DARPA's July 22 SBIR Release: Geopolitical Forecasting, LLM-Symbolic Fusion, and Radiation-Hard Memory — Three Topics That Close August 19

DARPA's Small Business Innovation Research topics are worth reading even if you never intend to apply, because they are one of the clearest public windows into what the agency thinks it will need in five years. On July 1, 2026, DARPA pre-released three FY26 topics under Release 4 of the 2026 SBIR Broad Agency Announcement; they opened for submission on July 22 and close on August 19, 2026 at 12:00 p.m. ET. That is a four-week live window — tight even by DARPA's standards — on three topics that, taken together, sketch the agency's current obsessions: anticipating events before they happen, making large AI models reason rather than just pattern-match, and building electronics that survive where nothing else does.

Here is what each topic funds, why it was written, and how a small team should think about a deadline this short.

Art of Novel Signals: forecasting the future from a knowledge graph

Topic DPA26BZ04-DV015, out of DARPA's Defense Sciences Office, asks performers to "develop and demonstrate a predictive/forecasting model that leverages Temporal Knowledge Graph Forecasting using In-Context Learning."

Unpack the jargon and the ambition is striking. A temporal knowledge graph represents entities — people, organizations, places, events — and the relationships among them, with each relationship stamped in time. "Forecasting" over such a graph means predicting which relationships will appear next: which actors will align, which events will follow which, how a situation will evolve. "In-context learning" means doing this the way a modern large model does — by conditioning on examples supplied at inference time rather than retraining for every new scenario.

Put plainly, DARPA wants a high-confidence geopolitical forecasting engine — a system that ingests multilingual, multimodal reporting and outputs calibrated predictions about what happens next, with the reasoning traceable through the graph rather than buried in a black box. The word "confidence" is doing real work in the topic title: the hard part is not producing a forecast, it is producing a calibrated one, where the model knows and communicates how sure it is. For a small team, the winning angle is rarely "we built a bigger model." It is a defensible method for calibration, provenance, and traceability — the properties that separate an intelligence tool an analyst will trust from a demo an analyst will ignore.

FALCON: making large models actually reason

Topic DPA26BZ04-DV016, from the Information Processing Techniques office, is FALCON — Fusion of Abstract Learning and Context-Optimized Neural-methods. It combines machine-learning methods with large language models for "interactive statistical analysis of large-scale data," explicitly across both enterprise and battlefield contexts.

FALCON is DARPA's bet on the current frontier problem in applied AI: large language models are fluent but unreliable at rigorous quantitative reasoning. They will describe a statistical analysis convincingly and get the arithmetic wrong. The topic's framing — fusing abstract (symbolic, statistical) methods with neural ones — is a direct response. The goal is a system where an analyst can interrogate a large dataset conversationally and get answers that are actually correct, because the LLM orchestrates rigorous statistical tools rather than hallucinating their outputs.

That "enterprise and battlefield" phrase matters for positioning. A team with a credible dual-use story — the same reasoning layer serving a commercial analytics customer and a defense user — fits the topic's intent better than a pure-defense or pure-commercial pitch. This is the kind of topic where a small company with real IP in neuro-symbolic integration or verified LLM tool-use can outcompete a larger firm bolting an LLM onto an existing dashboard.

Non-Volatile Memory for Extreme Environments: electronics that survive

The third topic, DPA26BZ04-DV017 (amended), comes from DARPA's cross-cutting "Multi X" effort and is the most physically demanding of the three. It asks performers to "develop and demonstrate a co-packaged temperature-hard (-250°C to +600°C) and radiation-tolerant NOR Flash memory system."

Consider that temperature range. -250°C is deep-space and cryogenic-instrument cold; +600°C is down-hole, hypersonic-skin, and jet-engine hot. Commercial flash memory is rated for a tiny fraction of that span and fails quickly under radiation. A single memory system that holds data across an 850-degree window and tolerates radiation is an enabling component for satellites, deep-space probes, hypersonic vehicles, nuclear-adjacent sensing, and downhole instrumentation — applications where you cannot swap the part and cannot afford a bit-flip.

This is not an AI topic; it is a materials-and-packaging topic, and it rewards a very different kind of small business — one with real fab or advanced-packaging capability and a plausible path to a manufacturable device, not just a lab demonstration. The word "co-packaged" is a hint: DARPA wants an integrated system, not an isolated memory cell that works in a probe station and nowhere else.

Why a four-week window is the real story

Across all three topics, the operative constraint is time. A topic that pre-releases July 1, opens July 22, and closes August 19 gives you roughly four weeks of live submission window — and if you first learned about it when it opened, you have effectively lost the three-week pre-release head start that serious competitors used to line up letters of support, confirm teaming arrangements, and draft.

That has three practical implications for any small business eyeing DARPA SBIR:

• The pre-release period is where proposals are won. DARPA publishes topics ahead of the open date specifically so you can contact the Topic Author with questions during the pre-release window — a channel that closes once the topic opens. Teams that engage early write to the author's actual intent; teams that show up on open day are guessing.
• Direct-to-Phase-II is the leverage point. DARPA routinely offers Direct-to-Phase-II paths for teams that can document prior Phase-I-equivalent work. The public topic page does not spell out award amounts or D2P2 eligibility for these three — that lives in the solicitation documents and instructions — so read the topic-specific instructions before assuming a standard Phase I. If you already have relevant results, the D2P2 route can compress your timeline dramatically.
• Four weeks favors the prepared, not the fast. You cannot originate credible DARPA-grade technical volumes, teaming, and commercialization plans from scratch in a month. The teams that win a compressed window are the ones whose technology already existed and who simply mapped it onto the topic. If the fit is real, move now; if you would be inventing the approach to fit the topic, this is not your cycle.

Reading the tea leaves

Step back and the three topics rhyme. Two of them — Art of Novel Signals and FALCON — are about making AI trustworthy for consequential decisions: forecasting with calibrated confidence, and reasoning with verifiable correctness. The third is about the hardware substrate that lets advanced systems operate where the mission actually is. This is a coherent picture of an agency that has moved past "can we build a large model" to "can we build systems whose outputs an operator will stake a decision on, running on electronics that survive the environment."

For the deep-tech small-business ecosystem, DARPA topics are also a forecasting signal in their own right. What DARPA funds at the SBIR stage today tends to define the requirements that flow into larger defense programs — and eventually commercial markets — years later. Even if none of these three is your topic, the pattern is worth filing away.

All three close August 19, 2026 at 12:00 p.m. ET. Submission runs through the DoD SBIR/STTR portal under the 2026 SBIR BAA. For the broader agency-by-agency picture, see our 2026 SBIR/STTR deadlines calendar and the earlier DARPA BTO topic analysis. If one of these fits work you have already done, the clock is the only thing standing between you and a proposal.