Mnemox Control

Mnemox Control is a broker-neutral Policy Evaluation Engine for agentic trading. It evaluates an untrusted order intent against an owner policy and sealed account, market, and instrument state, then returns deterministic proof-carrying evidence with 40 ordered rules.

This repository implements the v0.3 Policy Decision Point (PDP). It does not execute trades, hold broker credentials, authenticate policy owners, reserve exposure, or issue single-use order authorizations.

Evaluate a conformance input

import json
from datetime import datetime
from pathlib import Path

from mnemox_control import (
    InstrumentCatalog,
    MarketSnapshot,
    OrderIntent,
    PolicyBundle,
    TrustedAccountSnapshot,
    evaluate,
)

fixture = json.loads(Path("tests/conformance/v0.3/basic-allow.json").read_text(encoding="utf-8"))
inputs = fixture["inputs"]
result = evaluate(
    policy=PolicyBundle.model_validate(inputs["policy"]),
    intent=OrderIntent.model_validate(inputs["intent"]),
    account=TrustedAccountSnapshot.model_validate(inputs["account"]),
    market=MarketSnapshot.model_validate(inputs["market"]),
    instruments=InstrumentCatalog.model_validate(inputs["instruments"]),
    evaluated_at=datetime.fromisoformat(inputs["evaluated_at"].replace("Z", "+00:00")),
)

print(result.decision, result.content_hash)

An ALLOW result must not be sent directly to a broker. It is a static PDP result, not an execution authorization. A production Policy Enforcement Point must verify authority and revocation, atomically reserve capacity, issue and consume a single-use grant, submit idempotently, and reconcile broker truth.

What v0.3 guarantees

• Pure deterministic evaluation: no clock, storage, network, UUID, or randomness reads.
• Complete binding to policy, intent, account, market, instrument catalog, state version, and time.
• Full account-wide worst-case exposure; opposing pending orders never net.
• Exact Decimal quantities, prices, increments, notionals, and leverage.
• Known unsupported instruments deny instead of using an incorrect valuation formula.
• Strict monotonic reduce-only semantics with a data-driven exemption matrix.
• One structured result for every rule; dependent SKIP requires a prior DENY.
• DENY outranks ESCALATE, while all triggered evidence remains visible.
• Byte-identical Apache-licensed conformance vectors for independent implementations.

The full equations, rule order, temporal semantics, and conformance procedure are in docs/protocol/evaluation-v0.2.md. What v0.3 changes on top of it (a required require_protective_stop policy field and four protective-stop reason codes, 36 -> 40) is in docs/protocol/evaluation-v0.3.md. v0.3 is a breaking change: v0.2 policies and conformance vectors do not validate against a v0.3 engine.

Install

pip install mnemox-control

Python 3.12 or newer. The engine is AGPL-3.0-only; see Licensing below. The TradeMemory broker proxy (pip install "tradememory-protocol[proxy]") pulls it in as its policy engine.

Development

Mnemox Control requires Python 3.12 or newer.

python -m venv .venv
.venv\Scripts\python -m pip install -e ".[dev]"
.venv\Scripts\python -m pytest -q
.venv\Scripts\python -m ruff check .
.venv\Scripts\python -m mypy src
.venv\Scripts\python -m build

Hypothesis properties cover exposure monotonicity, pending-risk monotonicity, strict reductions, determinism, hash binding, decision precedence, and input immutability. Conformance tests compare the full result, canonical JSON, and SHA-256 byte for byte.

Security boundary

Content hashes detect mutation but do not prove issuer identity or current authority. v0.3 has no signature verification, policy registry, revocation lookup, atomic reservation, authorization grant, broker adapter, execution receipt, reconciliation service, or coverage proof. These are explicit future PEP/evidence layers, not implied capabilities of this kernel.

Do not place API keys, exchange secrets, broker sessions, or live-order access in this package.

DecisionReceipt remains importable for v0.1 compatibility but is deprecated. New integrations use EvaluationResult.

Licensing

• Python engine: AGPL-3.0-only, with separate commercial licensing available.
• Protocol specifications and conformance vectors: Apache License 2.0.
• Mnemox names and trademarks are excluded except for origin identification.

See LICENSE, COMMERCIAL-LICENSE.md, and docs/protocol/LICENSE.