The Historical Track

Status: BUILDING (2026-07-11). One of two independent extraction pipelines over the corpus; the other is the Conceptual Track. They share only the disambiguated text upstream and link at the graph (a concept-claim references a person/work entity this track built) — no shared logic, no shared ordering. See the unified pipeline for the stage machinery.

Aim

Build the factual layer — who existed, who did what, when, where, which tablets were revealed. Bounded, event/person-based, every fact cited and proof-gated. Entity types: person, work/tablet, place, group, event. This is distinct from organizing doctrine, which is the Conceptual Track’s job.

Ordering = textual rigor / reliability (NOT nominal authority)

The historical track ranks sources by how carefully the text was collected and edited — a factual criterion, deliberately distinct from the interpretive authority that governs the Conceptual Track. ’Abdu’l-Bahá’s authority is supreme, but Memorials of the Faithful as a text is Sohrab’s notes from oral story-telling rendered into English — it can carry transcription, translation, and memory errors — so for fact extraction it ranks below the rigorously-edited scholars. The graph is anchored by the most reliable texts first; looser ones resolve against them.

Duplicate ingests exist for several works (esp. pilgrim accounts and Ẓuhúru’l-Ḥaqq’s two copies) — dedup to one canonical per work before running.

Pipeline

Disambiguate   →   Extract mentions   →   Extract cited claims    →   Reconcile
(place/era/who,     (each reference to      (cited biographical claims,   (evidence-based person
 multilingual)       a person/work, its      proof-gated,                  resolution; to build)
                     identity left deferred)  English-canonical)

Gated: extraction runs only on text that has already been disambiguated. Processed in priority order; person-heavy books get the full pass, doctrinal books (Gate of the Heart) drop out — they belong to the Conceptual Track.

Key mechanisms

• Multilingual model routing — flash for English/Arabic/Hebrew (proven, cheapest); haiku for Persian (flash fails silently on it). Escalation ladder primary→fallback self-heals a passage the cheap model can’t parse.
• Continuation-on-truncation — the answer to Momen’s biblical-genealogy paragraphs: when a single extraction call’s output is truncated, the extractor issues continuation calls (“continue with the remaining claims, don’t repeat”) until the model closes cleanly. Captures arbitrarily dense paragraphs without shrinking output.
• English-canonical output over source text — a claim’s subject, relation, and object are written in English (unifying the multilingual corpus into ONE entity graph: Persian وحید → “Vaḥíd”); the proof span stays verbatim in the source language (the proof gate checks it against the source paragraph).
• Proof gate — every claim carries a verbatim proof-span present in the paragraph, else dropped (no anonymous facts).
• Deferred binding — a claim’s subject/object identities are assigned later by evidence-based reconciliation, never by literal romanization match (per the no-literal-name-binding doctrine).

Status

• Disambiguation: done + hardened (multilingual, dense-para truncation fixed, ~99% coverage on wave-1).
• Extraction (the mention + cited-claim stages): upgraded with multilingual routing + escalation + continuation-robustness + English output. To validate on English wave-1, then the Persian Ẓuhúru’l-Ḥaqq.
• Reconciliation: still to build (evidence-based person resolution / merge-split).