AI with Authority, from Application to Silicon

A five-week, single-researcher campaign from application code to a taped-out RISC-V processor, with every proof kernel-checked and no human RTL

Jason Hickey· 2026· DOI 10.48550/arXiv.2608.21356

The core problem

For sixty years, machine verification has been treated as a major cost overhead — affordable only for exceptional artifacts, and rarely for the ordinary software and hardware that industry actually ships. The premise of this work is that generative AI inverts that relationship. At AI speed, machine verification is not merely economical; it is essential to productivity. It functions as the incorruptible referee that lets one person safely direct autonomous machine work at scale. The reported campaign is deliberately extreme: in five weeks, one researcher on consumer AI subscriptions directed a small fleet of AI agents from application code, through a verified compiler and executive, down to a RISC-V processor taped out on a community silicon shuttle. No proof passed through human review, and no RTL was written by a human. The organizing discipline — the **Salt method** — rests on a proof kernel that no hallucinated proof can pass. Mathematical claims travel between agents as kernel-checked artifacts, and human attention is reserved for statements, designs, and rulings. Verification is stated link by link, from the Lean 4 kernel to SAT-checked equivalence at the silicon boundary. T

Innovation

The campaign completed in five weeks. One researcher, working on consumer AI subscriptions, directed a small fleet of AI agents from application code, through a verified compiler and executive, to a RISC-V processor taped out on a community silicon shuttle. Two negative results are as important as the positive one: no proof passed through human review, and no RTL was written by a human. The accounting is published in full. Theorem provenance is recorded for every artifact. A pre-registered token meter reports the compute cost. Human time is floor-bounded, so the human contribution is measured conservatively rather than generously. The error ledger's catch numbering runs to #256 — a monotone counter over the mathematics campaign's append-only flags ledger, maintained 2026-07-07 to 2026-07-20, with one number (#79) never assigned and later catches recorded un-numbered. Against that ledger of catches, the number of incorrect proofs reaching the record is zero. The verification chain held from the Lean 4 kernel to SAT-checked equivalence at the silicon boundary. The result is not a demonstration that verification is cheap in the abstract; it is a measured report that, at AI speed, ke

For sixty years, machine verification has been treated as a major cost overhead — affordable only for exceptional artifacts, and rarely for the ordinary software and hardware that industry actually ships. The premise of this work is that generative AI inverts that relationship. At AI speed, machine verification is not merely economical; it is essential to productivity. It functions as the incorruptible referee that lets one person safely direct autonomous machine work at scale.

The reported campaign is deliberately extreme: in five weeks, one researcher on consumer AI subscriptions directed a small fleet of AI agents from application code, through a verified compiler and executive, down to a RISC-V processor taped out on a community silicon shuttle. No proof passed through human review, and no RTL was written by a human. The organizing discipline — the **Salt method** — rests on a proof kernel that no hallucinated proof can pass. Mathematical claims travel between agents as kernel-checked artifacts, and human attention is reserved for statements, designs, and rulings. Verification is stated link by link, from the Lean 4 kernel to SAT-checked equivalence at the silicon boundary.

Why it matters

The significance of the result lies in the inversion it describes. For sixty years, machine verification was a cost overhead affordable only for exceptional artifacts. If generative AI makes verification economical at AI speed, then verification stops being a luxury and becomes the productivity mechanism itself: the incorruptible referee that lets one person safely direct autonomous machine work at scale. The Salt method is one concrete discipline for operating under that referee, with a proof kernel no hallucinated proof can pass. The trust boundary is the key architectural choice. By requiring that mathematical claims travel between agents as kernel-checked artifacts, the method removes natural-language assertion from the critical path. Human attention is reserved for statements, designs, and rulings — the places where judgment is irreplaceable — while the kernel adjudicates the rest. The error ledger's monotone catch counter to #256, with zero incorrect proofs reaching the record, is evidence that the boundary held under sustained adversarial pressure from the agents' own proposals. The result also has a silicon dimension. A RISC-V processor taped out on a community silicon shuttle, with no human-written RTL and SAT-checked equivalence at the boundary, suggests that the verified-compiler-and-executive path can be carried all the way to physical hardware. The taxonomy candidates for this work — Architecture, Cybersecurity, Network, Cryptography — reflect where kernel-checked equivalence and append-only provenance ledgers are most likely to matter next: anywhere a trust boundary must be stated link by link and defended without human review of every proof.

Who should read this

CS practitioners and researchers

Opening member content…