A reward you cannot fake
Most agent benchmarks ask a model to grade the transcript. This one reconstructs the ledger. A checkout saga runs under escalating chaos, and afterwards an oracle asserts properties a payments system may never violate — so an agent that serves fast garbage fails, however plausible it looked.
The Immortal Daemon · reinforcement learning · chaos · GRPO
View source All worksThe saga under chaos
Every dot is an order moving through reserve → charge → confirm → fulfil. Break things with the chaos controls and watch the invariant panel: a passive operator lets violations accumulate, while a competent one restarts, drains and compensates. The invariants are checked against reconstructed state, not against what the service claims happened.
The five invariants are the ones the real oracle asserts: no oversell, no double-charge, no lost orders, no money without goods, and fulfilment recovery once the queue drains.
Why microVMs make the group cheap
GRPO needs a group of rollouts from one state to compute a relative advantage. Snapshot the running application and its emulated cloud, then fork the snapshot: every branch starts bit-identical, so the spread in returns is caused by the policy rather than by environment variance. Press fork.
Measured on the GB10: microVM boot ~92 ms, snapshot ~230 ms, and a group of 16 forks in ~470 ms amortised — negligible against a multi-second chaos episode. The timings shown animate those figures.
The Immortal Daemon
The application is an e-commerce checkout service running a saga against an AWS-compatible backend. A fault injector escalates through process kills, resource pressure, backend outages, network partitions and clock skew; the agent restarts, scales, drains, degrades and fails over to keep orders flowing. After each episode an invariant oracle reconstructs ground truth and asserts what a checkout system may never violate: no oversell, no double-charge, no lost orders, no money without goods, full fulfilment recovery. An agent that serves fast garbage fails the oracle.
- SubstrateCubeSandbox microVMs on GB10
- Fork cost~92 ms boot · ~230 ms snapshot · 16 clones in ~470 ms
- Cloudfakecloud — 19 MB Rust AWS emulator, ~47 MB RAM
- TrainingGRPO with group-relative advantages over identical forks
HPC Cluster Simulator
A drag-and-drop, gamified training environment for building and operating GPU clusters. Trainees design a cluster, run workloads against it, break it, and learn why it broke.
legal-financial-modernbert-150m
A 150M-parameter ModernBERT encoder trained from random init — no pretrained checkpoint, no distillation — on public legal and financial documents, then turned into a Matryoshka embedding model whose vectors truncate without retraining.
Populace
A world-scale persona simulation: grounded personas on a real globe living daily routines and reacting to events, with a language model authoring what each kind of person makes of the news and a Go engine doing the per-capita work.