Legal work composes evidence and authority across adjudication, transactions, corporate governance, regulation, tax, and intellectual property. Adela separates grammar declaration, source reading, identity resolution, and product projection so probabilistic models propose candidates while a deterministic kernel owns what enters the matter.
01 / Grammar & Authority
Executable grammars learned from sources and rules.
Each document kind has its own syntax, promulgator, revision, and semantic types. When a known form or filing arrives, the runtime reuses its admitted declaration so a scan and a native PDF yield the same structure. When an unfamiliar document kind or a new court arrives, the runtime learns the grammar from unconsumed page remainder or compiles it from published rulebooks and filing portal specifications—admitting a revision only when every production type-checks and increases coverage.
02 / Witnessed Reading & Context
Propositions, provenance, and explicit uncertainty.
Reading under an admitted grammar binds values to source constituents with page-level witnesses and records any unconsumed material as explicit remainder. The runtime separates the context of production—who spoke, when, and under what attestation—from quoted exhibits and embedded records. A blank slot or an unreadable field remains an open dependency rather than a silent default.
03 / Scope-Indexed Resolution
From page mention to settled entity without losing the witness.
Identity resolves through four scopes: a witnessed mention on a page, a document entity within one source, a matter entity across the record, and a canonical person, organization, court, or authority in the workspace. When an identity settles, every role that actor fills updates across the matter while retaining the exact surface and page coordinate that introduced it. Conflicting ownership stays open for human review.
04 / Products, Events & Durable Reduction
One proposition store across views, timelines, and drafts.
Document views, object-centric event timelines, cross-matter search, and verified drafts are projections of the same settled propositions. Every reduction step is keyed by deterministic identity over versioned inputs, so when new evidence arrives, a rulebook updates, or a review question is answered, dependent work resumes and converges without duplicating records.
