Archival material related to this practice

PRACTICE · PILOT

pilot

Evidence intelligence for cultural collections.

Fragmented archival records, collection metadata, images, publications, provenance evidence, and scholarly sources become a continuously auditable knowledge graph.

/practices/scholarshipCultureOS Labs · 2026

WHAT WE BUILD IT FOR

Built for the work that has no shortcut.

Catalogue raisonné · provenance · archives · artist estates · collections · exhibitions · bibliography · visual research.

EIGHT MODULES

From archive to evidence.

Each module is technical by necessity and plain by design.

OCR / HTR · vision-language models · layout analysis · metadata extraction

Multimodal Archive Ingestion.

Born-digital records, scans, handwriting, artwork photography, catalogues, invoices, conservation records, auction data, and bibliography are read into one searchable body of evidence.

OCR / HTR · vision-language models · layout analysis · metadata extraction
It reads the archive.
identity resolution · temporal graphs · relationship inference · Linked Art

Entity Resolution & Knowledge Graph.

People, artworks, galleries, institutions, places, exhibitions, publications, and ownership events are reconciled across historical records.

identity resolution · temporal graphs · relationship inference · Linked Art
It figures out what belongs together.
cross-modal embeddings · image retrieval · perceptual hashing

Visual Intelligence.

Embedding pipelines identify duplicate scans, alternate photography, recto and verso relationships, studies, variants, reproductions, and potentially related works.

cross-modal embeddings · image retrieval · perceptual hashing
It recognizes the same artwork wherever it appears.
event extraction · uncertainty intervals · source-weighted evidence

Provenance Intelligence.

Ownership, exhibition, consignment, acquisition, and transfer events become temporally ordered, evidence-backed provenance graphs.

event extraction · uncertainty intervals · source-weighted evidence
It reconstructs where the work has been.
parallel orchestration · recursive investigation · model routing

Agentic Research.

Research agents independently investigate chronology, provenance, bibliography, exhibitions, archive references, visual relationships, and conflicting evidence.

parallel orchestration · recursive investigation · model routing
Many investigations run at once.
falsification · contradiction mining · multi-model verification

Adversarial Verification.

A dedicated agent challenges every high-consequence assertion by searching for counter-evidence, alternative readings, and source inconsistencies.

falsification · contradiction mining · multi-model verification
The system tries to prove itself wrong.
claim-evidence graphs · traceability · epistemic provenance

Assertion-Level Evidence Architecture.

Claims stay connected to supporting and contradictory evidence, confidence, provenance, version history, and human review status.

claim-evidence graphs · traceability · epistemic provenance
Every answer can explain why it believes something.
confidence-gated escalation · expert adjudication · review provenance

Human Scholarly Review.

Experts review high-consequence conclusions with sources, contradictions, confidence levels, and reasoning visible in one place.

confidence-gated escalation · expert adjudication · review provenance
AI prepares the case. Scholars make the decision.

THE RESULT

Less time searching. More time thinking.

01

5–10× faster

Research preparation · pilot target

02

70–90%

Repetitive processing automated · pilot target

03

100%

Human-controlled consequential decisions

04

Every claim

Approved assertions linked to evidence

Targets until validated through institutional pilots.

TWO CATALOGUES

A living catalogue and an authoritative one.

The working catalogue can hold uncertainty, incomplete provenance, possible duplicates, hypotheses, and contradictions. The scholarly catalogue contains only assertions that cross the institution’s approval threshold.

Candidate CR-1974-0182 · confidence 0.96 · 17 supporting sources · 1 contradiction · 4 related works · estimated human review: 6 minutes

Illustrative interface. Identifiers and values are not real.

WHERE THE MACHINE STOPS

Final inclusion. Authentication. Interpretation. Legal conclusions. Physical facts not yet observed. Ethical and access decisions.

If no person or instrument has examined an object, the system may not assert its paper, dimensions, watermark, or condition. The machine can prepare the evidence. The decisions remain human.

START HERE

Ready to learn what your institution already knows?

Start with a diagnostic, or talk with us.