DARPA's FALCON Topic Skips Phase I Entirely: $1.5M Direct-to-Phase-II for Fusing LLMs With Classical ML, Closing August 19
July 8, 2026 · 6 min read
Granted Research Team · Editorial policy
Most defense small-business topics ask you to prove an idea from scratch. DARPA's FALCON topic — DPA26BZ04-DV016, "Fusion of Abstract Learning and Context-Optimized Neural-methods" — assumes you already have. It is a Direct-to-Phase-II (DP2) solicitation, meaning there is no Phase I feasibility study, no $150K warm-up, no six-month runway to sketch an approach. DARPA is putting $1.5 million on the table for teams that can walk in with a credible architecture and leave with a working demonstration. The topic opened July 22 and closes August 19, 2026 — a roughly four-week window that will feel much shorter to anyone who has to assemble a DP2 package from a standing start.
The technical premise is one of the more honest problem statements DARPA has published this cycle. Large language models are general and context-aware but statistically loose. Classical machine learning is precise on structured, tabular data but blind to context. FALCON wants both in one workflow — the "statistical power of ML" fused with the "contextualization power of LLMs" — to enable interactive analysis of large-scale mixed data, whether that data lives in an enterprise warehouse or streams off a battlefield sensor network. If you have spent the last two years watching teams bolt an LLM onto a database and call it analytics, FALCON is DARPA asking for the version that actually holds up under statistical scrutiny.
The numbers that define the topic
- Topic ID: DPA26BZ04-DV016
- Structure: Direct-to-Phase-II (no Phase I)
- Award ceiling: $1,500,000
- Publication date: July 1, 2026
- Open date: July 22, 2026
- Close date: August 19, 2026 (12:00 PM ET)
- Eligibility: For-profit, U.S.-owned small business; 500 employees maximum including affiliates
- Prerequisite: Demonstrated prior research in emerging ML methods within the past three years
That last line is the real gate. Because there is no Phase I, DARPA replaces the feasibility study with an eligibility bar: you must document qualifying prior work. In practice, a DP2 requires you to show that a Phase-I-equivalent effort already happened — under other funding, internal R&D, or published research — and that it produced the technical foundation you now propose to mature. A team that cannot evidence three years of relevant ML research is not merely disadvantaged here; it is ineligible.
Why Direct-to-Phase-II changes the field
The DP2 structure is not a convenience — it is a filter, and it quietly reshapes who can realistically compete.
It rewards existing capability over promise. Traditional Phase I is where a scrappy team with a good idea earns the right to keep going. FALCON removes that door. The teams that win DP2 topics are almost always ones that were already building in the space and treat the award as fuel for a demonstration they were going to attempt anyway. If FALCON's problem statement reads like a description of your last 18 months of work, you are the intended applicant. If it reads like a pivot, the timeline is brutal.
It compresses the proposal calendar. A four-week window on a $1.5M DP2 is not enough time to invent an architecture, recruit a principal investigator, and register in the Defense SBIR/STTR Innovation Portal (DSIP) if you are starting cold. It is enough time to package existing results into a compelling Phase II plan. The registration overhead alone — SAM.gov, SBIR Company Registry, DSIP — can consume a week for the unregistered. Read the entitlements before the technical writing.
It raises the bar on demonstrated milestones. DARPA has published an aggressive base-year schedule (more on this below). DP2 reviewers expect you to hit month-6 functionality and month-11 enterprise-scale demonstrations. That cadence is only survivable if your codebase already exists in some form.
The base-year milestones tell you what "done" looks like
DARPA laid out the Phase II deliverable schedule explicitly, and it is the single most useful planning artifact in the topic:
| Month | Deliverable |
|---|---|
| 1 | Technical report on selected ML methods |
| 3 | Quarterly progress presentation |
| 6 | Initial functionality demonstration |
| 11 | Enterprise-scale demonstration across ≥2 datasets |
| 12 | Final report and documented software |
An optional year adds month-15 and month-18 checkpoints. Two things jump out. First, month 1 is a survey deliverable — DARPA wants a defensible rationale for which emerging ML methods you selected and why, not just an implementation. Second, month 11 demands generalization across at least two datasets from different domains. A model that shines on one benchmark and collapses on another fails the topic's core promise. Build your evaluation harness for cross-domain generalization from day one, because that is the milestone the whole award is graded against.
The hallucination requirement is not boilerplate
Buried in the technical requirements is a line that should shape your entire architecture: proposals must demonstrate "methods to mitigate possible hallucination in the workflow." DARPA is not asking for a disclaimer. In a system where an LLM contextualizes statistical output for an analyst — potentially an intelligence analyst or a commander — a confidently wrong interpretation is worse than no interpretation. Your proposal needs a concrete, testable hallucination-mitigation strategy: retrieval grounding, statistical cross-checks that flag when the LLM's narrative diverges from the ML layer's numbers, confidence calibration, human-in-the-loop gates, or some combination. Teams that treat this as a compliance sentence will lose to teams that make it an architectural centerpiece. DARPA's stated preference for open-source LLMs compounds this: you are expected to own and inspect the model layer, not treat a closed API as a black box you cannot audit.
Who should actually apply
The ideal FALCON applicant is a small AI/ML firm — under 500 employees — with a demonstrable three-year track record in methods for structured-plus-unstructured data fusion: think teams already working in retrieval-augmented analytics, tabular foundation models, neuro-symbolic reasoning, or LLM-over-database interfaces. If you have shipped or published in any of those areas, FALCON is a rare chance to convert existing IP into a $1.5M non-dilutive award without the Phase I gauntlet.
If you are earlier-stage, FALCON is likely the wrong door this cycle — but it is a signal. DARPA's Information Processing Techniques office is telegraphing sustained interest in trustworthy LLM-plus-ML fusion, and topics in this family tend to recur. The move is to start building and documenting the qualifying research now, so the next DP2 in this lineage finds you eligible.
The four-week playbook
- Confirm eligibility first. Can you document three years of relevant ML research? If not, stop — you are not eligible, and no amount of proposal polish changes that.
- Clear registrations in week one. SAM.gov, SBIR Company Registry, and DSIP accounts must be active before you can submit. This kills more DP2 bids than weak science does.
- Anchor on month 11. Write the technical volume backward from the enterprise-scale, two-dataset demonstration. Everything else is scaffolding for that milestone.
- Make hallucination mitigation a section, not a sentence. Name the mechanism, name the test, name the metric.
- Pick your two domains early. Generalization is graded; choose datasets that are genuinely different so a passing demo actually proves the claim.
For the broader picture of DARPA's summer SBIR activity, see our coverage of the June 3 DSO and BTO drop. FALCON is narrower than those topics but structurally harder: it does not ask whether your idea could work. It asks you to prove it already does — and gives you until August 19 to make the case.
Granted tracks live SBIR/STTR topics across DARPA, NSF, and the Department of Defense. Search current deep-tech opportunities and past awardees to benchmark your proposal at grantedai.com.