AI

AI is a layer over the whole file — it cuts across every capability, building on the data model and the boundaries already in place. One rule shapes it: it reads the record and writes back into the primitives the system already has. Every output is a proposal the reviewer edits, accepts, or rejects — so each suggestion is an auditable record, and a person stays accountable for the decision.

What it reads, what it writes

The inputs are the documents and their extracted text (the search sidecar). The outputs land in the platform’s primitives:

  • Issue extraction — read the whole file and propose the unified issue list: the adjudicated matters, each with its scope. This populates the Issue object; the reviewer curates it.
  • Relevance flagging — surface the documents and pages most relevant to a given issue, so a 2,000-page file opens to the evidence that matters first. Output is an annotation-like marker, scoped to the appeal.
  • Document consolidation — recognize that scattered documents belong together (a submission and its attachments, a form and its continuation) and associate them as one logical record.
  • Translation detection — flag non-English documents that need translation, the same markers the Reader already supports, applied automatically.
  • Task generation — scan the file for actionable items and propose the tasks they imply, feeding the workflow engine.
  • Issue-level reporting — analyze outcomes and trends by issue type, which only becomes possible once issues are first-class (see Reporting).

Why it rides on the existing model

Every capability above is a composition of things the system already does: read the document text, write a record keyed by appeal ID, respect the isolation boundary. That has three consequences worth stating:

  • Isolation holds for free. An AI suggestion is a record scoped to one appeal, written through the same path as a human annotation — it cannot reach across appeals any more than a reviewer’s note can.
  • Outputs are auditable. A flag, an issue, a task — each is a stored record with an author (the model) and a timestamp, visible and reversible like any other edit. Every action surfaces as a record you can read.
  • The contract is unchanged. AI-produced issues and tasks flow downstream through the same correspondence boundary as human-produced ones. The receiving system reads them the same either way.

In short

Every AI output is a stored record on the existing substrate: read the file, propose a record, with the reviewer as editor and the audit log as the account of what was suggested and what was kept.