Marty McEnroe

Palaestra Live

The standards roles above turn on a question that is hard to answer from a document: what actually happens at the interface between the grid and a very large load, when the load is run by someone with an incentive. Palaestra puts a person in that seat. You run a data center for a simulated day — buying power, working a battery, curtailing load, answering events as they arrive — and are scored against a certified optimal line. Where players consistently bend a technical rule is evidence about what that rule should say, which is the whole reason for building an instrument rather than a simulator.

The answers only mean something if the scoring is beyond argument, so the platform is built for that. Scenarios are data rather than code: a new one is a compiled pack, not a deployment. Heavy derivation happens offline in a compiler; the runtime engine is deterministic bookkeeping. Every published optimum is solved offline by mixed-integer linear programming and then replayed through the same engine that scores play, so the target is engine-verified rather than asserted. Re-scoring a stored action log must reproduce the stored score exactly, and that is enforced by tests rather than by intention.

Six scenarios are live and playable now. How the engine works, how scoring and stars are computed, and the interface every player and every agent speaks are documented in the open, alongside the authentication, network, and privacy surfaces.

AssemblyZero Active

Production multi-agent orchestration framework. Five-stage LangGraph StateGraph pipeline (Triage→Design→Spec→TDD→PR) coordinating 12+ concurrent Claude and Gemini agents. Cross-vendor adversarial verification — Claude generates, Gemini reviews — catches hallucinations across model families. Five-layer evaluation framework: execution-based verification, AST structural analysis, cross-model review, stagnation detection, longitudinal learning. 740 issues processed, 88% closure rate, 434 PRs merged, 5,583 tests.