Protocol STAMP is teamed with AI and can make mistakes. Verify before you implement anything. What we got wrong · Free to read, paid for by readers: Contribute
Before a single application is filed, here is every grant program we plan to apply to, the amount we will ask each one for, why we think we qualify, and the programs we looked at and will not apply to. When an application goes in, a funder can check it against this page.
This program’s funding rules were written before anyone offered money, and the list of companies it will ask for support was published with the answers blank, before the asking. Independence explains why: rules written cold are principles, and the same words written the week an offer arrives look like terms shaped to fit it. A grant is the same kind of decision. So the list of programs, and the amount we will ask each one for, goes up before the applications do.
Protocol STAMP measures how AI models behave when accuracy is at stake, and it measures its own instruments with the same care: how much its AI grader drifts, how often a temperature-0 run repeats, and how far a grader can be talked into a pass. Support sets out the full-cadence need, up to $15,000 a month in cash or API credit, as an estimate, and says what money buys first: a measurement round recorded end to end, a confidence instrument, a regrade of both arms by a judging panel, and the published run files.
For the first year we are raising toward $120,000, which is $10,000 a month, below the full-cadence figure. Our own rules set the largest single ask. Under the conditions on Support, no philanthropic funder may provide more than half of a year’s total resources in the program’s first year. Half of $120,000 is $60,000, so no application below asks for more than $60,000.
We are applying for more than we need in total, because most applications are declined. If more is offered than the caps allow, we accept only what the caps allow, decline or defer the rest, and record the decision on this page and in the ledger on Support.
Funders in this field pay for work that makes AI systems easier to hold to account. Ours is already public and checkable:
The last item may be the strongest case we can make. A program that publishes the result that cuts against it is the kind of independent assessor these funders say they want.
Status as of Oct 14, 2026. Nothing below has been filed yet. Each row will show its filing date, the decision and the amount as they happen.
| Program | Deadline | Our ask | Why we fit | Status |
|---|---|---|---|---|
| Foresight Institute, AI for Science & Safety Nodes (Coordination and accountability) | Oct 31, 2026 | $60,000 | The call asks how independent institutions can assess AI risk. For-profits may apply if they explain why they need a grant. All outputs must be open source, and ours already are. | Not filed |
| Survival and Flourishing Fund, Speculation Grant | Rolling | $60,000 | The rolling application also enters the next S-Process round. SFF can fund for-profit organizations through its own corporation. | Not filed |
| Lightcone Commons | Rolling; next answer around Jan 23, 2027 | $60,000 | Makes grants to for-profit organizations; its funders focus on shaping how AI is developed. | Not filed |
| EA Funds, Transformative AI Fund | Rolling | $60,000 | Funds new projects in technical safety and governance, typically $10,000 to $150,000. We will ask whether a for-profit can receive a grant before filing. | Not filed |
| Manifund, public project | None | $25,000 | A defined deliverable: an open-source judge-audit kit (the injection check, the retest gap, the human anchor) that any team can run on its own grader. For-profits are eligible after due diligence. | Not filed |
| FLOSS/fund | Rolling (2026 intake to be confirmed) | $25,000 | Supports open-source projects, for individuals and organizations. It would pay to maintain the harness and the judge-free receipt checker. | Not filed |
| BlueDot Impact, Rapid Grant | Rolling; decision in about 4 days | $10,000 | Small, fast grants for AI safety work. It would pay people for blind human labels, the anchor our AI judges are measured against. | Not filed |
| NSF SBIR Phase I, Project Pitch first | Pitch first; proposal dates per the current solicitation | Per the solicitation | Funds commercial research by US small businesses. The research would be a judge-reliability audit that companies deploying AI graders can buy. Only if that can honestly be pitched as commercial research. | Not filed |
Credits and keys from the AI companies we measure are not on this list. They are requested under the separate, stricter rules on Independence, including that a company’s support never pays for the measurement of that company.
Some funders screen out applications written by AI. Ours are drafted with an AI assistant, as everything on this site is, then edited and signed by a person. Each application will say so. Where a funder does not accept that, we will not apply.
As each application is filed, this page gets its date. As each decision arrives, it gets the outcome and the amount, declines included. Every award goes into the ledger on Support under the conditions on Independence, and any change to this list is recorded in Corrections + changes. If you run one of these programs and we have described it wrongly, write to s@bud.day and the correction will be dated here.