Radical openness.
By default.
Real-time telemetry on compute, safety, and capital. We believe that trust is a function of verifiability, not marketing.
Operational Telemetry
Safety Incidents
All incidents are publicly logged within 24 hours. We categorize events by severity and impact.
Capability Boundaries
We maintain a living registry of known failure modes ("System Card"). These are active research areas where our capabilities fall short.
View Research RoadmapScalable Oversight
UnsolvedWe cannot yet guarantee alignment verification as models become more capable than evaluators.
Deceptive Alignment
Open ResearchDetecting models that appear aligned during evaluation but may behave differently in deployment.
Long-horizon Planning
LimitedCurrent systems struggle with tasks requiring planning over extended time horizons.
Distribution Shift
ImprovingPerformance degrades in situations significantly different from training data.
We pick partners,
not the other way around.
In safety-critical systems, misaligned capital is a primary failure mode. We do not accept funding that implicitly or explicitly demands shortened timelines.
Our "Reverse Due Diligence" process ensures that every dollar on our cap table is legally bound to the mission of safe AGI, stripping away traditional investor rights in favor of mission integrity.
Governance Air-Gap
Investment capital purchases economic interest only. It confers zero governance rights, zero board seats, and zero influence over deployment timelines.
The Safety Veto
The Research Integrity Council holds unilateral power to halt commercialization if safety benchmarks are missed, regardless of investor liquidity needs.
Timeline Decoupling
Partners must agree to a multi-decadal horizon. We do not optimize for quarterly cycles or short-term exits.