open it full size. Figure 1. The division of labour across a run. The interpretive decisions at either end stay with the historian; the scholarly AI carries the evidentiary work in the middle, and every claim it produces remains traceable back to a retrieved document.

The problem

Large-scale digitisation of archives has created vast machine-readable archives that no individual scholar can realistically analyse. This ‘archival deluge’ presents a methodological challenge:

How can historians engage with an ever-expanding evidential record while maintaining the principles of historical source criticism?

Conventional AI systems are designed to optimise fluent responses rather than historically defensible claims. They often obscure provenance, flatten ambiguity, reconcile contradictory evidence without justification, and provide little transparency regarding how conclusions were reached.

Our central claim

AI systems cannot be reliably — and therefore safely — adopted in the humanities unless the epistemic requirements of scholarship are encoded directly into the system architecture.

So we propose a Archos — scholarly AI system : a source-grounded architecture that translates those requirements into an Epistemic Constitution, and then into retrieval and synthesis themselves.

Rather than optimising for persuasive answers, Archos is designed to produce transparent, inspectable and revisable scholarship. Historians remain in control of interpretation while the system assists with evidence discovery, organisation and documentation.

Read how the system works →

Two ways to answer a historical question

Conventional AI systems compared with the OxARCA scholarly AI system
Conventional AI systems Our scholarly AI system
Optimises for plausible answers Optimises for historical rigour
Hides uncertainty Preserves uncertainty
Opaque provenance Sentence-level traceability
Stateless Cumulative research record

Our Epistemic Constitution

Eight requirements of scholarly practice, encoded into retrieval and synthesis rather than left to the model's discretion.

  1. Traceability Every claim cites a specific source.
  2. Bounded agency Synthesis only from retrieved evidence.
  3. Language fidelity Source language reproduced verbatim.
  4. Ambiguity preserved Never silently resolved.
  5. Contradiction shown Conflicting sources kept side by side.
  6. Accumulation Every independent attestation stays visible.
  7. Contextual continuity Argument context carried forward.
  8. Revision & versioning A cumulative, revisable research record.

News

Collaborate with us

We are seeking collaborations with historians, archivists and digital humanists who would like to evaluate the system on their own collections — and help shape what trustworthy AI for historical research should look like.

Send us a collection