Home / Libraries / ML4T Live / Docs
ML4T Live
ML4T Live Documentation
Production trading with broker integrations
Skip to content

Public Claim Evidence

This register defines the public stable claims and the checks that support them. A passing source unit test alone is not enough when the claim concerns an installed artifact, sustained behavior, or an external paper account.

Area Public stable claim Evidence boundary
Portability Lifecycle version 1 callbacks and canonical intents are portable across the qualified backtest and live engines tests/contracts/test_causal_strategy_parity.py; minimum, locked, and maximum dependency profiles
Safety Canonical orders fail atomically; live safety may reject or reduce an intent without redefining it tests/unit/test_order_contract.py, tests/unit/test_safe_broker.py, tests/unit/test_secure_persistence.py
Feeds OKX and bar aggregation emit validated UTC events and halt on overload or continuity loss installed-wheel external OKX qualification; feed contract, queue, continuity, recovery, and stress tests
Brokers IB and Alpaca implement the documented snapshots, state transitions, canonical order input, capabilities, reconciliation, and six-hour paper-session contract broker contract and adapter suites; scripts/qualification/qualify_paper.py; fresh installed-wheel lifecycle and soak evidence
Performance The credential-free Linux reference workload sustains 100 events per second for one virtual hour with less than 25 MiB post-warmup RSS growth, no-op dispatch p99 below 10 ms, shutdown below 5 seconds, bounded queues, and exact event and intent checksums scripts/qualification/qualify_performance.py; tests/stress; candidate performance report
Platform The wheel and source distribution install with Python 3.12, 3.13, and 3.14 on Linux, macOS, and Windows; Python 3.15 prerelease installation and the non-hardware-dependent suite pass on all three operating systems before final support stable installed-artifact profiles; ecosystem qualification matrix
Maturity IB and Alpaca brokers plus the OKX feed are stable-supported; Alpaca, IB, DataBento, and generic CCXT feeds require experimental opt-in public claim scan, experimental-feed tests, retained six-hour OKX, IB, and Alpaca evidence tied to the candidate wheel hash

The complete credential-free command is documented in Candidate Qualification. Paper evidence is separate because untrusted pull requests never receive account credentials.