Deepfield: assessments reality can check

The situation

Strategic assessment usually ends in a slide deck. The deck is stale within a quarter. The world has moved by then, and the analyst's reasoning is locked in a PDF that nobody can ask another question of. Nobody ever finds out whether its judgments were right, either. The program office's choices are to live with the stale answer or pay for another that will be stale in three months.

We built Deepfield because we wanted to stop selling memos.

What we built

We built Deepfield with the Andrew W. Marshall Foundation. Its report follows the net-assessment tradition Andrew Marshall built at the Office of Net Assessment. It says what the competition is, how the balance stands and where it's moving, which asymmetries matter, what's being assumed, what could happen, and what we judge most likely. The question can be geopolitical, commercial, or personal.

Every assessment also ends in predictions written down before the fact. The system's job is to write good ones, score them honestly, and learn from the scores.

A run goes like this.

1. Ask. You enter a question. The system drafts a scope (the actors, the horizon, what's in and what's out) and a restated, answerable version of the question. You edit it and approve it, and approval starts the run.
2. Research. The system plans six to twelve searches across the actors and dimensions in the scope, fetches the sources, and checks for gaps, for at most two rounds and within fixed caps on searches, documents, and spending. Each source is stored as it was retrieved, not as a link. Then the evidence freezes. Nothing added after that point can reach a forecast.
3. Assess. A fixed chain of model calls writes the assessment from the frozen evidence alone. It diagnoses the competition and the balance, names the asymmetries and the assumptions (each with what would break it), and lays out two to four scenarios, at least one of which runs against the prevailing view. It ends in three to seven key judgments. Each carries a probability in words and, separately, a confidence level.
4. Propositions. Each key judgment is backed by one to three propositions. A proposition is a precise statement with a deadline and a named way of settling it. A separate check confirms that someone holding the named source on the deadline could answer it with no follow-up questions, and that it wasn't already settled when the research ran. A proposition that keeps failing is marked not checkable, with the reason, and the report lists it that way.
5. Forecast and commit. A forecaster puts a probability on every checkable proposition, taking the median of five independent draws. The forecasts go into an append-only, hash-chained ledger before anyone can know the answer, and from then on they can't change.
6. Review and release. The report stays marked "Draft, not released" until a named person releases it.
7. Resolve and score. When a proposition's date arrives, a resolver settles it without seeing any forecast. Then the forecast is scored against a status-quo baseline, and the key judgment it backs shows as holding, weakening, or contradicted.

Each run keeps the searches as issued, the sources as retrieved, every prompt, the model and provider that answered, and every response. The report is rendered from that record and nothing else, so anyone with access can see exactly what the system saw and said.

How a judgment is checked

Every statement in a Deepfield assessment either cites source passages or is marked as the system's own inference. A citation carries a verbatim quote, and the quote is checked mechanically. It has to appear word for word in the cited passage of the run's frozen source snapshot. A quote that fails the check goes back to the model with the failures listed, at most twice. If it still fails, the run fails. A citation is never quietly dropped.

The key judgments get a second kind of check, from the world. The propositions behind them are dated and settled by a named source, so a judgment that said "likely" can later be shown to have held or not. Every committed forecast counts toward the track record, including those from runs that were sent back, and the report states that track record with its sample size.

Then people read it. Before an assessment goes out, our analysts review the key judgments by hand against the sources and fix or cut what doesn't hold.

What we left out

The first version, retired in October 2026, ran sixteen stages, including a knowledge graph, competing-hypotheses reasoning, course-of-action generation, and wargaming. It grew faster than anyone could verify it. We rebuilt Deepfield around a small core that does the whole job end to end. Wargaming, courses of action, and competing-hypotheses matrices are out of the core, and each comes back only when it can show with evidence that it makes the assessments better.

What it replaced

The default alternative is a consultancy engagement that delivers a PDF. Deepfield is the infrastructure underneath that engagement, kept running. The same question can run again when the world changes, without re-paying for the analyst's learning curve. Any claim in it can be checked against the passage it quotes, and its predictions get scored.

What a similar engagement looks like

Deepfield is a platform we embed. A typical engagement runs 12 to 16 weeks. We set it up for your domain, work through the scope of your first questions with your team, and run the first real assessment together. At the end you keep the running system, the code, and your data.

It's a fit when you have a recurring assessment need and a requirement that the conclusions you act on be checkable against the evidence behind them. If your decision is one-shot and never repeats, Deepfield is too much infrastructure. If it's ongoing and the stakes are high, a platform you run yourself costs less over the first year than two consulting engagements, and its track record grows with every question you ask.