The automated agent's order lifecycle was cross-referenced against roughly 120 sources: the regulatory canon for automated trading controls, order-lifecycle engineering standards, and human-in-the-loop automation research. The review rated it the strongest of the three audits, found the lifecycle independently reimplementing much of what institutional standards require, and named the gaps plainly: the kill switch, per-order collars, and partial fills. The adopt-now tier shipped in v3.98.0.
The third audit took the automated agent's full order lifecycle, from staged intent to reconciled position, and cross-referenced it against roughly 120 sources of a different kind: the regulatory canon for automated trading risk controls (SEC Rule 15c3-5, MiFID II RTS 6, FINRA guidance, the FIA best-practice guides, and the SEC's Knight Capital order), the FIX-era order-state engineering standards, functional safety practice, and the human-in-the-loop automation literature. None of those regulations bind a retail customer trading through a broker; they are used as what they have become, the distilled, incident-tested consensus on what automated trading systems need.
The review rated this the strongest of the three audits. Several of the lifecycle's mechanics turn out to reimplement, sometimes verbatim, what the institutional canon requires: the intent hash is a faithful duplicate-order control of the kind 15c3-5 mandates; the fill-race guard on cancels matches the FIX "too late to cancel" semantics; day-only orders mirror the reasoning that led the NYSE to abolish native good-til-canceled orders; the mode ladder from shadow to semi-automatic to paper-only full automation is a private-scale version of the phased-deployment regime RTS 6 prescribes; and every-cycle reconciliation that blocks trading on any break exceeds the published baseline and directly implements the Knight Capital lesson.
The human-oversight design fared equally well. Typed production confirmation is the "hard stop" intervention class, the only approval-friction design with strong effect sizes in the decision-support literature (passive alerts are overridden at rates above 90%), and the assistant's stage-but-never-execute boundary matches where regulator thinking on AI agents landed: an accountable human for every order.
v3.98.0 shipped the adopt-now tier within days: per-order economic collars on the manual path, including a notional ceiling and a price-versus-mark deviation check; query-before-record on ambiguous order submissions, so a timeout is resolved by asking the broker rather than by assuming rejection; a daily positive heartbeat digest, because an alerting system whose silence is indistinguishable from its failure is itself a hazard (a lesson taken from process-industry alarm standards); a pre-breach loss warning at half the daily cap; and a written incident process with a manual kill drill. The remaining items, including cancel-working-orders-on-trip and first-class partial-fill quantity, are tracked in a review backlog with explicit tripwires rather than silently dropped.
Publishing an audit means publishing its praise as well as its findings. The review rated as ahead of documented practice: the verified reconciliation pass that refuses to treat an unreadable brokerage account as an empty one; the force-close discipline that requires independent verification of zero on every leg and never fabricates a profit-and-loss number; the safe-state asymmetry set, in which the machine-learning gate can only subtract autonomy and missing evidence always halts; and the honest simulated plane, which refuses to fabricate a fill rather than inventing one.
The complete bibliography is in the full review: Lifecycle vs. the Standards (PDF). Regulatory texts are used as design templates, not compliance obligations, for a retail system. Research commentary, not investment or legal advice.