Prove what the review didn't miss

The question from the client's general counsel was polite, almost apologetic. The review team had just worked through 1.2 million documents for a regulatory production, on time, under budget, and she asked the modest thing nobody had planned for: "Can you tell me what it missed?"

Silence. Not because the answer was zero, but because nobody knew, and everyone in the room understood that "we worked very hard" is not a number.

Every large document review has this problem, whether it is litigation discovery, a regulatory second request, or diligence on a thousand contracts before an acquisition. Human review of everything is physically impossible at modern volumes, so teams filter by keyword, sample what they can, and hope. The hope is the weak part. When opposing counsel or a regulator challenges the production, "we did our best" collapses immediately, and the redo costs more than the original review.

You cannot read a million documents. You can prove what you didn't miss.

The method is called technology-assisted review, and courts have accepted it for over a decade, but the part that answers the general counsel's question is the statistics, not the AI.

It works like this. A classifier learns from reviewer decisions which documents look responsive. The model itself is unglamorous; logistic regression over text features does respectably, and the active-learning loop that feeds it, where the machine keeps surfacing the documents it is least sure about for human eyes, is the same algorithm you can read in scikitlearn's documentation tonight. The classifier ranks the corpus, humans review down the ranking, and at some point the team decides the rest is discard.

Then comes the step that makes it defensible: the elusion sample. You draw a random, statistically sized sample from the discard pile, the documents no human will otherwise ever read, and review just that sample completely. The proportion of responsive documents found in it estimates exactly how much responsive material eludes the review, with a confidence interval you can state out loud. Combine that with the richness of the original collection and you get a recall figure: we found an estimated 96 percent of everything responsive, plus or minus, and here is the math.

That sentence, with numbers in it, separates a production that survives a challenge from one that becomes its own litigation.

The discard pile is where the risk lives, so the discard pile is what you sample.

(Lawyers were early to this not because they love statistics, but because they are the one profession whose mistakes get cross-examined.)

The same architecture, classifier plus human review of the uncertain band plus statistical validation of the reject pile, applies anywhere a system screens high-stakes documents at volume. Privilege screens before a production. Contract repositories checked for a problematic clause after a regulation changes. Compliance teams triaging communications. In each case the question that matters is never "how accurate is the model," it is "what got through, and can you bound it."

We shipped this pattern to a team screening a portfolio of roughly 40,000 contracts, where human eyes ultimately read fewer than a tenth of them; the elusion sample bounded what slipped through at under 2 percent, and that figure was the entire reason the work was accepted. The client's outside counsel did not ask a single question about the model, and asked four about the sampling protocol. Wise ones, too.

If your organization reviews documents at a scale where nobody reads everything, and today the answer to "what did we miss" is a shrug, that is a fixable condition rather than a fact of nature. TensorLabs builds review pipelines where the validation is designed in from the first day, so the uncomfortable question from the general counsel gets a number for an answer. Preferably a small one, with a confidence interval attached.