Historical Knowledge Fabric

Estimated reading time: 6 minutes

OCR makes historical pages readable. It does not make historical knowledge trustworthy.

The hard problem begins after extraction. An AI-assisted knowledge system has to preserve what a source says, which entity it refers to, what we infer from it, what conflicts with it, and how that judgment changes over time. That is epistemic state management.

Preserve the distance between what the source says and what the system believes.

My test case is Hyman Fried, my great-grandfather. I had already resolved his arrival and presence in New York. The harder problem came next: deciding which relationships belonged to him without turning plausible kinship into asserted fact.

Entity Resolution Is Not Relationship Resolution

In Hebrew he is Chaim. In New York he is Hyman. The 1973 stone says חיים ב”ר מענדל, Chaim son of Mendel. The 1942 draft card names Meryszczów, while other American papers round his origin to Przemyślany. Those variations can bind to one person.

You can correctly resolve Chaim and Hyman to one person and still be wrong about every edge attached to that node.

In 1928, Jides Fried appears in a Przemyślany marriage register as a daughter of Mendel Fried and Finkel Schaffel. Rudie’s 1929 marriage act names the same parents. Those records establish a Galician household while Hyman is already living in New York.

The relationship to Hyman is indirect. Jides is a register-backed child of Mendel and Finkel. Her placement as Hyman’s sister is strongly supported, but his own Galician birth record is still missing and a competing reading leaves his mother’s identity open.

The node is resolved. The sibling edge remains a revisable interpretation.

Unknown Must Remain Representable

In 1922, a Mechel Fried arrives on the Paris and names an uncle Chaim Fried at 754 Sackman Street. That contact line is a claim about a relationship. The census at that address points to a different Hyman, and the later checks do not supply a bridge. The uncle edge stays parked.

Meanwhile the same registers contain another Chaim Fried, house 52, married to Ruchel Zeisel, with a daughter also named Jides. Same given names. Same town cluster. Separate household. A second Mechel, son of Schyje Mendel Dier and Blime Fried, marries in Przemyślany in 1924. Keep him separate from the New York passenger until a bridge exists.

Two Jideses. Two Chaims. Two Mechels. That is the stress test. A system optimized for plausible completion will resolve ambiguity faster than the records warrant.

An embedding-first retrieval system may correctly surface both Mechels as semantically similar. The failure happens when similarity becomes identity without a separate entity-resolution decision. The fabric keeps semantic proximity from becoming identity.

Unresolved belongs in the model as a durable state. Relationship hypotheses need explicit statuses such as accepted, disputed, parked, and rejected. Identity needs its own resolution state. A scalar such as sibling_probability: 0.82 cannot tell us what supports the relationship, what conflicts with it, what is missing, or what would change the conclusion.

The Five Distinctions

I call the portable model The Five Distinctions:

Layer	Stored object	Hyman case
Source	Immutable artifact and citation	1928 marriage register image
Extraction	Human or machine observation	Mendel Fried; Finkel Schaffel
Claim	Atomic source assertion	Jides is their daughter
Identity	Explicit entity binding	This Jides differs from Jides in house 52
Interpretation	Revisable synthesis	Jides is probably Hyman’s sister

This change from Evidence to Source matters. In genealogical practice, a source contains information. That information becomes evidence when it is evaluated for relevance to a research question. The artifact remains the source; evidence is the relevant information brought to the question.

Together, the five layers make a Historical Knowledge Fabric: a claim-centric provenance architecture in which sources remain unchanged, claims retain citations, identity bindings stay explicit, and interpretations can be revised without rewriting earlier layers.

A conventional knowledge graph records relationships. A provenance-preserving claim graph also records who asserted each relationship, from which source, how the entities were resolved, what contradicts it, and whether the current interpretation accepts it.

What This Means Computationally

A small object makes the separation visible:

claim:
  id: claim-1928-jides-parents
  subject: mention-1928-jides
  predicate: child_of
  object: [mention-1928-mendel, mention-1928-finkel]
  source: przemyslany-marriage-1928-17
  extraction: "Mendel Fried; Finkel Schaffel"
  status: accepted
identity_binding:
  mention: mention-1928-jides
  entity: jides_fried_01
  status: resolved
interpretation:
  relationship: sibling_of(hyman_fried_01)
  status: supported_not_proved
  missing: hyman_galician_birth_record

The final family tree should be a view over these objects. Treating the tree as the datastore makes the source trail disappear.

Here, accepted means the claim faithfully represents what the cited source says. The inferred sibling relationship keeps its own status.

Microsoft’s GraphRAG indexing pipeline is a useful technical bridge. It extracts entities and relationships from source text and can optionally extract claims. That makes unstructured material available for graph-based retrieval. The Historical Knowledge Fabric addresses the next obligation: preserving provenance, identity state, disagreement, and interpretation history when those extracted objects are considered for canonical knowledge.

Vector search retrieves. Adjudication belongs in a separate layer.

A Fabric Has to Remember Its Reasoning

The idea sits beside established provenance work rather than replacing it. W3C PROV represents the entities, activities, and agents involved in producing information so people and systems can assess quality, reliability, and trustworthiness. Historical and AI-assisted research add a demanding edge: claim-level disagreement, explicit entity bindings, unresolved relationships, and interpretations that remain open to revision.

Three established patterns sharpen the boundary. OpenLineage records operational lineage through datasets, jobs, and runs. SQL:2011 temporal modeling separates valid time from system time, while Dublin Core gives the source itself a creation date. Nanopublications separate an atomic assertion from its provenance and publication metadata. A Historical Knowledge Fabric brings those disciplines to a different question: why does the system believe this relationship now, and what would revise it?

Genealogists will recognize the discipline. The Genealogical Proof Standard calls for source citation, analysis and correlation, resolution of conflicting evidence, and a reasoned conclusion. A Historical Knowledge Fabric gives AI-assisted research a data architecture capable of preserving the distinctions that proof process depends upon.

The storage boundary matters. Put scans and record images in immutable artifact storage. Put citations, extractions, claims, identity bindings, exclusions, interpretations, and tests under version control. The artifact store preserves the records. The repository preserves the reasoning state.

That extends the architecture across this series. Knowledge as Code versions the knowledge model. Negative Knowledge preserves rejected hypotheses. A Historical Knowledge Fabric separates sources, assertions, identity decisions, and conclusions so that a new record can change the household reading without flattening its history.

The Agent Test

Ask an agent a harder question than “Who was Hyman Fried’s sister?”

Ask: “Why do you believe these two people were siblings? What evidence contradicts that conclusion? What new record would change your answer?”

Passing all three is the test for preserved knowledge. A fluent answer alone is only a retrieved conclusion.

Madam I’m Adam

This continues the thread from Negative Knowledge: rejected hypotheses preserve what proved wrong. A Historical Knowledge Fabric preserves the distance between what a source says and what the system believes.