The Defense Advanced Research Projects Agency's CLARA program — short for the high-assurance AI initiative — funds research aimed at AI systems whose behavior can be formally proven. Up to $2M per award over 24 months, with a hard requirement that all software produced be released under Apache 2.0. Eligible proposers include universities, research organizations, and small businesses; proposals are due April 10, 2026. Coverage from Granted AI.
The unusual ask: prove the AI works
Most federal AI procurement focuses on benchmarks and demos. CLARA is different — it funds research into AI systems where behavioral guarantees are formally verified, not just empirically observed. That makes the program a natural fit for firms working on:
- Formal verification of neural network behavior
- Constraint-based reasoning systems
- Hybrid symbolic/neural architectures
- Provable-bounds reinforcement learning
Open-source-mandatory is the differentiator
The Apache 2.0 requirement is unusual for DARPA. It signals that the agency wants the research output to circulate broadly — including potentially to commercial customers — rather than be locked into specific defense programs. For small firms, this creates a useful asymmetry: the IP stays open, but the firm retains its know-how, customer relationships, and proprietary tools built around the open-source release.
Firms with experience commercializing open-source projects (think GitLab-style company-around-open-source) are well-positioned for CLARA.
The broader DARPA small-business path
CLARA isn't the only DARPA AI funding currently open. Two other programs worth tracking:
- I2O Office-Wide BAA. Individual awards range from $500k to $5M depending on scope. Full proposals due November 30, 2026. Broad theme: information innovation.
- SBIR/STTR direct-to-Phase II. DARPA recently awarded a direct-to-Phase II SBIR contract to CoVar, a small AI/ML firm, for the Predictive Psychological Architectures for Decision-Making (PPADM) program. Direct-to-Phase II skips Phase I — a meaningful acceleration for firms with prior R&D credentials.
What CLARA proposers actually need
From the public solicitation summary:
- Original research approach with quantifiable assurance metrics
- A clear path to publicly demonstrable software releases
- Apache 2.0 license commitment for all software outputs
- Realistic 24-month milestone schedule
- Team capable of formal-methods + ML work (rare combination)
What to do this week
- If your team has formal methods + ML capabilities, the April 10 deadline is tight but achievable with focused effort. Pull the BAA solicitation from DARPA and identify which sub-program fits.
- If CLARA's timeline is too short, the I2O BAA's November 30 deadline gives you 6 months of preparation time for a similar profile of work.
- Consider partnering: CLARA explicitly welcomes universities-with-small-business teams. The combination of academic methodology and commercial productization is exactly what DARPA wants.