Precedents research for drugs and biologics

Guidance tells you what the agency says it wants. Precedent tells you what it has actually accepted — and the gap between those two is where regulatory strategy lives. A guidance document describes an endpoint in general terms; the review of the drug approved on that endpoint three years ago tells you what data package satisfied a specific division, what the reviewers pushed back on, and what got negotiated. One is a policy statement. The other is evidence.

That's why the questions that matter most sound like this: Why did the last twenty products in this class do X and none of them do Y? Has FDA ever accepted this surrogate endpoint in this population? What did the CRLs in this therapeutic area actually object to, and what fixed them? Has anything been approved despite missing its primary endpoint — and how? None of those have an answer in a guidance library. All of them have an answer in the record.

And underneath it sits the question nobody asks out loud: have we exhausted it? Not "did we find some precedent" but "is there something we haven't seen that a reviewer will raise, or that would have changed our approach if we'd known?" That anxiety is the actual job.

The material is scattered across document types that were never designed to be read together:

• Review packages (SBAs, summary reviews, discipline reviews): the richest source by far, because they record disagreement: what the statistical reviewer thought, where clinical and CMC diverged, what the division concluded anyway.
• Complete Response Letters: precedent about failure, which is often more informative than precedent about success.
• Approval letters, labels, and post-marketing requirements: what was granted, and what conditions came attached.
• EPARs and CHMP opinions: the EU parallel, including refusals and re-examinations, which is where you learn how the same data package fared with a different authority.
• Advisory committee materials and votes: the arguments in the open, plus the questions the agency chose to ask.
• Clinical trial records and protocols: what was actually run, including the programs that quietly stopped.
• Guidance, ICH, and authority Q&As: the policy frame the above gets interpreted against.
• Designation records (orphan, breakthrough, fast track, PRIME): what evidentiary bar cleared for what population.

Three properties make this genuinely hard to search. Most of it is scanned image PDF with no machine-readable text — a crawler and a keyword search both come back empty on documents that plainly contain the answer. The vocabulary doesn't match either: you don't know how the agency phrased your concept, so you can't enumerate the synonyms to search for. And the answer is almost never in one document; it's a pattern across thirty, which surfaces only once all thirty have been read.

Manual research (Drugs@FDA, EMA's site, ClinicalTrials.gov). Everything is public and every citation is real. But hunting patterns across twenty or thirty products by hand takes a week, so the scope gets cut to five, and the pattern you needed lived in the twenty-five you skipped. This is the honest baseline, and its failure mode is the sample.

Guidance and rules libraries (Cortellis-class suites, IQVIA Regulatory Intelligence). Excellent at what changed and what the current requirement is in a given market. Structurally not precedent tools: they hold curated summaries of policy, while the reviews and letters where decisions actually got recorded sit outside the library. Useful stack neighbors, wrong instrument for this job.

General AI alone (ChatGPT, Claude, Gemini, Copilot). Genuinely useful for orienting on a well-documented topic. But precedent research is where its structure bites hardest: it answers from memory plus a shallow web pass, the review packages are invisible to that pass, and its economics assume seconds-not-minutes. You get the famous precedent while the complete record stays buried, and the citations often don't survive being opened. Copilot compounds it by grounding in your own tenant when the whole point is the world outside it.

Consultants. The right call when the hard part is judgment — a novel pathway, a division relationship, someone accountable in the room. But hours don't scale to exhaustive reading, so the wide scan becomes a billable line item or doesn't happen. The strongest arrangement is a consultant making the call on top of research that was actually exhaustive.

In-house AI. Loading approval PDFs into an internal bot demos well and degrades structurally: OCR for the scanned back catalog, weekly refresh across authorities, and an evaluation harness proving the agent didn't fall back on memory. Build AI where it differentiates you — a model predicting whether your candidate works is core IP; reading the public record isn't.

Rhizome. Rhizome is a focused, AI regulatory coworker built specifically for this job: what has your regulator said and decided, and what does it mean for your program. Our agents can only state facts pulled from primary-source documents; memory doesn't count as a source, and we benchmark them on that discipline. So the answer to "has FDA accepted this endpoint?" is assembled from the reviews themselves. A hard precedent question reads hundreds of documents, sometimes thousands, with up to 1,000 citations in a single answer, and each one opens to its exact passage inside the 500-page review, so you land on the sentence that decides it.

The corpus is what makes exhaustiveness possible: >45 million documents across >75 health-authority databases in >12 markets as of August 2026 (on pace for >150 databases and >25 markets by year end), with every line indexed and every scanned image processed, because that's where the substance of a review package actually is. No hallucinations reported in 18 months: a track record, not a guarantee, which is exactly why every claim is openable. Free tier and published pricing, and it connects into Claude, ChatGPT, or Copilot over MCP if that's where your team already works.

Cons, honestly: precedent research helps least where there is little precedent — a genuinely first-in-class mechanism with nothing comparable in the record. Interpretation remains yours. And nothing in the public record substitutes for the confidential parts of someone else's submission.

1. State the decision, not the keyword. "Can we use this surrogate endpoint in this population?" beats searching for the endpoint's name.
2. Ask across document types at once. The pattern crosses reviews, CRLs, labels, EPARs, and trial records, so the search has to as well.
3. Go wide before narrow. Thirty comparable programs, then the interesting five. Starting narrow is how you miss the counterexample.
4. Hunt the negatives deliberately. Refusals, CRLs, failed trials, approvals despite missed endpoints. Negative precedent is under-searched and disproportionately decision-relevant.
5. Reconcile the sources. The review, the published paper, and the trial registry entry often disagree. The disagreement is information — a good answer surfaces it and lets you weigh which source governs.
6. Open every citation you plan to rely on. Non-negotiable before anything reaches a regulator, whatever produced the draft.