What's actually true about the program
Checked against NSF's solicitation (NSF 26-510, updated Oct 6, 2026) and the reauthorization record. The viral post was mostly right.
Eligibility, rule by rule
Tick items as we confirm them. Saved for the whole team.
His concerns, answered
Guessed from his chat with Toolbelt-Strategy and the usual objections to SBIR. Each one gets an honest verdict.
What we'd actually pitch
Not "an AI operations platform." A research question with unproven feasibility, which is the thing NSF funds.
Can an AI agent's repeated behavior be automatically compiled into verified, deterministic code, with a cheap typed gate that knows when the compiled path is safe and a monitor that catches when a model or provider change breaks it?
Why NSF cares
- Cuts compute and energy per task (AI topic AI6, sustainable AI)
- Same answer every time, auditable (AI7, trustworthy AI)
- Small US businesses stop depending on one model provider
Why it's genuinely risky
- Real agent traces may resist synthesis
- A wrong "safe" call returns a confident wrong answer
- Drift is hard to detect without ground truth
Why we can do it
- Code steps, classifiers, storage and compression already run in production (confirmed with Ryan)
- Real multi-org traces to learn from
- Leads asking for exactly this, in their words
Prior art we must differentiate from: model routing (FrugalGPT, RouteLLM) and response caching still send every request to a model or replay stored text. Neither turns behavior into verified code, and neither gives a business the same answer every time. That's our two-sentence difference.
Draft Project Pitch
NSF's four sections, with live character limits. Edit in place, then copy. Bracketed items need founder input; nothing here is invented.
Is it worth one day?
Plug in your own odds. NSF doesn't publish invite or award rates for this cycle, so the defaults are placeholders, not data.