Documentation
How Nomylax evaluates an intent, why a block cannot be argued with, and what happens between approval and settlement.
The model
Nomylax sits between a human treasury and an autonomous agent. The human owns the capital. The agent proposes economic actions. Nomylax decides whether each one is financially permissible, and only then does anything settle. The agent may reason about money. It never holds unrestricted authority over money.
The Financial Constitution
Each agent has a constitution: maximum transaction, daily and monthly budgets, allowed assets, approved recipients, unknown recipient behaviour, emergency reserve, failure threshold, velocity threshold, risk threshold and permission expiry. Editing it appends a new version with a hash over the canonical policy. Nothing is replaced in place, so every decision can be traced to the exact rules that produced it.
Enforcement order
Checks run in a fixed order. A failed hard check blocks. A failed soft check is held for owner review. No model sits anywhere in this path: every check is a comparison against a value the owner wrote. A language model cannot approve a transaction that violates policy, because the model is never asked.
- 01Allowed asset
- 02Single transaction limit
- 03Daily budget
- 04Monthly budget
- 05Recipient allowlist
- 06Emergency reserve
- 07Agent state
- 08Failure threshold
- 09Spend velocity
- 10Risk threshold
- 11Transaction simulation
Nomylax Risk Score
Seven weighted signals produce a single value from 0 to 100: unknown recipient, policy deviation, amount anomaly, spend velocity, contract risk, liquidity risk and operational failures. Bands are LOW below 30, MODERATE to 54, HIGH to 74 and CRITICAL above. A weighted mean can let benign signals dilute a dangerous one, so a severity floor applies: one extreme signal floors the score at HIGH, two or more floor it at CRITICAL. Every score reports its components and any escalation applied.
Execution on Base
Simulation runs before signing and a failure stops the request. After confirmation the recipient balance is compared against the approved amount. A confirmed transaction whose state does not match is reported as unverified rather than as success. Financial operations fail closed: when the executor is not configured, the request is refused, never faked.
- 01Simulate
- 02Submit
- 03Confirm
- 04Verify state
Shadow Mode
An agent runs its real logic against real policy with simulated settlement. Budgets deplete on a copy, so limits behave exactly as they would in production while no capital moves. The report gives requested value, what would be approved, what would be blocked, violations, critical events, projected burn, and the point at which the agent would have entered Safe Mode.
Safe Mode
An agent moves itself down through the states AUTONOMOUS, WATCH and SAFE MODE on a critical risk event, on reaching its failure threshold, or on exceeding its velocity ceiling. It can never move itself back up. Recovery is an owner action. While Safe Mode is active, every request is refused before risk is even scored.
Connecting an agent
Implement one method. An adapter returns pending intents; Nomylax does the rest. The endpoint is registered against the agent in storage and validated on every call. A URL in a request body is ignored, so credentials cannot be redirected to a host of the caller’s choosing. Agents may also call the decision API directly and receive the same answer.
Authentication
A connected wallet address proves nothing on its own. The owner signs a server issued challenge, the nonce is consumed on first use so a captured signature cannot be replayed, and the server issues an http only session cookie. Nomylax never asks for a seed phrase and never holds an unrestricted private key in the browser.
Repository documentation
Architecture, security, threat model, risk model, economics, agent strategy, incident response, API reference, Base integration and deployment notes live in the docs/ directory of the repository.