DARPA Wants to Predict Conflict Two Weeks Out From Radio Static: Inside the $2M 'Art of Novel Signals' Direct-to-Phase-II SBIR

Buried in DARPA's July 2026 Release 4 SBIR topics is one of the most ambitious — and most demanding — forecasting problems the agency has put in front of small businesses in years. Art of Novel Signals: Predicting and Forecasting with High Confidence (topic DPA26BZ04-DV015) asks companies to build a system that can predict geopolitical events two weeks in advance at 90% precision, using data the open web never captured: multilingual radio broadcasts, local-language news, and community reporting from the world's most under-observed regions. The award is $2 million, the deadline is August 19, 2026 at 12:00 PM ET, and — critically — it is Direct-to-Phase-II (DP2) only. There is no Phase I to ease into. You either arrive with a working forecasting pipeline or you do not apply.

This topic pairs with the machine-learning-and-LLM topic FALCON, which shares the August 19 close date, but Art of Novel Signals is a fundamentally different bet. FALCON is about reasoning over structured and unstructured data. Art of Novel Signals is about collecting a signal that almost nobody else has and turning it into early warning. Understanding that distinction is the first step to deciding whether your company should chase it.

The problem: the open web is running out of new information

The topic opens from a premise that has become conventional wisdom inside frontier AI labs and is now driving defense R&D: high-quality, open text is effectively exhausted. Models trained on the same crawled corpora are hitting diminishing returns, and for the specific task of forecasting instability in places like Central and Southeast Asia, East and Northeast Africa, and South America, the open web was never a good source to begin with. Conflict often builds in languages, on platforms, and through channels that global data pipelines simply do not index.

DARPA's insight is that multilingual radio, local news, and community reporting from data-sparse regions represent a large, almost entirely untapped reservoir of high-value, time-sensitive information — signal that "was never in the training distribution before" and that synthetic data cannot manufacture. The whole topic is organized around capturing that signal and converting it into predictions. Four technical components define the work:

1. A radio data collection engine — capturing multilingual broadcasts and local news via direct streaming and software-defined radio receivers.
2. Automatic speech recognition (ASR) for low-resource languages — fine-tuning state-of-the-art models to reach usable accuracy on oral, under-documented languages at roughly 50 hours of transcribed audio per language.
3. A synthetic-data strategy — balancing real and synthetic data so that adding language coverage stays affordable.
4. Forecasting integration — converting transcribed audio into temporal knowledge graphs and running an in-context-learning forecasting pipeline over them for geopolitical early warning.

That fourth piece is the intellectual core. Rather than retraining an embedding model every time the world changes, the winning system must run temporal knowledge graph forecasting using in-context learning — updating its predictions from new events without re-training. The precision target is unusually specific and unusually high: extend the useful forecasting horizon from today's roughly five-day window out to fourteen days while pushing precision to at least 90%, and cut false-positive rates by roughly half along the way.

Why "Direct-to-Phase-II only" changes who can win

Most SBIR topics run Phase I first — a small feasibility award (typically $150K–$300K over six months) that lets a company prove a concept before a larger Phase II build. Art of Novel Signals skips that entirely. DP2 means DARPA is only interested in teams that can already document feasibility. To qualify, a proposer must show, up front:

• A working in-context-learning forecasting pipeline that operates over temporal knowledge graphs without retraining embeddings.
• A baseline precision of at least 80%, validated independently — not self-reported.
• Demonstrated ingestion of multilingual, multimodal open-source signals, including radio.
• Generalization to events and regions the system has not seen.
• The ability to deploy across three strategic regions.

Read that list carefully: it is not a research proposal, it is an audit. DARPA is asking companies to prove they have already built most of the hard part. The 80% baseline is the gate. If you cannot independently validate 80% precision on real multilingual signal today, this topic is not addressable in the six-week window, and no proposal-writing effort will close that gap.

For companies that do clear the bar, the Phase II structure is an 18-month base period plus a 6-month, $500,000 option, with a milestone cadence that reveals exactly what DARPA will be watching:

The option period pushes toward an operational pilot with real user feedback and rapid language onboarding for contingencies — the tell that DARPA wants a transition, not a paper.

Eligibility and the realistic candidate profile

The formal eligibility is standard SBIR: a for-profit, U.S.-owned and controlled company with 500 or fewer employees, counting affiliates. But the practical eligibility is far narrower than the legal one. The teams that can credibly win Art of Novel Signals cluster into a few profiles:

• Signals-intelligence and OSINT startups that already run collection infrastructure and have solved the low-resource ASR problem for at least a handful of languages.
• Applied-AI firms with a temporal knowledge graph or event-forecasting product — political-risk monitors, supply-chain-disruption predictors, humanitarian early-warning shops — that can retarget their pipeline onto radio.
• Academic spinouts with published forecasting benchmarks who have already been operating over event graphs.

If your company is none of these but the space is strategically interesting, the correct move is not to write a long-shot DP2 proposal in six weeks. It is to build toward the next window: stand up a radio-collection prototype, publish an independently validated precision number, and be ready when DARPA re-releases an adjacent topic. DARPA pre-releases SBIR topics on the first Wednesday of every month, and forecasting is a durable agency interest — this will not be the last bite.

The commercial thesis beyond defense

DARPA is explicit that Phase III transition can be funded from private or non-SBIR government sources, and the dual-use case here is strong. The same engine that gives a combatant commander two weeks of warning before regional instability is directly valuable for political-risk monitoring, supply-chain resilience, and humanitarian operations — three markets that already pay for exactly this kind of foresight and are chronically starved of ground-truth from data-sparse regions. A company that clears the 80% gate is not just winning $2 million; it is building a defensible data moat around a signal source competitors cannot easily replicate, because the barrier is physical collection and low-resource ASR, not model architecture.

Bottom line

Art of Novel Signals is a topic where the proposal is almost secondary to the prior work. DARPA has written the qualification criteria as a checklist of things you must have already accomplished — an 80% independently validated precision baseline over real multilingual radio signal being the hard floor. For the small set of SBIR-eligible firms that have quietly been building geopolitical forecasting over event graphs, the August 19, 2026 deadline is a rare chance to convert a research capability into a funded, transition-oriented $2 million program with a clear commercial afterlife. For everyone else, it is a detailed public blueprint of where DARPA thinks the next generation of early warning is going — and a checklist worth building against for the next release.

Tracking DARPA and defense SBIR deadlines? Granted monitors federal forecasting and AI topics across DARPA, the services, and the civilian agencies, and matches them to your company's capabilities. See the full SBIR and STTR 2026 deadline calendar for every agency's windows.