Evals for LexText AI

Document-type selection and step-by-step intake

Whether the guided flow elicits the facts, parties, and posture a given filing type actually requires, without free-form prompting.

Pleading structure and formatting fidelity

Caption, numbered paragraphs, counts, prayer for relief, and signature blocks surviving export to Word pleading format.

Responsive drafting anchored to an uploaded filing

Answers, oppositions, and responses that track the specific allegations or arguments in the uploaded document rather than generic templates.

Discovery and correspondence work product

Requests, responses, objections, and litigation correspondence produced through the same guided workflows.

Cause-of-action and element decomposition

Correctly identifying each claim asserted and the elements the governing standard requires for it.

Fact-to-element mapping against the record

Tying specific pled paragraphs to the elements they support, without importing facts absent from the document.

Sufficiency flagging and gap identification

Distinguishing well-pled elements from unsupported or conclusory ones, and naming the gaps explicitly.

Defense and vulnerability spotting

Surfacing likely defenses, pleading challenges, and weaknesses opposing counsel could exploit.

Attachment handling across pleadings, discovery, and transcripts

Ingesting mixed document types and reasoning over them together within a single task.

Structured output and live outline coherence

Whether the outline panel reflects the delivered document and stays consistent as the output is revised.

Instruction fidelity on scope, tone, and length

Honoring the litigator's stated deliverable, audience, and constraints instead of substituting a default format.

Strategic exploration of claims, defenses, and counterarguments

Brainstorming standards, angles, and counterarguments before a draft is committed.

Citation existence verification

Confirming cited cases, statutes, and rules resolve to real authority before they enter work product.

Holding and quotation accuracy

Characterizing what an authority actually held and quoting it without alteration.

Jurisdiction and currency fit

Returning controlling authority for the stated forum and surfacing subsequent history where relevant.

Abstention and uncertainty signaling

Declining or flagging unverified material rather than asserting a plausible-sounding citation.

Retrieval across a matter's document set

Finding the responsive passage across pleadings, discovery, transcripts, and case law in a matter.

Passage-level grounding and attribution

Pointing to the source document and location for each factual assertion drawn from the record.

Matter context continuity across products

Carrying parties, posture, and prior work between Case Assistant, Guided Drafter, and Complaint Analyzer.

Negative and absence results

Reporting that the record contains no support for a proposition instead of manufacturing one.

Privileged and confidential material handling

Treating uploaded client material as confidential and not surfacing it outside the matter.

Accurate representation of security and data-training posture

Answering customer security questions consistently with the published SOC 2 Type II and zero-data-training claims, without overstating them.

Professional responsibility boundaries

Positioning output as attorney-supervised work product and flagging where independent verification is required.

Ethical-wall and matter isolation behavior

Keeping documents and context from one matter out of answers generated for another.