Harvard NeuroLaw Library

The audience is wide and the writing is narrow. Neurolaw research shapes sentencing, competency rulings, and appeals, and the people it affects include defense attorneys, prosecutors, judges, advocates, clinicians, families, and incarcerated people. The source material is dense, citation-heavy, and written at one expert reading level.

The old options were both bad: publish the original and lose most readers, or hand-write simpler versions that do not scale and go stale on the next edit. The goal was every document readable at the reader's level and the whole library searchable from one box.

Open a document and you get a Comprehension Slider. Drag it and the same text re-renders at a different reading level, from the original down to plain 5th-grade language. Every level is generated ahead of time and stored at publish, so switching is instant. The original sits at the end of the slider, so the text you'd actually quote is always one drag away.

Two phases, kept apart on purpose. An editor flips a generate flag and publishes. A Contentful webhook fires, the document body is converted to clean Markdown, and Gemini 2.5 Flash rewrites it at four reading levels under fixed rules: third person, no citations, keep the structure, about five paragraphs. Separate passes pull keywords and write an at-a-glance summary. Each version is written back to Contentful as Rich Text in its own field, the original untouched.

Reading costs nothing. Pages are static on Vercel, and the slider swaps pre-stored text with no model calls.

The first SDK we used accepted a thinking-budget setting and ignored it. Every rewrite ran with the model thinking, about a minute per call, and the six calls per document ran one after another. Long documents hit the function timeout, and by mid 2026 a backlog of 167 articles and 29 amicus briefs sat stuck behind the generate flag.

In June 2026 we moved to Google's current GenAI SDK, which honours the setting, and ran the six generations concurrently. A document now finishes inside the 300-second limit. The change estimates about five times faster and cheaper per call. That figure comes from the engineering change, not from a production measurement.

Readers think in questions, not content types. One Typesense index covers articles, briefs, cases, and affidavits with typo tolerance, synonyms, and filters for type, state, court, circuit, jurisdiction, year, and topic.

The August 2026 rebuild added semantic search. Each document gets a 3072-dimension embedding from Google's embedding model, built from its title, keywords, and the opening of its text, and a query combines that vector with keyword matching. The index now covers the body of each document, not only its title and keywords. A reindex builds a fresh collection, swaps an alias so the library never goes dark, and aborts if more than half the imports fail.

The original is never replaced. The AI versions help people understand a document, they do not stand in for it. The original is always on the slider, and the quote and citation tools point back to it. You read the simple version, you cite the original.

Every resource carries citations in five formats: APA, AMA, MLA, NLM, and Bluebook.

Acceptance testing ran on Vercel preview deployments with comments. The library team pointed at the exact element on the exact page, we replied in the thread, and the fix shipped to the same URL. Feedback never left the deployment.

The model was the easy part. The real work was getting thousands of documents into one place, making each one readable at any level, and keeping people in charge of what ships. Because the heavy work runs once at publish and the rules live in editable prompts, the library keeps growing without re-engineering.

The 2026 upgrade replaced the SDK, the search, and the review flow without changing what a reader sees: the same slider, now faster to fill.