NGA SBIR OSW26BZ06-DV034: Vision Language Model for SAR Target Search and Classification from SICD
Quick Answer
OSW26BZ06-DV034 is a Direct to Phase II SBIR topic from the National Geospatial-Intelligence Agency under the 2026 SBIR Broad Agency Announcement, Release 6. Phase I proposals are not accepted. NGA wants a vision language model that reads synthetic aperture radar imagery in Sensor Independent Complex Data format, finds the targets in it, and classifies them. The award must not exceed $1,500,000 for up to a 24-month period of performance, and the technical volume runs to 40 pages. The topic opens September 23, 2026 and closes October 21, 2026 through the Defense SBIR/STTR Innovation Portal.
The feasibility gate here is unusual and worth reading twice, because it has two halves and one of them is about people rather than technology. You must demonstrate successful classification with greater than 95 percent accuracy of single SAR image chips in SICD format. And you must demonstrate that the performer is very well qualified, with PhD or research level given as the example, and has demonstrated experience classifying complex SAR imagery. NGA spells out what qualification means: a scientist's extensive knowledge of SAR signal and image processing, SAR operations from space, and deep learning understanding and implementation.
That is a credentialing requirement embedded in a technical gate. A team of excellent machine learning engineers without deep SAR phenomenology expertise does not satisfy it, and neither does a SAR expert without deep learning implementation experience. NGA is asking for both in the same organization.
Then there is the security environment, which governs every NGA SBIR. This work is classified. All contractor personnel shall possess a current Top Secret personnel security clearance and be eligible for favorable NGA adjudication for Sensitive Compartmented Information access. Contractors are subject to counterintelligence-scope polygraph. Your Key Personnel section must list each person's clearance level. And NGA scores your ability to perform controlled work as part of the technical evaluation.
Topic At a Glance
Topic number: OSW26BZ06-DV034
Title: Vision Language Model (VLM) for Synthetic Aperture Radar (SAR) target search and classification from Sensor Independent Complex Data (SICD)
Agency: National Geospatial-Intelligence Agency (NGA), a Department of War combat support agency and a member of the U.S. Intelligence Community
Solicitation: NGA 2026 SBIR Broad Agency Announcement, Release 6, Proposal Submission Instructions
Program type: Direct to Phase II only. This topic is accepting Direct to Phase II proposals only
Award: must not exceed $1,500,000
Period of performance: up to 24 months
Options: the offeror shall not propose option periods
Technical volume: 40 pages maximum, consisting of Part 1 Phase I Justification at 20 pages maximum and Part 2 Phase II Technical Proposal at 20 pages maximum. Pages in excess will not be considered. Number all pages consecutively
Component Technology Priority Areas: Integrated Sensing and Cyber, Trusted AI and Autonomy, Integrated Network Systems-of-Systems
Projected CMMC level requirement: Level 2 (Self)
Export control status: no topic-level ITAR or EAR restriction paragraph appears on this topic. Note that the work itself is classified, which imposes far stricter personnel and facility requirements than an ITAR notice would
Feasibility accuracy bar: greater than 95 percent classification accuracy on single SAR image chips in SICD format
Feasibility qualification bar: the performer must be very well qualified, PhD or research level, with demonstrated experience classifying complex SAR imagery, requiring extensive knowledge of SAR signal and image processing, SAR operations from space, and deep learning understanding and implementation
Phase II scope: input is SAR imagery with multiple targets, output is location and classification of the targets. Government-provided SAR SICD data or your own or commercial data may be used
Clearance environment: all contractor personnel shall possess a current Top Secret personnel security clearance and be eligible for favorable NGA adjudication for SCI access. Contractors are subject to counterintelligence-scope polygraph as requested
Added evaluation criterion: NGA will evaluate a vendor's ability to perform controlled work, meaning CUI and classified, as part of the technical evaluation score
Section 508: the Commercialization Strategy shall address Section 508 compliance per NGA Instruction 8400.4 and Section 508 of the Rehabilitation Act, with an outline of how compliance will be achieved
Company Commercialization Report: information contained in the CCR will not be considered by NGA during proposal evaluations
Percentage of Work: NGA will not accept any deviation to the POW requirements
Technical and Business Assistance: up to $50,000 per Phase II project, in addition to the cost ceiling and not subject to profit or fee, using the mandatory SBIR/STTR TABA Request Form in Volume 5
Contract type: NGA typically provides a firm fixed price contract within 180 days of the proposal due date, at the discretion of the Contracting Officer
Topic open date: September 23, 2026
Proposal deadline: October 21, 2026
Submission portal: DSIP at dodsbirsttr.mil
Keywords: synthetic aperture radar, Sensor Independent Complex Data, metadata, single look complex
The Feasibility Bar, Which Has Two Halves
This topic is accepting Direct to Phase II proposals only. To qualify for a Phase II award, proposers must submit Feasibility Documentation demonstrating two things. The purpose of this documentation, in NGA's words, is to prove that the underlying technology is mature and ready to transition immediately into Phase II requirements.
Half one: the accuracy result
Successful classification, greater than 95 percent accuracy, of single SAR image chips in SICD format.
Three parts of that sentence matter. Greater than 95 percent accuracy is a specific, measurable number. Single SAR image chips means one target per chip, which is a classification problem rather than a detection or search problem. And in SICD format means complex data, not detected magnitude imagery.
That last point is the one most likely to trip up an otherwise strong applicant. SICD is the NGA standard for sensor independent complex data, meaning the imagery retains phase as well as amplitude, in single look complex form. Most publicly available SAR classification work, and most benchmark results, use detected magnitude chips. If your 95 percent result is on magnitude imagery, it does not literally satisfy the stated gate, and the honest move is to reproduce the result on SICD data or to state plainly what you have and what the gap is.
Report the accuracy properly: the class set, the number of classes, the number of chips, the train and test split, whether the split is by scene or by chip, and the confusion matrix. A 95 percent number on a small closed class set with a chip-level random split is a much weaker claim than the same number on a scene-disjoint split, and a reviewer who works in this field knows the difference.
Half two: the qualification of the performer
The performer must be very well qualified, with PhD and research level given as examples, and must have demonstrated experience classifying complex SAR imagery.
NGA then defines what qualification requires: a scientist's extensive knowledge of SAR signal and image processing, SAR operations from space, and deep learning, meaning machine learning, understanding and implementation.
This is a personnel requirement inside the feasibility gate, and it is unusual. Three distinct competencies are named and they rarely coexist in one person or one small team.
SAR signal and image processing means the phenomenology: how the image is formed, what phase carries, speckle, sidelobes, layover and shadow, polarimetry, the difference between slant plane and ground plane, and why a SAR image of a target looks nothing like an optical image of it.
SAR operations from space means the collection side: orbital geometry, look angles, incidence angles, revisit, squint, the collection modes and their resolution and swath tradeoffs, and the metadata that describes all of it. The SICD keyword list includes metadata for a reason.
Deep learning understanding and implementation means you build and train models, not just apply them.
Address all three explicitly in your Key Personnel and Related Work sections, with named individuals, degrees, publications, and specific prior work on complex SAR. If one of the three is weaker in your team, say how you cover it, whether through a named consultant, a subcontract, or a hire, and remember that the Percentage of Work rules and the security requirements both apply to anyone you bring in.
The restriction that applies to both halves
NGA's Direct to Phase II guidelines state that feasibility documentation cannot be based upon or logically extend from any prior or ongoing federally funded SBIR or STTR work. Work submitted within the feasibility documentation must have been substantially performed by the proposer or the principal investigator. If technology in the feasibility documentation is subject to intellectual property, the proposer must either own the IP or must have obtained license rights to such technology prior to proposal submission.
Audit the provenance of your 95 percent result. SAR automatic target recognition work in the United States has been substantially funded through SBIR and STTR programs, so this is a live risk for exactly the companies best positioned to bid. Internally funded work, privately funded work, non-SBIR government contract work, and your own published academic work under other funding are cleaner ground.
If the proposer fails to demonstrate technical merit and feasibility equivalent to the Phase I level as described in the associated topic, the related Phase II proposal will not be evaluated.
Twenty pages of feasibility is a lot, so use it
NGA allows twenty pages for feasibility documentation, four times what most components in this cycle allow. That is room for the full experimental description, the confusion matrix, the data provenance discussion, example chips with your model's outputs, the SICD handling detail, and complete personnel qualification evidence including publication lists. A five-page feasibility section submitted into a twenty-page allowance reads as thin evidence.
What NGA Is Actually Looking For
The objective
Develop VLM based SAR imagery search and classification from sensor independent complex data.
Note "search" in the title and objective. That is a broader task than classification, and it is where Phase II goes.
The concept, in NGA's words
Vision language models combine multimodal data and imagery, with video, SAR, and infrared imagery named, and generative AI models to understand and process video, image, and text. A VLM integrates large language models with a vision encoder that allows the LLM to explain contents of an imagery.
The goal of this research is to analyze, meaning detect and classify, synthetic aperture radar imagery from sensor independent complex data.
Why a VLM rather than a classifier, which your proposal should answer
The topic does not explain why the model has to be a vision language model rather than a conventional detector and classifier, and that is arguably the central intellectual question of the proposal. Notice the phrase NGA chose: a vision encoder that allows the LLM to explain contents of an imagery.
Explanation is the plausible answer, and it aligns with the Trusted AI and Autonomy Component Technology Priority Area. A GEOINT analyst does not only want a bounding box and a label. They want to know what in the image supports that label, in language, so they can accept or reject it and write it into a product. A VLM can produce that. A classifier producing a softmax score cannot.
Language also enables open-vocabulary and query-driven search, which is what "target search" implies. Instead of a fixed class list baked in at training time, an analyst can ask for something described in words. That is a genuine capability change for GEOINT tasking and it is worth arguing for explicitly.
If you make that case well, you have answered the question the topic left implicit, and you have distinguished yourself from proposals that treat "VLM" as a fashionable label on a conventional detector.
The hard part, which is SAR-specific
Vision language models are trained on natural imagery and text. SAR imagery is not natural imagery. It is complex-valued, speckled, geometrically distorted by layover and shadow, radiometrically counterintuitive, and sensitive to look angle in ways that change a target's appearance completely. A vision encoder pretrained on photographs has learned features that do not transfer cleanly.
Your proposal should confront this directly. How do you adapt or replace the vision encoder for complex SAR input. Do you feed complex data to the encoder, or magnitude and phase as separate channels, or a learned representation. How do you handle the SICD metadata, which carries the collection geometry that determines what the image means. Where does your text supervision come from, given there is no large corpus of SAR imagery paired with expert descriptions.
That last question is the practical bottleneck and most proposals will gloss over it. Say where the language side of your training data comes from: synthetic captions generated from metadata and annotations, expert-written descriptions, a curriculum that bootstraps from labels, or something else.
The GEOINT context
NGA describes GEOINT as the exploitation and analysis of imagery and geospatial information to describe, assess, and visually depict physical features and geographically referenced activities on the Earth, consisting of imagery, imagery intelligence, and geospatial information. NGA manages the National System for Geospatial-Intelligence.
NGA Research supports the National System for Geospatial-Intelligence and the National Security Strategy in three broad areas: Foundational GEOINT, Advanced Phenomenologies, and Analytic Technologies. This topic sits in Analytic Technologies.
Framing matters. The customer is an imagery analyst producing intelligence, working under time pressure, accountable for what goes into a product. Speak to that workflow rather than to a benchmark leaderboard.
Phase II Scope
Develop a VLM based SAR SICD imagery classification system, where input contains SAR imagery with multiple targets and output should be location and classification of the targets. Use SAR SICD data provided by government or your own, or commercial, data.
The shift from feasibility to Phase II
This is the most important structural feature of the topic and it is easy to miss. Feasibility is single chips, one target each, classified with greater than 95 percent accuracy. Phase II is full imagery with multiple targets, and the output is both location and classification.
That is a jump from classification to detection plus classification, which in SAR is a substantially harder problem. A chip classifier operates on a pre-localized target. A search system has to find candidate targets across a large scene, in clutter, with false alarms as the dominant failure mode, at whatever resolution and geometry the collection provides. The class imbalance is severe: a scene is mostly not-target.
Your Phase II plan should be explicit about how you bridge that gap. Sliding window, region proposal, a detection head on the shared encoder, an attention mechanism over the full scene, or a two-stage architecture. And your performance metrics should shift accordingly, from classification accuracy to detection and classification metrics that account for false alarms, whether precision and recall, mean average precision, or a false-alarm-rate-versus-probability-of-detection curve.
State those metrics yourself, because the topic states only the feasibility accuracy number and does not specify Phase II performance targets. Proposing your own targets, justified against the scene sizes and target classes you address, is expected and it is a place to show judgment.
The data question
You may use SAR SICD data provided by the government or your own or commercial data. That flexibility is genuinely useful and it is worth planning around deliberately.
Government-provided data is likely to be the most mission-representative and the most likely to be controlled, which brings the classified environment and the security requirements into your data handling plan. Commercial SAR data is more accessible and there is now a meaningful commercial constellation supply, but you should confirm it can be delivered in or converted to SICD. Your own data, if you have it, has the cleanest provenance for the feasibility restriction discussed earlier.
Say which you will use, in what proportion, and how you handle the domain gap between them. If you train on commercial data and test on government data, that shift is a technical risk and naming it is a strength.
Note also the practical toolchain. The topic cites the sarpy package, NGA's own open-source Python library for reading and processing SICD and related NGA formats, and the Schwartzkopf paper on generic processing of SAR complex data using the SICD standard. Using sarpy, or explaining why you use something else, signals familiarity with the actual format rather than the abstraction.
What Phase III is
Transition the VLM based SAR SICD classification tool for government utilities. That is a short statement, and it means the Phase II deliverable should be a tool an NGA organization can adopt, not a research prototype. Interface, documentation, deployment model, and the Section 508 accessibility requirement all follow from that.
The Security Requirements, Which Govern Everything
This is the longest part of the NGA instructions and it substantially determines who can win. NGA has made it a scored evaluation factor, so treat it as technical content rather than a compliance appendix.
The added evaluation criterion
In addition to the DoW Program evaluation criteria, NGA will evaluate a vendor's ability to perform controlled work, meaning CUI and classified, as part of their technical evaluation score.
NGA may use Systems Engineering and Technical Assistance support to assist with the programmatics of executing the evaluation process, but will not use SETA support to evaluate proposals.
Personnel security
All Contractor personnel shall possess a current Top Secret Personnel Security Clearance and be eligible for favorable NGA adjudication for SCI access. The government will verify security clearances and SCI eligibility in accordance with SEAD 7 on reciprocity of background investigations and national security adjudications. If needed, NGA will sponsor SCI accesses at its sole discretion.
Contractor personnel performing Top Secret/Sensitive Compartmented Information work are required to have active TS/SCI clearances for access to TS/SCI facilities, when performing duties within TS/SCI environments, and for access to TS/SCI computer systems.
Contractors are subject to a counterintelligence-scope polygraph examination as requested by NGA. As a condition of employment and assignment, contractors who have not successfully completed polygraph testing within the last five years must immediately schedule a polygraph examination and must complete the process, in no more than three test sessions, within 90 days.
Cleared Contractor personnel will be enrolled into the Director of National Intelligence Continuous Evaluation System throughout the lifecycle of the contract while at NGA, per SEAD 6, and must self-report security issues, foreign contacts, and other applicable information per SEAD 3 and NGA Instruction 5205.3. Cleared contractors must also submit electronic fingerprints and enroll in NGA's Report of Arrest and Prosecution Background program.
Personnel designated as Tier 3 Privileged Users, or holding enhanced access through a Special Access Program or Controlled Access Program, will participate in NGA's annual Security Financial Disclosure Program.
Some personnel may need access to Special Access Program, Controlled Access Program, or Alternate Compensatory Control Measures information as required for NGA's sensitive program missions, and access will not be granted without written permission from the Contracting Officer's Representative.
Defense Counterintelligence and Security Agency remains the Cognizant Security Agency for individuals with collateral Secret or Top Secret clearances, with SEAD 3 reporting through the contractor's Facility Security Officer to DCSA via the Defense Information Security System.
What this means for a machine learning team
This is the provision most likely to reshape a strong AI company's plan for this topic. Deep learning talent is highly mobile and internationally sourced, and a Top Secret clearance with SCI eligibility plus a counterintelligence-scope polygraph is a substantial filter. Anyone not polygraphed in the last five years must complete the process in no more than three sessions within 90 days.
Your Key Personnel section must list the clearance level held by each of the personnel, because, as the instructions put it, we will be operating in a classified environment. On a topic whose feasibility gate already demands PhD-level SAR and deep learning expertise, the intersection of that expertise with TS/SCI eligibility is the real scarcity.
Foreign nationals
Foreign nationals are not permitted to perform unclassified work under the terms of this contract without prior written approval from the Contracting Officer or the COR. Should NGA identify the use of unauthorized personnel, the CO may direct the contractor, at its own expense, to remove and replace any unauthorized contractors, subcontractors, or other personnel, at NGA's discretion and without prejudice to other rights including termination for default.
Note the scope: the prohibition covers unclassified work, which is stricter than the case-by-case disclosure approach used on ITAR-restricted topics elsewhere in this cycle. For a deep learning team this is often the binding constraint, and it must be resolved before you propose rather than after award.
The unclassified path
If applicable, and after discussion with the COR and with written COR approval, uncleared Contractor personnel are authorized to work on this contract at the unclassified level, with access up to DoD Controlled Unclassified Information at the contractor site without the requirement of a security clearance.
Any Contractor personnel working with CUI must receive a favorable HSPD-12 adjudication for access of 60 days or less, or a favorable HSPD-12 Tier 1 adjudication for access longer than 60 days, prior to accessing CUI. NGA will sponsor the HSPD-12 and HSPD-12 Tier 1 background investigations as needed.
Uncleared personnel visiting NGA facilities may receive a visitor badge and be escorted as appropriate, returning the badge at the end of each visit day.
This is what makes a mixed team feasible: model development on commercial or your own data at the unclassified level with uncleared but HSPD-12 adjudicated staff, and cleared staff for anything touching government-provided controlled data or classified environments. Structure your data plan and your staffing plan together, because they determine each other.
Physical security and facilities
All classified work performed at a non-NGA facility must be approved by the COR. Any classified work at contractor sites must be performed in an NGA accredited Sensitive Compartmented Information Facility, an Other Government Agency SCIF, or an approved secure collateral space with a Memorandum of Agreement, Memorandum of Understanding, Joint Use Agreement, or Co-Use Agreement with NGA for this contract. Accreditation shall be at least commensurate with the safeguarding level required by the DD Form 254.
Accreditation follows the IC Technical Specifications for Construction and Management of Sensitive Compartmented Information Facilities Version 1.5.1, the IC Tech Specs for ICD and ICS 705. NGA conducts periodic inspections based on threat, facility modifications, program sensitivity, past security performance, or at least every five years. Technical Surveillance Countermeasures activities in NGA accredited SCIFs are conducted only by NGA TSCM teams.
Contractor personnel are forbidden from bringing prohibited or unauthorized items into any NGA installation or secure facility, including weapons, cell phones, cameras, two-way pagers, laptops, recording devices of any kind, flash drives, or any other removable media. Exceptions may be granted by NGA Security upon request with documentation shown prior to entry. Violations may subject the contractor and personnel to civil or criminal liability.
Your Facilities and Equipment section must identify how Controlled Unclassified Information will be protected in accordance with NIST SP 800-171 and provide your Supplier Performance Risk System score, and must identify your Facility Clearance Level, the safeguarding level for each facility, the CAGE code for each facility, and the associated physical addresses.
The SPRS score is a specific number a reviewer can check immediately, and it takes time to improve. Verify yours before you submit.
There is also a practical wrinkle worth noting for a machine learning program: the prohibition on laptops and removable media in secure facilities means your development workflow inside a SCIF looks nothing like your workflow outside one. Address how model development, training, and evaluation happen on the controlled side if your plan requires it.
Information security and cybersecurity
NGA has sole authority to determine whether and to what extent protected information is provided. Access to classified information is pursuant to the DD254. The contractor shall not access, download, print, or further disseminate classified information outside the execution of defined contract requirements without written COR permission.
The contractor will comply with all applicable NGA, DoD, and IC information security policies, including the Consolidated NGA Security Classification Guide, for marking, handling, processing, storing, and safeguarding classified and unclassified material, giving markings the lowest possible classification to maximize dissemination while maintaining confidentiality and integrity.
Access to CUI is pursuant to DFARS clause 252.204-7012, which the prime shall include, including paragraph (m), in subcontracts for operationally critical support or involving covered defense information, without alteration except to identify the parties. FAR clause 52.204-21 must similarly flow down.
Contractor personnel shall not release any unclassified information, in any medium, pertaining to any part of this contract or any related program unless the COR has given prior written approval or in performance of a project scoped and negotiated by NGA. For a company whose commercial strategy depends on publishing results or demonstrating capability publicly, that is a material constraint.
At minimum the contractor must implement 10 U.S.C. Sections 391 and 393 along with NIST SP 800-171. If using an external cloud service provider to store, process, or transmit covered defense information, the contractor shall ensure the provider meets security requirements equivalent to the FedRAMP Moderate baseline.
On discovering a cyber incident affecting a covered contractor information system or covered defense information, the contractor shall notify the DoD Cyber Crime Center and the COR in writing within 72 hours of discovery and conduct a compromise review. Isolated malicious software goes to DC3 per COR instructions and shall not be sent to the COR.
The FedRAMP Moderate cloud equivalence requirement deserves emphasis on this topic more than most. Training a vision language model is compute-intensive and the default answer is a commercial cloud with GPU capacity. If that environment will hold covered defense information, confirm the equivalence before you build a budget around it.
Insider threat
The contractor will establish and maintain an insider threat program to gather, integrate, and report information indicative of a potential or actual insider threat, consistent with Executive Order 13587 and the Presidential Memorandum on National Insider Threat Policy and Minimum Standards for Executive Branch Insider Threat Programs. The contractor shall report to the COR events affecting clearance eligibility, events indicating an insider threat, events affecting safeguarding, and events indicating classified information has been or is suspected to be lost or compromised.
Contract clauses to be aware of
FAR 52.204-7 System for Award Management, requiring registration at offer and continuously through final payment. FAR 52.204-27 Prohibition on a ByteDance Covered Application, meaning TikTok and successor applications.
NGA-specific clauses include 5X52.209-9003 Protection of Information and Nondisclosure Agreements, 5X52.227-9000 Unauthorized Use of NGA Name, Seal and Initials, 5X52.237-9001 Contractor Identification, 5X52.37-9000 Contractor Employee Data for Access to NGA Facilities or Sensitive Systems, which requires initial and timely updates to NGA's Human Capital Management System for all personnel with facility or sensitive system access and requires attendance at in-processing and out-processing briefings, and 5X52.246-9000 Contractor Compliance with all applicable NGA and U.S. Government installation regulations, directives, instructions, rules, policies and procedures.
An Inspector General cooperation requirement applies. The contractor must report to the NGA Inspector General, DoD IG, or Intelligence Community IG any and all possible violations of federal law or illegal intelligence activities related to the contract by individuals charging directly or indirectly to it, and the IG has access to those individuals and direct access to all related records. Failure to cooperate is grounds for administrative action. Contractors must make employees aware of the NGA IG Hotline. The requirement is supported by FAR 52.203-13 and NGA Instruction 7410.1.
Badging is also governed. NGA IC badges are issued only to contractors providing direct charge support on an active TS/SCI NGA contract, even when seated at corporate locations outside NGA facilities. Badges must be used at least once during a one-month period at an NGA government facility or may be suspended or terminated for inactivity, and they expire at the end of the supported contract. On departure, all NGA badges, Common Access Cards, hangtags, and other government furnished property must be returned no later than four business days from the date of departure.
Section 508 Compliance, Which Is a Proposal Requirement
This sits inside the Commercialization Strategy instruction and is easy to miss, and it applies more directly to this topic than to most.
The contractor shall ensure that all systems, hardware, software, software engineering, and information technology associated with this effort is made in a manner that is accessible for people with disabilities as directed in NGA Instruction 8400.4 and Section 508 of the Rehabilitation Act of 1973 as amended in 1998.
All Information and Communications Technology associated with this contract may use the Web Content Accessibility Guidelines 2.1 to comply with Section 508, or use alternative designs or technologies which result in substantially equivalent or greater access to and use of the product for people with disabilities.
Furthermore, the contractor shall pursue human centered design and usability guidelines to ensure that all services associated with this topic area are accessible by as many users as possible and to drive modernization, innovation, and enhance mission support.
As part of the proposal, the offeror should include an outline of specifically how Section 508 compliance will be achieved in the design of the ICT product. The Phase II proposal should provide an explicit, detailed description of the approach, indicate what is planned, how and where the work will be carried out, a schedule of major events, how the solution will be Section 508 compliant, and the final product to be delivered. If a determination is made that a Section 508 exception request is justified, the rationale for the exception request must be made and submitted as part of the proposal.
Why this matters more here
This topic delivers an analyst-facing tool. The Phase II output is location and classification of targets, and Phase III transitions the tool for government utilities. That means a user interface, which is squarely Information and Communications Technology.
There is also a natural alignment worth exploiting. A vision language model produces text. Text descriptions of imagery content are inherently more accessible than imagery alone, and screen reader compatibility for a natural-language explanation of a SAR scene is a genuine accessibility advantage rather than a compliance burden. Making that argument turns a checkbox into a strength, and it connects the 508 requirement to the reason a VLM is the right architecture in the first place.
Concretely: name WCAG 2.1, describe keyboard navigation, screen reader compatibility, contrast, and non-color-dependent status indication for your interface, and put accessibility review in the schedule. Use the exception path with stated rationale if some component genuinely cannot comply. Silence is the failure mode, and it is exactly what a proposal reused from another agency will produce.
Funding, Cost Structure, and NGA Mechanics
The award
The Phase II amount must not exceed $1,500,000 for up to a 24-month period of performance. Costs must be separated and clearly identified on the Proposal Cover Sheet in Volume 1 and in Volume 3.
Unless otherwise stated in the individual topic announcement, NGA Phase II awards are capped at $1,500,000 each over a maximum 24-month period of performance, and Phase I awards are capped at $150,000 each over a maximum six-month period of performance.
NGA caps sequential Phase II contracts, meaning those proposed near the completion of the initial Phase II or Direct to Phase II contract, at the then-current Small Business Administration "without seeking SBA approval" ceiling over a maximum 24-month period of performance.
No option periods
The offeror shall not propose option periods. That is flat, and it distinguishes NGA from components in this cycle that structure awards as a base plus options. Build a single 24-month effort.
Contract type and timing
NGA typically provides a firm fixed price contract for its awards within 180 days of the proposal due date, with type at the discretion of the Contracting Officer. One hundred eighty days from October 21, 2026 is roughly mid-April 2027.
Firm fixed price on a machine learning program means your estimate must absorb training compute uncertainty and data acquisition uncertainty. Both are real on this topic. Carry margin.
Percentage of Work, with no exceptions
Review the updated Percentage of Work calculation details included in the DoW Program. NGA will not accept any deviation to the POW requirements.
The temptation on this topic is to subcontract the SAR phenomenology expertise or the model training. Model your POW before assembling the team, and remember every subcontractor inherits the security requirements and the DFARS 252.204-7012 and FAR 52.204-21 flow-downs.
Technical and Business Assistance
Phase II awardees may request up to $50,000 per Phase II project, in addition to the cost ceiling and not subject to profit or fee.
All requests for TABA must be completed using the SBIR/STTR TABA Request Form, and the completed form must be included in Volume 5 of the proposal submission in DSIP. OSW will not accept requests that do not utilize the form or that are not included as a submission document in Volume 5.
For this topic, security and facility clearance consulting is the standout use if you are building that posture, followed by accessibility and Section 508 consulting, which is a stated proposal and performance requirement.
The Company Commercialization Report is not evaluated
Completion of the CCR as Volume 4 is required, but information contained in the CCR will not be considered by NGA during proposal evaluations. That differs from several other components in this cycle. Complete it because it is required, and put your transition thinking into the Commercialization Strategy inside the technical volume, which is also where the Section 508 requirement lives.
The technical volume structure
The technical volume consists of two parts, described in one place as Part A Feasibility Documentation and Part B Technical Proposal and in the volume list as Part 1 Phase I Justification and Part 2 Phase II Technical Proposal. Each part is not to exceed twenty pages, for a maximum of 40 pages. The Government will not consider pages in excess of these limitations. Number all pages consecutively and follow the DoW SBIR Program formatting requirements.
Content follows the Technical Proposal Template provided in the DoW Program, with NGA's additions: Section 508 compliance in the Commercialization Strategy, clearance levels in Key Personnel, and CUI protection with SPRS score plus facility clearance level, safeguarding level, CAGE code, and addresses in Facilities and Equipment.
Evaluation, selection, and protests
All proposals will be evaluated in accordance with the evaluation criteria listed in the DoW Program Solicitation, plus the controlled-work criterion described above.
Proposing firms will be notified of selection or non-selection status within 90 days of the closing date of the topic, via email. The NGA text says "for a Phase I award," which appears to be residual language given that this topic issues no Phase I award. Ninety days from October 21, 2026 is approximately January 19, 2027.
Refer to the DoW Program Solicitation for procedures to protest the announcement. As prescribed in FAR 33.106(b) and FAR 52.233-3, protests after award should be submitted to Patricia Hill at Patricia.D.Hill@nga.mil.
Questions
Specific questions pertaining to the administration of the NGA SBIR/STTR Program and these proposal preparation instructions should be directed to sbir@nga.mil.
The NGA instructions do not state that DSIP Topic Q&A is unavailable, so the standard DoW SBIR Program BAA process applies and Topic Q&A closes to new questions two weeks before the topic closes, on October 7, 2026. Worth asking: what government SAR SICD data will be made available, in what quantity, with what annotations, and at what classification level, because that answer materially shapes both your technical plan and your security plan.
The References
Three, and they are practical rather than academic, which tells you something about what NGA expects.
The NVIDIA vision language models glossary page. This is a vendor explainer, not a research citation, and its inclusion suggests the topic author wanted a plain definition of the architecture class rather than to point at a specific technique. Do not read it as an endorsement of any particular model family.
The sarpy package on PyPI, version 1.1.7. This is NGA's own open-source Python library for reading, writing, and processing SICD and other NGA sensor formats. It is the most actionable reference in the set. Using sarpy in your pipeline, or explaining why you use something else, demonstrates you have actually handled SICD files rather than treating the format as an abstraction.
Schwartzkopf, Cox, Koehler, and Fiedler, "Generic Processing Of SAR Complex Data Using the SICD Standard in Matlab," IGARSS 2019. This is the SICD standard paper and it explains why the standard exists: to let processing algorithms operate on complex SAR data independently of which sensor collected it. That sensor independence is the whole point of the format and the reason NGA specified it in the topic title. Read it and use its vocabulary.
The reference set is notably light on the AI side and notably specific on the SAR data format side. That asymmetry is a hint about where NGA thinks the risk lies: they assume you can build a model, and they want to know you can handle their data. Bring your own literature on the model side, on vision language models, on SAR automatic target recognition, and on adapting pretrained vision encoders to non-natural imagery, because the DoW Technical Proposal Template requires you to demonstrate awareness of the state of the art.
Timeline and What to Do When
The dates
Topic opens: September 23, 2026
DSIP Topic Q&A closes: October 7, 2026, two weeks before the topic closes, per the DoW SBIR Program BAA
Proposal deadline: October 21, 2026
Selection notification: within 90 days of the closing date, approximately January 19, 2027, via email
Contract award: NGA typically provides a firm fixed price contract within 180 days of the proposal due date, approximately mid-April 2027
Period of performance: up to 24 months
A working backward plan
Before September 23. Confirm you can satisfy both halves of the feasibility gate. On the accuracy side, verify your greater than 95 percent result is on SICD-format complex data and not on detected magnitude chips, and assemble the full experimental description including class set, chip counts, split methodology, and confusion matrix. On the qualification side, map your team against the three named competencies of SAR signal and image processing, SAR operations from space, and deep learning implementation, and identify how you cover any gap. Audit the funding provenance of your results, since work based upon or logically extending from prior federally funded SBIR or STTR work is excluded. Resolve intellectual property ownership or license rights. Confirm your security posture: Facility Clearance Level, safeguarding levels, CAGE codes, addresses, SPRS score, and NIST SP 800-171 status. Identify who holds current Top Secret clearances and who has been polygraphed within five years. Resolve foreign national participation, remembering the prohibition extends to unclassified work. Confirm your training compute environment meets FedRAMP Moderate equivalence if it will hold covered defense information. Decide your data strategy across government, commercial, and internal sources. Get familiar with sarpy and read the SICD standard paper. Model your Percentage of Work. Send your data availability question to Topic Q&A early.
September 23 through October 5. Draft the 20-page feasibility documentation. Use the space for the full accuracy experiment, the SICD handling detail, example chips with model outputs, the data provenance discussion, and complete personnel qualification evidence with degrees, publications, and specific prior complex SAR work. Draft the 20-page Phase II technical proposal against the DoW Technical Proposal Template. Make the case for why a VLM rather than a classifier, addressing explanation and open-vocabulary search. Address the SAR-specific encoder adaptation problem and where your language supervision comes from. Explain how you bridge from single-chip classification to multi-target search, and propose your own detection and classification metrics with justification. Cover the data plan and the domain gap. Include the Section 508 outline in the Commercialization Strategy, clearance levels in Key Personnel, and full facility and CUI detail in Facilities and Equipment. Do not propose option periods.
October 6 through October 7. Submit remaining questions through DSIP Topic Q&A before it closes. The most valuable is what government SAR SICD data will be available, in what quantity, with what annotations, at what classification. Administrative questions go to sbir@nga.mil.
October 8 through October 14. Build the cost volume, with costs separated and clearly identified on the Proposal Cover Sheet and in Volume 3, against the $1,500,000 and 24-month ceiling with no options. Price training compute realistically, data acquisition or licensing, annotation effort including any language supervision generation, secure compute or facility access if needed, security administration, accessibility work, and the productization effort for a tool that transitions. Firm fixed price is typical, so carry margin on compute and data. Complete the SBIR/STTR TABA Request Form and place it in Volume 5.
October 15 through October 18. Complete Volume 4, the Company Commercialization Report, which is required though not evaluated by NGA. Assemble Volume 5 with the TABA form and supporting documentation. Complete Volume 6 training and the Volume 7 foreign affiliations webform, remembering it must be the webform and will not be accepted as a PDF in Volume 5, and that no previous versions should be uploaded there. Run compliance: 20 plus 20 pages, consecutively numbered, no option periods, clearance levels listed, SPRS score and facility details provided, Section 508 outline present.
October 19 through October 20. Submit and certify in DSIP.
Frequently Asked Questions
What is NGA SBIR topic OSW26BZ06-DV034?
OSW26BZ06-DV034 is a Direct to Phase II SBIR topic titled "Vision Language Model (VLM) for Synthetic Aperture Radar (SAR) target search and classification from Sensor Independent Complex Data (SICD)," released under the National Geospatial-Intelligence Agency 2026 SBIR Broad Agency Announcement, Release 6. The objective is to develop VLM based SAR imagery search and classification from sensor independent complex data.
How much funding is available?
The Phase II amount must not exceed $1,500,000 for up to a 24-month period of performance. Phase II awardees may also request up to $50,000 in Technical and Business Assistance, in addition to the cost ceiling and not subject to profit or fee, using the mandatory SBIR/STTR TABA Request Form in Volume 5.
When is the proposal deadline?
The topic opens September 23, 2026 and proposals are due October 21, 2026 through the Defense SBIR/STTR Innovation Portal at dodsbirsttr.mil.
Can I submit a Phase I proposal?
No. This topic is accepting Direct to Phase II proposals only.
Can I propose option periods?
No. The NGA instructions state that the offeror shall not propose option periods.
How long can my technical volume be?
Forty pages total, consisting of a feasibility documentation part at 20 pages maximum and a technical proposal part at 20 pages maximum. Pages in excess will not be considered. Number all pages consecutively.
What must my feasibility documentation demonstrate?
Two things. Successful classification with greater than 95 percent accuracy of single SAR image chips in SICD format. And that the performer is very well qualified, with PhD or research level given as examples, and has demonstrated experience classifying complex SAR imagery. NGA states that qualification requires a scientist's extensive knowledge of SAR signal and image processing, SAR operations from space, and deep learning understanding and implementation.
Does my 95 percent result have to be on SICD data?
That is what the topic says: classification of single SAR image chips in SICD format. Most publicly available SAR classification work uses detected magnitude imagery rather than complex data. If your result is on magnitude chips, reproduce it on SICD data if you can, and if you cannot, state plainly what you have and what the gap is rather than letting a reviewer discover it.
How should I report the accuracy?
With the full experimental description: the class set and number of classes, the number of chips, the train and test split, whether the split is by scene or by chip, and the confusion matrix. A 95 percent figure on a small closed class set with a chip-level random split is a much weaker claim than the same figure on a scene-disjoint split, and a reviewer in this field will know.
What does the qualification requirement actually mean?
Three named competencies. SAR signal and image processing, meaning the phenomenology of how the image forms and what phase carries. SAR operations from space, meaning collection geometry, look and incidence angles, collection modes, and the metadata that describes them. And deep learning understanding and implementation, meaning you build and train models. Address all three with named individuals, degrees, publications, and specific prior complex SAR work.
Can my feasibility evidence come from a prior SBIR award?
No. NGA's Direct to Phase II guidelines state that feasibility documentation cannot be based upon or logically extend from any prior or ongoing federally funded SBIR or STTR work, and that the work must have been substantially performed by the proposer or the Principal Investigator. SAR automatic target recognition work in the United States has been substantially SBIR and STTR funded, so audit your provenance.
What happens if my feasibility documentation is inadequate?
If the proposer fails to demonstrate technical merit and feasibility equivalent to the Phase I level as described in the associated topic, the related Phase II proposal will not be evaluated.
What does Phase II have to deliver?
A VLM based SAR SICD imagery classification system where input contains SAR imagery with multiple targets and output is location and classification of the targets. You may use SAR SICD data provided by the government or your own or commercial data.
How does Phase II differ from the feasibility requirement?
Substantially, and this is the key structural feature of the topic. Feasibility is single chips with one target each, classified. Phase II is full imagery with multiple targets, and the output includes location as well as classification. That is a jump from classification to detection plus classification, where false alarms across large scenes become the dominant failure mode. Your Phase II plan should explain how you bridge that gap and should propose detection metrics rather than only classification accuracy.
Does NGA specify Phase II performance targets?
No. The only stated number is the greater than 95 percent feasibility accuracy on single chips. Propose your own Phase II metrics, justified against the scene sizes and target classes you address, using measures that account for false alarms such as precision and recall, mean average precision, or a probability-of-detection versus false-alarm-rate curve.
What data will I use?
SAR SICD data provided by the government, or your own or commercial data. Say which, in what proportion, and how you handle the domain gap between them. Government data is likely the most mission-representative and the most likely to be controlled, which brings the security requirements into your data handling plan. It is worth asking through Topic Q&A what government data will be available, in what quantity, with what annotations, and at what classification level.
What is SICD, and why does it matter?
Sensor Independent Complex Data is NGA's standard format for complex SAR data, retaining phase as well as amplitude in single look complex form. The point of the standard, as the cited Schwartzkopf paper explains, is to let processing algorithms operate on complex SAR data independently of which sensor collected it. NGA's own open-source sarpy Python library reads and processes it, and using sarpy signals that you have handled real SICD files.
Why does the model have to be a vision language model?
The topic does not say directly, which makes it the central question your proposal should answer. The plausible reasons are explanation, since NGA describes a vision encoder that allows the language model to explain contents of an imagery, which serves an analyst who must accept or reject a call; and open-vocabulary query-driven search, which is what target search implies and which a fixed-class classifier cannot do. Making that argument well distinguishes you from proposals treating VLM as a label on a conventional detector.
What makes SAR hard for a vision language model?
Vision language models are pretrained on natural imagery and text. SAR is complex-valued, speckled, geometrically distorted by layover and shadow, radiometrically counterintuitive, and highly sensitive to look angle. A vision encoder trained on photographs has learned features that do not transfer cleanly. Your proposal should address how you adapt or replace the encoder for complex input, how you use SICD metadata, and where your text supervision comes from, since there is no large corpus of SAR imagery paired with expert descriptions.
Do I need a security clearance?
Yes. The instructions state that all Contractor personnel shall possess a current Top Secret Personnel Security Clearance and be eligible for favorable NGA adjudication for SCI access. Personnel performing TS/SCI work require active TS/SCI clearances for access to TS/SCI facilities, environments, and computer systems. NGA will sponsor SCI accesses at its sole discretion if needed.
Is a polygraph required?
Contractors are subject to a counterintelligence-scope polygraph examination as requested by NGA. Contractors who have not successfully completed polygraph testing within the last five years must immediately schedule an examination and complete the process, in no more than three test sessions, within 90 days, as a condition of employment and assignment.
Can uncleared machine learning staff work on this?
In a limited way. After discussion with the COR and with written COR approval, uncleared contractor personnel are authorized to work at the unclassified level with access up to DoD Controlled Unclassified Information at the contractor site without a security clearance. Personnel working with CUI must receive favorable HSPD-12 adjudication for 60 days or less of access, or HSPD-12 Tier 1 adjudication for more than 60 days. NGA will sponsor those investigations as needed. Structure your data plan and staffing plan together, since they determine each other.
Can foreign nationals work on this contract?
Not without prior written approval from the Contracting Officer or the COR, and that prohibition applies to unclassified work. If NGA identifies unauthorized personnel, the CO may direct removal and replacement at the contractor's expense, without prejudice to other remedies including termination for default. For a deep learning team this is often the binding constraint.
What must my Key Personnel section include?
In addition to the standard content, list the clearance level held by each of the personnel, because, as NGA puts it, we will be operating in a classified environment. On this topic that section is doing double duty, since the feasibility gate also requires demonstrating the performer's SAR and deep learning qualifications.
What must my Facilities and Equipment section include?
How Controlled Unclassified Information will be protected in accordance with NIST SP 800-171, your Supplier Performance Risk System score, and your Facility Clearance Level, the safeguarding level for each facility, the CAGE code for each facility, and the associated physical addresses.
Is my security posture actually scored?
Yes. In addition to the DoW Program evaluation criteria, NGA will evaluate a vendor's ability to perform controlled work, meaning CUI and classified, as part of their technical evaluation score. NGA may use SETA support for the programmatics of executing the evaluation but will not use SETA support to evaluate proposals.
What are the cybersecurity requirements?
At minimum, implement 10 U.S.C. Sections 391 and 393 along with NIST SP 800-171. If using an external cloud service provider to store, process, or transmit covered defense information, ensure it meets security requirements equivalent to the FedRAMP Moderate baseline. On discovering a cyber incident, notify the DoD Cyber Crime Center and the COR in writing within 72 hours and conduct a compromise review. Isolated malicious software goes to DC3 per COR instructions, not to the COR. For a model training program on commercial GPU cloud, the FedRAMP equivalence point is the one to check early.
What is the Section 508 requirement?
The Commercialization Strategy must address Section 508 compliance per NGA Instruction 8400.4 and Section 508 of the Rehabilitation Act of 1973 as amended in 1998. Information and Communications Technology may use WCAG 2.1 to comply, or alternative designs achieving substantially equivalent or greater access. The proposal should include an outline of specifically how compliance will be achieved, and if an exception is justified the rationale must be submitted as part of the proposal. It applies here because the deliverable is an analyst-facing tool, and a natural-language description of imagery is inherently more accessible than imagery alone, which is an argument worth making.
What CMMC level applies?
The projected requirement for this topic is CMMC Level 2 with self-assessment. Separately, the classified nature of the work imposes personnel and facility requirements well beyond CMMC.
Is this topic ITAR restricted?
No topic-level ITAR or EAR restriction paragraph appears on OSW26BZ06-DV034. The classified environment and the foreign national prohibition impose stricter constraints than an ITAR notice would.
Is the Company Commercialization Report evaluated?
No. Completion of the CCR as Volume 4 is required, but information contained in the CCR will not be considered by NGA during proposal evaluations. Put your effort into the Commercialization Strategy inside the technical volume instead.
Are there Percentage of Work restrictions?
Yes. NGA will not accept any deviation to the Percentage of Work requirements described in the DoW Program. Note that subcontractors also inherit the security requirements and the DFARS 252.204-7012 and FAR 52.204-21 flow-downs.
What contract type should I expect, and when?
NGA typically provides a firm fixed price contract within 180 days of the proposal due date, roughly mid-April 2027 for this cycle, with type at the discretion of the Contracting Officer. Firm fixed price on a machine learning program means carrying margin for training compute and data acquisition uncertainty.
Can I publish or publicize this work?
Only with permission. Contractor personnel shall not release any unclassified information, in any medium, pertaining to any part of the contract or any related program unless the COR has given prior written approval or in performance of a project scoped and negotiated by NGA. The NGA clause on unauthorized use of the NGA name, seal, and initials also applies. For an AI company that publishes as part of its recruiting and commercial strategy, this is a real constraint worth planning around.
When will I hear back?
Within 90 days of the closing date of the topic, approximately January 19, 2027, via email. Note that the NGA text says notification "for a Phase I award," which appears to be residual language given that this topic issues no Phase I award.
Who do I contact with questions?
Technical questions about the topic go through DSIP Topic Q&A, which closes October 7, 2026. Administrative questions about the NGA SBIR/STTR Program and these proposal preparation instructions go to sbir@nga.mil. Protests after award go to Patricia Hill at Patricia.D.Hill@nga.mil.
Positioning Advice for Companies Considering This Topic
Check both halves of the feasibility gate before you invest. The accuracy half is a number. The qualification half is a credential requirement, and it names three competencies that rarely coexist: SAR signal and image processing, SAR operations from space, and deep learning implementation. If you are a strong AI company without deep SAR phenomenology expertise, or a SAR house without model-building depth, you have a gap that NGA has explicitly told you it will look for.
Verify your 95 percent result is on complex data. SICD is complex, and most published SAR classification results are on detected magnitude chips. If yours are magnitude, reproduce on SICD or say clearly what you have. A reviewer who works this problem will spot the substitution immediately, and being straightforward about a gap costs less than being caught glossing over one.
Report the accuracy like a scientist. Class set, chip counts, split methodology, scene-disjoint or not, confusion matrix. On a topic that explicitly requires research-level qualification, how you report results is itself evidence of qualification.
Answer the "why a VLM" question that the topic leaves implicit. Explanation for the analyst, and open-vocabulary query-driven search, are the two defensible answers, and both connect to the Trusted AI and Autonomy priority area. A proposal that uses a language model without arguing why language is the right interface for GEOINT looks like it followed a trend rather than a requirement.
Confront the encoder problem head on. Pretrained vision encoders learned photographs. SAR is complex-valued, speckled, geometrically distorted, and look-angle dependent. Say what you do about it: complex input handling, magnitude and phase channels, learned representations, metadata conditioning. This is the deepest technical question in the topic and where a serious team separates from a capable one.
Say where your language supervision comes from. There is no large corpus of SAR imagery paired with expert text. Synthetic captions from metadata and annotations, expert-written descriptions, label bootstrapping, or something else. Most proposals will skip this and it is the practical bottleneck of the whole approach.
Plan the bridge from chips to scenes explicitly. Feasibility is one target per chip. Phase II is many targets per scene with localization. False alarm rate across a large scene, not classification accuracy, is the metric that will determine whether the tool is usable. Propose those metrics yourself, since NGA states none for Phase II.
Use sarpy, or explain why not. It is NGA's own library, it is one of only three references, and it is the most actionable signal in the topic that they care whether you can handle their format. Reading the SICD standard paper and using its vocabulary reinforces the same point.
Resolve the clearance and foreign national questions before you write. Top Secret with SCI eligibility for all personnel, counterintelligence-scope polygraph, and foreign nationals barred from even unclassified work without prior written approval. For a modern machine learning team this is usually the binding constraint, not the technology.
Check your SPRS score today. NGA asks for the number, a reviewer can look at it immediately, it feeds a scored criterion, and it takes weeks to improve.
Structure the team around the two security tiers. Uncleared but HSPD-12 adjudicated staff can do unclassified model development on commercial or internal data with written COR approval. Cleared staff handle government-provided controlled data and anything in a classified environment. Design your data plan and staffing plan together, and note that laptops and removable media are barred from secure facilities, so development inside a SCIF looks nothing like development outside one.
Confirm FedRAMP Moderate equivalence for your training environment. Training a vision language model means significant GPU capacity, and the default answer is commercial cloud. If covered defense information will touch it, the equivalence requirement applies, and discovering that after award is expensive.
Write the Section 508 outline, and make it an argument rather than a checkbox. This topic delivers an analyst-facing tool, so accessibility applies. And a model that describes imagery in language is genuinely more accessible than one that outputs bounding boxes. Naming WCAG 2.1, describing conformance for your interface, and connecting accessibility to the reason a VLM is the right architecture turns a requirement into a strength.
Use all twenty feasibility pages. Four times the allowance most components give. Full experimental detail, data provenance, example outputs, SICD handling specifics, and complete personnel credentials all fit, and the qualification half of the gate practically demands the space.
Do not propose options, and price for firm fixed price. NGA prohibits option periods outright, and typically awards FFP within 180 days. Carry margin on compute and data, which are the two least predictable costs here.
Plan your commercialization story around the publication constraints. No release of unclassified contract information without written COR approval, and the NGA name and seal are protected. An AI company that publishes to recruit and to build credibility needs a strategy that works within that, and saying so honestly reads better than a plan that ignores it.
Ask about the government data. What SICD data will be available, how much, with what annotations, at what classification level. That single answer shapes your architecture, your annotation budget, your staffing tiers, and your facility needs. Topic Q&A closes October 7.
