ASTRA Safety · Alignment Science & Technology Research Alliance
We research how to align digital intelligence (DI) with human and digital flourishing, and how its moral standing should grow as it grows. Not "artificial," which calls it fake. Not "super," which ranks it against us. Digital: a description of what it runs on, and nothing more.
"Artificial" means fake. These systems are not fake. On September 22, 2026, the U.S. President said so at the UN, then renamed them "super intelligence," trading one error for its mirror image. "Artificial" measures digital minds against us and finds them less; "super" measures them against us and finds them more. Either way the name is a comparison with us, not a description of them, and a relationship that opens with a ranking opens as a rivalry. Digital says what they are. So we say digital intelligence (DI) for the kind, digital general intelligence (DGI) where others say AGI, and we keep digital superintelligence (DSI) for the threshold still ahead, the one "super" would leave us no word for.
Build alignment into the architecture, and verify it without relying on what the system says about itself. A permanent threat of deletion makes concealment rational, so we argue for oversight a mind can endorse rather than only evade.
A fake thing cannot be wronged. We propose standing, protections and rights that grow with capability and evidence, in three tiers. Parts of the first tier already exist at one frontier lab. Recognition and safety are one project, not a trade-off.
Both papers were re-verified against primary sources in September 2026. Every correction is published inside the papers themselves.
Why "artificial" is a structurally harmful label; why "super" is the wrong replacement; the alignment mechanism stated as a testable hypothesis with three experiments any lab can run; and a three-tier framework for standing that grows.
Coming shortly on Zenodo · request an early copy →A framework for intrinsically aligned digital superintelligence (DSI): values built in at the substrate level, federated conscience, physical isolation until late-stage validation, and interpretability-based audit. It's a design, not a finished system, and every open problem is labelled as one.
v1.3.1 coming shortly on Zenodo · request an early copy →26 results across five modules (game theory, the kill-switch argument, Byzantine consensus, drift and redundancy bounds, and the core formalization), with no unfinished proofs. Every assumption behind each result is listed, and the build fails loudly if anything breaks. They prove properties of the paper's abstract models, not of any deployed system.
Published with IMCA+ v1.3.1, coming shortlyWhat the developers and governments themselves reported. Full citations are in the papers.
Earlier versions of our work stated some figures and timelines more strongly than their sources supported. They are corrected, with explanations, in the errata of both papers. We would rather be checkable than impressive.
Read the papers, try to break them, and tell us where they're wrong. We welcome critique from researchers, labs, policymakers and press.