DEOS Computing
Technical papers on the DEOS architecture and the systems built on top of it, authored by Jean N Pijeau.
A shared sequencer for Ethereum L2s built on the DEOS kernel. Every state transition is logged to a per-CPU event ring and made provable through a three-tier ZK proof pipeline that lands on-chain as a Groth16 verification. Ordering is decided by a hybrid VDF before transaction contents are decryptable, which removes content-based MEV.
A working primitive for LLM inference where the kernel emits a cryptographic certificate identifying the prompt, weights, tokenizer, architecture, and output bytes of every run. Any third party with the public weights can re-execute and confirm the certificate is honest, without trusting the operator who produced it.
A runtime for AI agents that logs every decision, API call, and data access into a tamper-evident execution trace. The trace can be replayed and verified by an outside party, which matters in settings where the operator's own log files would not be accepted as evidence.
A verifiable computation layer for blockchain applications. Where typical oracle networks run the same job on many nodes for redundancy, DCO runs it once and produces a proof of correctness via deterministic replay, with an optimistic fraud-proof fallback. The paper reports roughly an order of magnitude lower cost at equivalent security.
The formal argument that the DEOS kernel produces identical execution traces on replay across multi-core hardware, including the cryptographic and architectural conditions under which the claim holds. Measured runs on Kernel v0.2.0 produce hash-equivalent output.
The executive-level proof that DEOS replays userspace execution bit-for-bit. Walks through the kernel mechanisms that eliminate each classical source of nondeterminism, with the test methodology used to demonstrate the property.
The formal model behind the determinism claims. Separates two distinct guarantees, replay determinism on the same host and cross-host reproducibility, gives the conditions each requires, and lists the failure modes that fall outside scope.
How DEOS gets bit-identical floating-point results across CPUs, compilers, and machines. Covers IEEE 754 mode pinning, SIMD reduction order, FMA contraction policy, and the compiler flags that make AI inference outputs reproducible across hosts.
The DEOS-native AI model format and runtime contract. How models are packaged, addressed, and run on the deterministic kernel so every inference is replayable from a content hash, with prompt, weights, tokenizer, architecture, and output bound together.
A reputation model rooted in DEOS substrate receipts rather than self-reported scores. Each claim about an actor's behavior points back to a signed execution receipt that records what actually happened, so reputation can be checked against evidence instead of negotiated.