The $1.5M SBIR Where Your Clearance Is a Scored Technical Criterion. NGA's DV033 Locks Out Almost Every AI Company in America.

There is a category of federal solicitation that looks like an open competition and functions like an invitation list. The National Geospatial-Intelligence Agency just posted a clean example of it, and the mechanism is worth understanding even if you never bid on this particular topic — because the same structure is spreading across intelligence-community SBIR.

OSW26BZ06-DV033, "Agentic AI Based Cognitive Radar for GEOINT Mission," is a Direct-to-Phase-II-only SBIR topic under the NGA 2026 SBIR Broad Agency Announcement, Release 6. It opened September 23, 2026 and closes October 21, 2026 at 12:00 PM ET on the Defense SBIR/STTR Innovation Portal. The award is up to $1,500,000 over a maximum 24-month period of performance, plus up to $50,000 in Technical and Business Assistance.

By SBIR standards that is a substantial award, and the technical problem is genuinely interesting. Neither fact will matter to most readers, because of what sits between the topic and the money.

The security posture is not a compliance annex. It is scored.

Standard Department of War SBIR evaluation criteria apply to DV033 — plus one more: controlled-work capability, meaning your demonstrated ability to perform CUI and classified work, including your security posture, as a scored technical factor.

That single addition changes the genre of the document. In a normal SBIR, clearances are an administrative hurdle you clear after selection. Here, your facility clearance level, your personnel's clearance status, and your CAGE code and SPRS score are inputs to the technical score itself. A brilliant cognitive radar architecture from a firm with no facility clearance does not lose on a technicality at the end. It loses on the scoresheet.

The specific requirements:

• Current Top Secret clearances with SCI eligibility for all contractor personnel performing classified work. Not clearable. Current.
• Polygraph within 90 days. Personnel who have not completed polygraph testing in the last five years must immediately schedule an examination and complete the process — in no more than three test sessions — within 90 days.
• An NGA-accredited SCIF or approved secure facility with formal agreements in place, disclosed with facility clearance level, safeguarding level, CAGE code, and physical addresses.
• A disclosed SPRS score and a stated NIST SP 800-171 compliance approach for protecting CUI.
• No foreign nationals, and — this is the part teams misread — the prohibition extends to unclassified work on the contract. That is stricter than standard ITAR practice. Prior written COR approval is mandatory.

For an AI company, that last item is often the decisive one. The talent market for machine learning research in the United States runs heavily on visa-holding researchers. A firm whose core modeling team includes foreign nationals cannot simply partition them onto the unclassified workstream, because on this contract there is no unclassified workstream they are permitted to touch.

Your Key Personnel section must list the clearance level held by each named person. Not "clearance in process." Held.

The feasibility rule that will disqualify the best-qualified bidders

Direct-to-Phase-II topics require proof that you have already done the Phase I work on your own. DV033's version of that requirement contains a trap.

You must demonstrate a developed concept for an agentic AI cognitive radar with subsystem architecture detail, and prove the core capability — detecting interference signals and mitigating them by transmitting different waveforms — against quantitative targets. And then the constraint: the feasibility documentation cannot be based upon, or logically extend from, any prior or ongoing federally funded SBIR or STTR work.

Cognitive radar in the United States has been substantially funded through SBIR and STTR for over a decade. The firms most likely to have a working adaptive waveform prototype are, disproportionately, firms that built it on SBIR money. The rule does not care that the work is excellent. It cares about provenance.

Audit your intellectual property lineage before you write a page. Ask, for each figure and dataset you intend to put in the feasibility volume, which contract paid for the work that produced it — and be able to answer with internal R&D, commercial revenue, or non-SBIR government funding. If the honest answer is a prior SBIR, that evidence cannot carry your feasibility case, no matter how good it is.

What NGA actually wants built

The objective is an agentic AI cognitive radar that automatically selects radar parameters, waveforms, and exploitation methods for resilient GEOINT applications. Concretely: a radar that senses its own electromagnetic environment and decides — without a human in the loop — which frequency, which bandwidth, and which pulse repetition frequency to transmit.

Phase II is simulation-only. You prototype in modeling and simulation, develop the parameter-selection software, and analyze performance in a test environment. Hardware is Phase III work. Anyone proposing a build-and-fly program has misread the scope.

The stated performance targets are where proposers should slow down. NGA asks for detection and classification to improve from below 70% to 95% after mitigation, and for SINR to improve from below 70% to 95%. The second metric is non-standard — signal-to-interference-plus-noise ratio is conventionally expressed in decibels, not as a percentage. Submit a DSIP Topic Q&A question before that window closes on October 7, 2026, and in the meantime report both formulations in your proposal so your claim is legible under either reading.

NGA also cites three foundational works it expects you to engage: Haykin's 2010 paper on cognition-enabled radar, Guerci's Cognitive Radar: The Knowledge-Aided Fully Adaptive Approach, and the Bell et al. 2015 software-defined radar demonstration. Read that citation list as an instruction. The evaluators want your agentic-AI contribution positioned explicitly against that lineage — what modern learned policies do that knowledge-aided adaptive methods did not.

The proposal structure is unusually generous, which is a hint

The technical volume runs 40 pages — 20 for feasibility justification and 20 for the Phase II technical proposal. That is roughly four times the typical SBIR allocation, and pages over the limit are not evaluated.

An unusual page budget is a statement about what the evaluators expect to read. Twenty pages of feasibility is room for full experimental conditions, architecture diagrams, interference scenarios, and before-and-after performance data. Teams that submit eight pages of feasibility because that is what they are used to writing will be compared against teams that submitted twenty.

Three more structural rules worth flagging:

No option periods. The instruction is explicit that the offeror shall not propose them. Design a single 24-month effort with the full scope committed.

Percentage of Work is non-negotiable. NGA states it will not accept any deviation to the PoW requirements. Build your PoW allocation before you assemble the team, not after — and remember that subcontractors inherit every security and DFARS flow-down obligation you carry.

Section 508 accessibility belongs in your commercialization strategy. All information and communication technology must be accessible per WCAG 2.1 or a justified equivalent. This is the kind of requirement that gets skipped in a technical-heavy proposal and then reads as carelessness.

Note also that the Company Commercialization Report is required but not evaluated by NGA — so the commercialization argument has to live inside the technical volume, not in the form. And information control is tight: contractor personnel may not release unclassified information about any part of the contract in any medium without prior written COR approval, and use of the NGA name and seal is restricted. Plan a commercialization narrative that does not depend on publicizing the customer.

The timeline, and the sibling topic

• Opens: September 23, 2026
• DSIP Topic Q&A closes: October 7, 2026
• Proposals due: October 21, 2026
• Selection notification: approximately January 19, 2027 (within 90 days)
• Contract award: approximately mid-April 2027 (within 180 days)

Administrative questions go to sbir@nga.mil; technical questions go through DSIP Q&A.

DV033 has a sibling on the same release: OSW26BZ06-DV034, seeking a vision language model that fuses multimodal AI to detect and classify targets in synthetic aperture radar imagery from sensor-independent complex data — large language models joined to vision encoders for automated SAR interpretation. A team with SAR expertise and the right security posture should read both before choosing, because the clearance investment amortizes across them.

The strategic read

The honest framing of DV033 is that NGA is not running a competition for the best cognitive radar idea. It is running a competition among organizations that have already paid the price of admission to classified work — and using an SBIR topic to find which of them has non-SBIR-funded adaptive waveform technology sitting on a shelf.

If you are inside that set, the economics are excellent: $1.5 million for a 24-month simulation program, in a mission area with obvious Phase III paths, evaluated partly on infrastructure you already own.

If you are outside it, the useful conclusion is not "try anyway." Clearances cannot be acquired in 32 days, and a proposal that leaves the security section aspirational will be scored on that aspiration. The realistic move is to build a relationship with a cleared prime now, note that the SBIR/STTR reauthorization signed in April 2026 expanded national-security and foreign-ties vetting across the entire program — meaning this vetting posture is becoming more common, not less — and target the unclassified topics on the same release. Department of War Release 6 carries 44 topics across 33 SBIR and 11 STTR, all closing October 21, and most of them do not ask what your SPRS score is.