Technical report
Lineage: Evidence-Linked Due-Diligence Drafting with Bounded Model Authority
Abstract
Language models can help assemble corporate due-diligence files, but a fluent draft is not sufficient evidence that its claims are supported or that the system respects approval boundaries. This report describes Lineage, a prototype that combines deterministic evidence collection, source-linked drafting, personal-data redaction and human authorization. A fixed jurisdiction-specific workflow gathers corporate ownership, legal-entity and screening records. Each generated claim carries a source identifier and quotation; a deterministic check verifies quotation presence, while a separate model judge assesses support. The repository contains a 240-row screening dataset, 30 ownership cases, 40 safety cases, 46 redaction passages and 26 policy questions. Recorded citation experiments produced 40 files containing 497 claims, of which 491 were judged supported. All 40 files passed the citation checks. A separate safety baseline reports zero failed runs across 40 cases repeated eight times, including clean controls. These results describe a small, author-maintained test collection and are not a general safety guarantee. A one-time GCP deployment exercised authentication, access controls, sign-off and audit verification with fixture-generated drafts, not live model drafting. The report explains the architecture, metric denominators, observed unsupported claims, evidence limitations and experiments required before broader operational claims are justified.
Citation
Deavyansh Gautam. (2026). Lineage: Evidence-Linked Due-Diligence Drafting with Bounded Model Authority. Technical report.