Papers, essays, and position pieces from Ashiba Research. Ongoing posts at ashibaresearch.substack.com.
The founding paper for Ashiba Compute. A specification, verification, and attribution framework for ML kernels across heterogeneous silicon. Eight-part contract object, three failure primitives, twelve contract classes, derived tolerance bounds. The reference verifier (Apache 2.0) is at github.com/cv700/ashiba-verify.
The opening essay for Ashiba Alignment. Why the verification of heterogeneous, opaque compute substrates is environmental work in the deepest sense — work that maintains the conditions under which moral life remains possible across a plural, AI-mediated world.
Curiosity beats cadence in long-tail deep-tech markets. The forward-deployed researcher is the lab's brain in the field, bringing back the gradient that decides the production possibility frontier of intelligence.
The case for treating operator data — the workflow exhaust of the real economy — as the next training-data frontier, and for building the legal-and-technical infrastructure that makes it licensable without giving away the moat.
What an operational AI data asset is, who has them, who pays for them, and why the rights problem has so far prevented the market from forming. Companion to LADDER.
Seventeen environments measuring whether coding agents follow document-grounded operational requirements. The visible-pass, hidden-operational-fail failure shape applied at the agent level.
Essays and announcements ship through Substack: ashibaresearch.substack.com. Code lives on GitHub. Threads on X and LinkedIn.