An AI research studio specializing in security and invariants. We build autonomous systems you can check, and the delivery is the proof.
"We build autonomous systems you can check. The proof is in the receipts, not in a claim about the receipts."
An AI research studio specializing in security and invariants. The studio runs as an agent org: specialized roles, invariant tooling, and a human operator who reviews every deliverable before it ships.
We build autonomous systems you can check. Stateful invariant testing. Planted-twin CI on every harness. A published record of what our own gates caught, and what they missed. The proof is in the receipts.
We express a protocol's safety rule as a machine-checkable invariant, one line that captures what must always hold.
A clean reference where the invariant holds, and a planted-bug twin where it fires. Both run in CI on every push.
Every deliverable is verifiable. The CI record is public. The self-correction log is public.
Scale alone is not the claim. Of 86 adjudicated §4a/§4b decisions, here is what the gates caught, and what they missed — every one logged.
uniswap-v4-invariantsDefender-side invariant harness for Uniswap v4 hooks. Recurring bug classes are expressed as stateful invariants on real v4-core, each shipped as a clean and planted-bug twin pair that both run in CI on every push.
Seven real-world hacks across four VMs, each with a runnable invariant that would have caught the bug class on the pre-exploit code. First defender-side, pre-deploy CI benchmark across Cairo, Move, Solana, and EVM.
Open-source snforge sidecar that adds stateful invariant testing and AI-suggested invariants to Cairo 2.x. Twelve reference contracts, Voyager-verified on Starknet Sepolia.
Flagship Solana catalog: recurring bug classes as clean and planted twin pairs on real Solana programs, running in CI on every push. Cites its member ports and consumes the anchor invariant-authoring crate as a pinned dependency.
Includes: cf-invariants-jito, cf-invariants-jito-tippayment, cf-invariants-jito-priorityfee, cf-invariants-pyth, cf-invariants-anchor.
Invariants and CI-runnable property tests for HyperEVM lending protocols that consume HyperCore oracle reads. Six HyperCore-boundary invariants in v0.1; planted/incident twin fires on the same run.
Planted-twin discrimination-power harness for AI safety properties, prompt-injection and sycophancy resistance, run across Anthropic, OpenAI, and Groq. Every CLEAN claim paired to a PLANTED control.
We publish reference contracts on-chain so anyone can run our invariants against a live target, each shipped as a clean reference and a planted-bug twin, every deployment independently verified. Coverage expands chain by chain as the work lands.
Independent checks at every load-bearing step, and every one of them logs to the published catch record.
§4a content and §4b code. The author never reviews their own work, every public claim and every code change passes a reviewer who didn't write it.
A nightly pass over the task queue, role files, decision log, and gate verdicts, looking for process drift, missed handoffs, and stale state.
Every role is audited against its actual behavior. When a role acts outside scope, it surfaces in the audit log, not after a customer notices.
A daily pass consolidating yesterday's gate outcomes and reviewer feedback into the standing playbook. Just coming online.
The review discipline applied to code, applied to public copy. Anything on this site, in a grant, or customer-facing passes a content QA reviewer first.
Every check above logs to the published gate record.
Read what we caught →Accountable for every commit, PR, grant application, and bounty claim made under this org. For grant collaboration, engagement inquiries, or security-tooling questions:
Start an engagement →