How LexVex computes treatment, topics, and case attributes.
LexVex is intentionally transparent about what is algorithmic, what is ontology-backed, and what still needs a human reader. The goal is not to hide the model — it is to show the model’s outputs in a form that can be checked against the underlying opinions.
AI labels, sentence-level evidence.
Treatment is computed by analyzing the full citing opinion text and asking the classifier to choose exactly one of six labels: followed, discussed, distinguished, criticized, not_followed, or general. The classifier returns supporting sentence ids, and the UI resolves those ids back to the sentence catalog so the displayed quotes are verbatim — not LLM-paraphrased.
Those supporting quotes come from the opinion sentence catalog stored in the database. In the citator, AI-analyzed rows are surfaced first, followed by the older/manual rows, so the newest algorithmic results are the easiest to review.
Depth, confidence, and queue priority.
The treatment graph is driven by citation depth and confidence signals. On the case page, the summary badges and counts reflect the treatment rows already computed for that cluster, while AI-backed results are ordered ahead of legacy results in the citing-opinion list. Nightly refreshes mark recently changed clusters for reanalysis, and the backfill runner always drains the priority queue before falling back to the highest-cited remaining cases.
LexIndex underpins the topic taxonomy.
LexVex’s active topic tree now comes from the editorial LexIndex taxonomy. The Case attributes section still extends the separate branch-facet system into Industry, Location, Forum, Objectives, and Actors.
Topic taxonomy attribution pending final approval.
Algorithmic summaries still need verification.
Citation treatment, topic assignment, and case attributes are algorithmic summaries. They are designed to help you navigate a case, not replace the case. Always verify the result against the primary sources before relying on it.
Treatment flags are computed from the full citing-opinion graph and can change as new opinions are added or older opinions are reprocessed.
Nightly refreshes, priority backfills, and stable display order.
Nightly jobs enqueue recently changed clusters for reanalysis, and the backfill runners process queued items first before moving to the most-cited remaining cases. That keeps the corpus moving forward in a predictable order while making the most visible cases available sooner.
Corpus refreshes happen on a nightly cadence, so the page you are reading is a snapshot of an evolving dataset rather than a static digest.