DeepField

AN INDEPENDENT RESEARCH PROGRAM

Applications

DeepField is one instrument, and an instrument is defined by what it can be pointed at. This is the full catalog of what it’s being built to do. One caveat governs every line of it: the entries describe what the system does, and whether it does it well enough to sell is the open empirical question the current work exists to answer. Parts of the machinery run today. Most of it is under construction. None of it is sold until the record earns it. Each entry carries a small diagram, computed at build time from a model stated in the build script. They show what the product is. They are not measurements, and every one says so on its face.

A · CORE PREDICTION PRODUCTS

Commodity price and supply predictions
Forward predictions on commodity prices and supply conditions, copper, oil, and wheat first, delivered as full probability distributions, not point guesses, with stated confidence, a stated horizon, and the causal chain behind each call written out.A price history line followed by a forecast fan of nested probability bands widening over the horizon, with a dashed median: the forecast is a full distribution, computed from 400 simulated continuations of a stated process.The grey line is the past. The blue fan is the forecast: every path the model considers plausible, with the middle path dashed. The wider the fan, the less certain the call.
Regime-change and tipping-point early warning
Detection that a market is leaving its current regime, the statistical character of the world changing, before the break is obvious. Three stacked detectors: hidden Markov regime detection, critical slowing down monitoring, and a neural early-warning-signal detector. The product is time: knowing a transition is forming weeks before consensus does.A noisy series that breaks regime near the end. A warning line placed by a fixed statistical rule fires 83 steps before the break, and the span between warning and break is shaded: the product is time.One noisy series, left to right in time. The rule fires at the solid rose line; the break arrives at the dashed one. The shaded span is the head start.
Cross-domain cascade theses
Multi-step causal calls that connect domains: Chinese import regulation, Red Sea shipping stress, and smelter maintenance season composing into a copper supply disruption within sixty days. Signals propagate across domains and the causal chains assemble the thesis. This is the flagship intellectual product: the call nobody else’s siloed research produces.Three stacked series: a shock in one domain propagates to a second domain 22 steps later and reaches copper supply 49 steps after the shock, with each crossing marked. The thesis is the composed, multi-step call.Three domains stacked in time. A shock crosses the top series first, moves the middle one 22 steps later, and reaches copper supply 49 steps after the start.
Time-priority access
The core paid product. Every published prediction releases on a staged schedule, earliest to the subscribers who underwrite the work, then successively wider audiences, public last. What the early tiers buy is being early on a verified divergence, in the window before consensus converges. Capacity is capped per domain, so the alpha isn’t sold to fifty competitors at once.A decaying consensus-gap curve with three release markers: first release with 100 percent of the gap left, a wider stage at 33 percent, public at 9 percent. Early access buys the window before the gap closes.The blue curve is how far ahead of consensus a call is, decaying as the world catches up. Each marker is a release stage; earlier stages get more of the gap.
Confirmatory predictions
Predictions that passed the confidence gate but agree with consensus: no divergence, so no headline value, sold as validation instead. Independent, verified confirmation of a thesis the client already holds. Low glamour, high retention.Two overlapping probability curves: consensus dashed, and an independent forecast almost on top of it, 0.4 sigma apart, below the publication bar. Agreement is sold as verified confirmation instead of published.Two bell curves: the dashed one is what the crowd expects, the blue one is the independent forecast. They nearly coincide, so the call sells as confirmation, not news.
Lifecycle-managed predictions
Every prediction is a live position, not a press release: confirmed, revised, or withdrawn as the world moves, with every transition signed, timestamped, and pushed to subscribers as it happens. Clients act on predictions knowing they’ll hear immediately when the system changes its mind.A monitored basis-health line dips below a dashed review trigger; a state band below switches from live to under review at that rule-fired moment, then to revised, signed and timestamped.The wobbling line is the health of a prediction’s reasoning, watched over time. When it crosses the dashed trigger, the status bar below changes state automatically.

B · INSTITUTIONAL DECISION TOOLS

Risk surface aggregation
Cross-prediction exposure analysis: these seven of your positions share one causal dependency, and if that node breaks, they break together. It answers the institutional question no single prediction answers: what am I exposed to across everything DeepField is telling me?An eight by eight correlation grid computed from a stated factor model, with a four by four block outlined: four positions sharing one causal node, which break together if that node breaks.Each square is how strongly two positions move together, darker meaning more. The outlined block is four positions that secretly depend on the same thing.
Prediction replay
Deterministic reconstruction of what the system believed at any past moment, with lookahead structurally impossible: show me how this would have behaved in March 2022. The due-diligence unlock: a fund audits the machine’s historical behavior instead of trusting a marketing deck.A belief line with an uncertainty band reconstructed up to a red cursor labeled the query, March 2022. Everything right of the cursor is ghosted: the reconstruction cannot see it, by construction.The blue band is what the system believed, rebuilt moment by moment up to the red cursor. The grey right half is the future the reconstruction is not allowed to see.
Correlation cluster display and alerts
A live view of which active predictions are correlated, through shared causal ancestry or shared regime exposure, with alerts when the cluster structure shifts: portfolio-level situational awareness on the prediction book itself.Two ring networks of the same ten predictions, last week and this week. In the second, five new correlation edges are drawn solid in rose and five vanished edges are drawn dashed, ten changes in all, and an alert fires on the structure shift.Each dot is an active prediction; lines connect ones that move together. Solid rose lines are new connections this week, dashed ones just vanished. That shift triggers the alert.

C · THE SIGNAL FEED

The quant data feed
A licensed, machine-readable data product for systematic desks that want inputs to their own models, not finished predictions: per-domain novelty scores, regime states with confidence, consensus-gap time series, cascade alert intensity, and aggregate risk-surface statistics. It sells features derived from DeepField’s instruments: never the predictions, never the instruments themselves.Four aligned machine-readable tracks reacting to one marked event: a novelty score, a regime-state stripe, a consensus-gap series, and a cascade alert intensity curve.Four data streams reacting to the same moment, marked by the dotted line: a strangeness score, a regime flag, the gap to consensus, and a cascade alarm.

D · COMMISSIONED AND BESPOKE

Scenario studies
Commissioned forward simulations at institutional depth: model the global cascade if the Strait of Hormuz closes for thirty days, or stress our supply chain against a copper shock. The simulation machinery pointed at one client’s question, delivered as a study.Two forecast fans sharing identical randomness until a marked shock labeled: the commissioned question, chokepoint closes 30 days. The scenario ensemble lifts 18 percent above baseline at the median.Two futures simulated with identical luck. Grey assumes nothing happens; rose assumes the chokepoint closes for thirty days. The gap between them is the scenario’s cost.
Commissioned domain expansion
A client funds DeepField to build coverage of a domain it doesn’t yet watch, a commodity, a region, a regulatory environment: sensing build-out, data validation, shadow calibration, then published-domain status. The expansion cost becomes revenue, and the domain then becomes permanent product.A calibration uncertainty curve falling as one over the square root of accumulated observations, crossing four phase markers: sensing build-out, data validation, shadow calibration, published domain.The curve is how uncertain the model is about a new domain, falling as observations accumulate through each build phase until the domain is fit to publish.

E · POLICY-ADJACENT PRODUCTS

Policy-consequence forecasting
Forecasts of what a government action does to markets: a tariff, an export control, a sanctions package, a subsidy, a strategic-stockpile action, and the price, supply, and substitution response that follows. Policy events become upstream signals, and the causal machinery projects the consequence. Buyers: trading houses, corporates, sovereigns.A marked policy event followed by three impulse-response curves at different speeds and signs: price reacting fastest, supply dipping slower, substitution rising slowest.The vertical line is a government decision. The three curves are what follows: price jumps fastest, supply sags slower, substitution builds slowest of all.
Contingent-outcome forecasting
Not who wins: what copper, grain, and energy do conditional on each outcome. Per-branch forecasts on an externally resolved trigger, an election, a policy vote, a court ruling, under a hard rule: no likelihood over the trigger outcomes themselves is ever published or derivable from what is. It captures what markets actually trade: the consequence, not the horse race.A trigger node labeled election, ruling, or vote splits into two branches, each ending in its own forecast distribution for copper. A note states no likelihood is ever published on the trigger itself.The circle is an event nobody forecasts here, an election or a ruling. Each branch carries its own full forecast of what copper does if that outcome happens.
Sovereign import-exposure and food and energy security profiles
Country-level exposure analysis: a nation’s import-dependency graph run through the risk-surface machinery. Your food security has concentrated exposure to these three supply chains, and here is what stresses them. Buyers: governments, multilaterals, development banks.Horizontal bars of one nation’s food import shares by supply chain. The top three chains, highlighted, carry 76 percent of supply, with a concentration index computed from the shares.Each bar is one supply chain’s share of a nation’s food imports. Three chains carry 76 percent of the total: the map of where security concentrates.
Political-risk instability indices
Country and region instability regime states with confidence, at coarse granularity, country by month, licensed as a signal-feed family: priors for insurers and reinsurers setting political-risk and trade-credit premiums. A hard granularity floor holds: never sub-national, never event-level.A coarse grid of six countries by fourteen months, each cell shaded calm, elevated, or unstable. The granularity floor is the design: country by month, never a district, never an event.Each cell is one country in one month, shaded by how unstable its regime state is. The grid is deliberately no finer than that.
Chokepoint disruption probability
Disruption-probability monitoring on the physical arteries of trade: straits (Hormuz, Malacca, Suez, Panama), rail corridors, refining capacity, port throughput, as binary event predictions of the form throughput below a stated level for a stated number of days. It feeds directly into cascade theses. Sold to trading houses, shippers, corporates, sovereigns.A strait throughput series dips below a stated level long enough to meet a written event definition, with the qualifying span shaded. Below, a rolling ensemble probability of that event rises ahead of it.The top line is traffic through a strait; the shaded span is where it stays below the stated level long enough to count as an event. The blue line is the odds of that event, updating as evidence builds.
Policy-adjacent commissioned expansion
The commissioned-expansion mechanism opened to policy domains: a government or institution funds coverage of a region, corridor, or policy environment it cares about, within the families DeepField accepts and never within the refused ones.A flow diagram: a government request passes through a written acceptance charter, routed either to accepted families and funded permanent coverage, or to a refused branch listing influence operations, individuals, and crowd-level unrest.A request for new coverage flows through a written charter: accepted families get built and kept; the refused classes stay refused at any price.

F · THE META-PRODUCT

The verified record itself
The thing that makes the other eighteen sellable at a premium. Every prediction cryptographically committed before anyone sees it, suppressed and private ones included, so changing the books would show, even if we were the ones changing them. Calibration published bucket by bucket, and accuracy reported separately on the calls where DeepField disagreed with consensus, on a cohort sealed before outcomes were known. No prediction vendor on earth publishes these numbers. Every claim is checkable by someone who doesn’t trust us, and that is the moat: the record is itself the product.A commit-and-reveal timeline where the hash is on the books before the outcome, next to calibration bars in matching colors: committed confidence in rose against realized frequency in blue, bucket by bucket, with a computed mean gap of 2.8 percent.Left: a prediction’s fingerprint goes on the books before the outcome, so it can’t be rewritten quietly. Right: the confidence that was committed, in rose, next to what actually happened, in blue, bucket by bucket.

REFUSED WORK

There is work DeepField won’t take. The refusals are written down in advance, so they don’t get renegotiated when the money shows up: no influence operations, no message or persuasion testing, no unrest prediction sharp enough to point at a crowd, nothing about a named person or a small group, no election horse-race for its own sake. Some of this pays well. The answer stays no. The unit of study is the crowd, never the person, and the record is worth more than the revenue.

Each application is the same instrument pointed at a different question. The record is what makes them worth anything.