Vision Language Model (VLM) for Synthetic Aperture Radar (SAR) target search and classification from Sensor Independent Complex Data (SICD)

• Five new Q&As added (Q5–Q10 in the new numbering, plus a renumbered sensor question): The Q4 "sensor independent" question was moved up, and Q5 now answers that the Phase II system must demonstrate performance across SICD data from multiple SAR sensors, which can be government or commercial X-band data.
• Target scope clarified: The customer expects target detection, location, and classification, possibly with a count of targets; no language-based VLM functionality (natural-language query, analyst interaction) is expected.
• "Location" defined: Target location means position in image space or geometric space, depending on metadata availability.
• 95% threshold definition: The Government gives only a general definition of 95% accuracy, with no prescribed metric, class count, or train/test method; offerors are asked to frame it around an IC/DoW problem.
• Technical volume format approved: Part A (Feasibility) may be organized under Appendix A format with a Part A/Part B split as a subset of the Phase II Statement of Work, rather than the 12-section DoW template for both parts.
• Phase II evaluation (new Q10): Evaluation will look at classification accuracy of 95% or above, which may depend on data; no thresholds were given for recall, false alarms, localization, or processing time, and it is unclear whether a Government test set will be used.
• Input data type settled: Offerors should use SICD scenes and classify the targets they find feasible.

The Q&A was substantively updated, mostly by answering questions previously pending:

• Personnel clearance (new Q1 answer): All Phase II personnel should hold a final Top Secret clearance eligible for SCI adjudication by NGA. The offeror must have a Top Secret facility clearance and SCIF access, and NGA expects performers with established TS/SCI facilities and personnel. This is a significant eligibility barrier for applicants without existing TS/SCI capability.
• Classes for the >95% feasibility requirement (new Q2 answer): 95% accuracy on any target class is acceptable, including ground, maritime, air, or mixed classes. Offerors may define their own approach.
• Clearance timing, classified data, and unclassified data (new Q3 answer): Proposals will be judged on technical merit, expertise, and execution in a proper facility. NGA may provide SICD data at various classification levels. Unclassified data use is acceptable for development.
• System input (Q4 answer, previously pending): Use SICD scenes rather than only chips, and classify the targets you consider feasible.
• Previously posed questions (Q5 and the Technical Volume format): Answers now read "Yes, that is acceptable," approving the proposed Part A/Part B structure. Q5 in the prior version's list appears to be restated here as the earlier questions on accuracy definition, location, VLM functionality, and sensor independence, which remain unanswered in this version.

Why it matters: applicants without TS/SCI facility clearances and personnel now face a stated expectation that may exclude them, while the technical feasibility and format questions are largely resolved.

• Six new Q&As added (previous version empty, so all content is new), covering:
  • Q1: whether the >95% Direct-to-Phase-II accuracy threshold has a prescribed accuracy definition, class/sample minimums, or train/test methodology
  • Q2: meaning of "location" for the Phase II requirement (image-space, geospatial, or both)
  • Q3: whether the Government expects language-based VLM functions such as natural-language query or analyst interaction
  • Q4: meaning of "sensor independent" (multi-sensor SICD performance vs. a single Government-designated source)
  • Q5: whether the VLM must ingest SICD chips or whether larger SICD scenes are acceptable input
  • Q6: Technical Volume format for Part A (Feasibility) and Part B (Phase II Proposal); the answer confirms applicants may use the DoW 12-section format for Part B and a topic-driven outline for Part A.
• Answers are provided only for Q6 (approved) in this text; Q1 through Q5 show no response shown here, so they may still be pending.

Status changed from status: Pre-Release to status: Open.

New opportunity: Vision Language Model (VLM) for Synthetic Aperture Radar (SAR) target search and classification from Sensor Independent Complex Data (SICD)