IRON LAKE

Product · Litigation

Privilege review,
attorney in the loop.

Model-assisted classification across a document corpus, with every call routed to a reviewing attorney. The deliverable is not a pile of tagged documents — it is a privilege position that survives a challenge.

▸ Running today on an active Florida probate matter.

The problem

1. The volume is inhuman and the judgment is not delegable.

A modest matter produces tens of thousands of emails. Somewhere in them are a few hundred that are privileged. An attorney has to decide which, because the decision is a legal one — but reading everything to find them is the least valuable use of the most expensive hour in the building, and the fatigue that sets in around hour forty is itself a source of error.

2. A wrong call in either direction is expensive.

Withhold something that is not privileged and you invite a motion to compel and a credibility problem with the court. Produce something that is privileged and, depending on the jurisdiction and how you handled it, you may have waived it — sometimes across an entire subject matter. This is not a domain where "mostly right" is a product.

3. So the model cannot be the one who decides.

A classifier that is right 95% of the time is genuinely useful and completely unacceptable as the final word. The 5% is the malpractice. The only defensible arrangement is one where the machine narrows and a licensed professional decides — and where the record shows exactly that, line by line, if anyone asks.

What it does

Narrow, review, produce.

Ingest
Loads the corpus with provenance intact — every document keyed to its source, nothing overwritten, so the collection is reconstructable.
Propose
Model-assisted classification proposes a call and a basis for it, against the privilege categories the matter actually uses. Proposals carry confidence and are never treated as decisions.
Review
Every proposal is routed to a reviewing attorney, who confirms, overturns, or escalates. The attorney's decision is what enters the record. Disagreement with the model is captured, not smoothed over.
Produce
Privilege log, Bates ledger, and issue binder generated from the decisions — not from the proposals. What opposing counsel receives is an artifact of human judgment with a machine's reach behind it.

The record is the deliverable

A privilege log is either defensible or it is worthless. So the log does not record what the system concluded — it records what a person decided, when, and on what basis, with the machine's proposal preserved beside it.

;; every call carries who decided it, and when
{
  "bates":        "MATTER-0004182",
  "proposed":     "attorney_client",
  "confidence":   0.91,
  "decided":      "attorney_client",
  "decided_by":   "reviewing_attorney",
  "decided_at":   "2026-07-14T16:22:09Z",
  "basis":        "counsel in cc; legal advice sought"
}

;; the model proposes. it never asserts privilege.
;; the log records a decision a person made.

That shape matters under challenge. When the question is "how did you decide this," the answer is a named attorney, a timestamp, and a stated basis — with an auditable trail showing the model narrowed the field and did not make the call.

Status, plainly. The review service, classification pipeline, privilege log, Bates ledger and issue binder are deployed and in use on a live matter. A legacy batch of model-call errors is still being reprocessed; the current workbook is not represented to counsel as a completed log. We would rather say that here than have it discovered.

Running a document-heavy matter?

If you are staring at a corpus and a deadline, we would like to hear what you are up against. We are deliberately taking few matters while the review workflow hardens.