The Method

DeepField is an attempt to compose six lines of established science into one instrument. Each is public. Each is tested. The insights that demand their composition are our own. Whether they compose is the open question.

THE INSIGHTS

Events are made twice. Once in the structure of the world, where pressure builds, and once in the crowd, where reaction amplifies or dampens it. The gap between a shock and its consequence is human behavior. Systems that model only one half keep being surprised by the other.

The signals are individually visible and collectively invisible. The early signatures of most significant events sit in plain sight: in some feed, some filing, some market. No analyst can hold them all at once, and no institution watches them together. The instrument that can is not smarter than the analysts. It is wider.

Agreement is worthless. A prediction that matches consensus is already priced into the world and changes nothing for whoever holds it. Value lives only in calibrated disagreement. An honest instrument must therefore be willing to stay silent most of the time, and be judged entirely on the occasions it speaks.

A prediction is a position, not an announcement. The field publishes forecasts the way people release balloons: whatever happens next is not their problem. We hold that a published prediction stays owned. It must be monitored against the world, revised when its basis weakens, and withdrawn in public when its basis breaks. Accountability is not the press release. It is the maintenance.

Publishing changes the world being predicted. A forecast that moves its subject contaminates its own scoreboard. Every public forecaster has this problem. Almost none accounts for it. An instrument that intends to be trusted must measure its own shadow.

STRUCTURAL SIMULATION

The world’s significant events emerge from structure: supply chains, institutions, markets, and the causal links between them. DeepField models this architecture explicitly, drawing on the causal inference tradition of Judea Pearl and the complexity economics of the Santa Fe Institute. The object is not a forecast of a number but a living map of how pressure moves through a system.

BEHAVIORAL SIMULATION

Systems move because people move. DeepField simulates populations as statistical archetypes: crowds, cohorts, and decision-makers with bounded information, calibrated only against aggregate public data. It models no real individual. It tracks no one. The unit of study is the crowd, never the person.

EARLY-WARNING DETECTION

Complex systems often announce their transitions before they arrive: variance rises, recovery slows, correlations tighten. This is the science of critical slowing down, developed by Marten Scheffer and others, joined with Bayesian change-point detection in the tradition of Adams and MacKay, and with learned detectors trained on critical transitions. DeepField listens for these signatures continuously.

CALIBRATED UNCERTAINTY

A prediction without an honest error bar is a headline. DeepField’s outputs carry distribution-free confidence intervals in the tradition of conformal prediction, and its record is scored the way forecasting science demands: by calibration, bucket by bucket, in the discipline of Brier and the Good Judgment Project.

PRE-REGISTERED EVALUATION

Every central claim is stated before it is tested, with defined conditions under which we will conclude we were wrong. Phase gates decide what advances. Kill criteria decide what stops. The system is built to survive the discovery of its own failure, and to report it.

REFLEXIVITY

A system that publishes predictions about the world becomes part of the world it predicts. DeepField is designed to account for its own influence rather than pretend it has none. What that requires is a subject of the research itself.

THE MACHINERY

In working terms, the instrument runs on the modern toolchain: Bayesian online change-point detection to separate genuine breaks from noise; hidden Markov regime models that sense when a system stops behaving like its own past; structural causal discovery and transfer entropy to recover who moves whom; contagion models from epidemiology for how shocks propagate through networks; deep-learned detectors trained on the signatures of critical transitions; agent-based behavioral simulation at population scale; probabilistic programs that produce full posterior distributions rather than point estimates; conformal calibration that wraps every output in intervals with guaranteed coverage; and adversarial challenge that attacks each candidate prediction before it is trusted. Naming the parts costs nothing. The instrument is the wiring, and the wiring is not public.

THE INSTRUMENT

Each of these six lines was validated alone: in published papers, on historical data, under conditions its authors could control. None of them was built to run beside the others, continuously, against a live world that pushes back. That is the actual problem. Signals arrive faster than certainty. Errors in one method become inputs to the next. A system that speaks about the future changes the future it speaks about. Making six sound methods behave as one honest instrument is the engineering substance of DeepField. It is original work, and it is the part of the program this site will not describe. Parts of it are patent pending. The rest is held closer still.

SIMULATION, NOT SURVEILLANCE

The behavioral layer models people, not persons. It simulates statistical archetypes calibrated only against aggregate, public data. It builds no profile of any individual, tracks no one, and holds no identifying information. The unit of study is the crowd, never the person. This is a design commitment, not a disclaimer.

The work these six lines draw on is public. It is collected, with citations, in the Foundations.

The methods are public. The question is whether they compose.