Trust, Layered: Warranted AI Knowledge Extraction for Interpretive Corpora, Demonstrated on Jung's Collected Works (dataset, pipeline, and paper)

A warranted atlas of C. G. Jung's Collected Works, graph v2.2.0: 807 nodes, 3169 edges, 4280 references over nine volumes. Each reference is anchored to a numbered paragraph with a verbatim quote of at most 25 words, gated by a model of a different family from the proposer, and cross-vendor audited (95% of references). A normalization layer (260 relations) and a relation-family layer (ruleset v0.7.2, calibration record pipeline/family_calibration/records/2026-08-13_v072_run4.json, PASS) inherit no warrant from the claim beneath. Every claim carries a warrant profile: the checks it survived, who made them, and which nobody has made. v2.2.0 adds CW 5, Symbols of Transformation, mined end to end, after v2.1.0 added CW 11, Psychology and Religion: West and East. Includes the full transaction ledger, all experiment inputs, verdicts and adjudications (E0-E10, Studies B and C), the family calibration records, and the James pilot. The paper, Trust, Layered, stays as published with v2.0.0 and describes that release: 589 nodes, 2091 edges, 2678 references. Graph v2.2.0 · ruleset v0.7.2 · calibration record 2026-08-13_v072_run4 (PASS). Succeeds v2.1.0. Copyrighted source texts are not included. Quotations are at most 25 words, under quotation exceptions (see docs/LEGAL.md). The code is MIT and the dataset's original structure and metadata are CC BY 4.0. Both licences exclude the verbatim quotations, which remain with their rights holders (see NOTICE). Live atlas: https://symbolicworld.observer