Architecture Before Guardrails
The same properties that inspire discussions about deterministic execution and formal verification must be rigorously applied throughout the coordination layers of the stack.
Local deterministic execution is necessary, but it does not solve multiparty coordination.
Soon, high-stakes environments will be dominated by AI agents, machines, and robots, outnumbering humans by orders of magnitude.
Architecture
Left: requirements—what the stack must provide. Right: a clean architecture on Zenon / Project Zeno. Rows align by layer.
Separate notary order from content-rich execution. No third-party bookies.
Pillars notary the order. Momentums = global order + finality, clean of content.
Policies, permissions, and audit rules. Decides which combinations of determinism and verification a task is allowed to use.
Pillars assign the multisig admin key; they do not perform protocol maintenance directly. Changes are public before activation.
Independently checks a claim. No discretionary execution. Does not author order.
Settlement + watchers. Re-run defined transitions; do not author order.
Detect failure and return to a known-good state. When bounds are exceeded: stop, refuse, or signal—do not guess.
Safe refusal / fail-closed. When bounds are exceeded: stop, refuse, or emit an explicit signal—do not guess.
Runtime for actions. Committed path: defined transition, or fail.
Domains + executors. Committed path: domain state transitions. Upstream modes OK.
Interfaces, contracts, side-effect declarations, permissions, and throughput metering.
Domain services with contracts, side-effect limits, and metered host bounds.
Versioned state checkable against a root without scanning the full store.
Account-chains + Sparse Merkle Trees, not the full global store.
Execution modes
Creative and predictive layers may plan and propose—those modes are permissible. The committed path—money, safety, legal commitment, medical dosing, infrastructure, multi-party coordination—needs a defined transition, or fail, as a first-class option, not bolted on with prompting.
You can loosen constraints upstream. What you cannot do is impose coordinated deterministic execution on a stack that already fused order and content-rich execution. That degree of definition was foreclosed by the architecture, not by a missing prompt.
Mission-critical systems need architecture, before guardrails.
Defined transition, or fail. Reproducibility and watcher re-runs are downstream of that definition—same inputs, same state, same tools can only re-run what was already specified. No silent failures or improvisation.
Coordination, not just local determinism
If defined transitions matter in local environments, arbitrary EVM/MEV style execution at the coordination layer cannot logically suffice. Those designs fuse ordering and content-rich execution: the same operator that decides order also exercises discretion. Ordering becomes a market. Conflict of interest is native, not accidental.
In an MEV environment, a watcher can still say “given this sequence, the transition is correct.” That is audit after the fact—it never answers who got to sequence the world, and why. Perfect audit under a captured ordering path is still a trust-dependent coordination regime.
Coordinated deterministic execution between parties requires environments that minimize third-party interference by separating concerns. Consensus orders. Executors compute.