Skip to account-risk research
Risk Management for TradersUnits / Funding / Scenario limits

How to inspect a risk system

Trace a risk decision from account inputs to order state, failure behaviour and deployed cloud evidence before trusting a control.

Original edition: 10 October 2026. Expanded edition: 13 October 2026.

Educational risk arithmetic, not personal sizing advice. Stops do not guarantee fills. Stress inputs are invented assumptions without assigned probabilities; margin, shorts and derivatives need their own contracts.

Write the control's authority and operating lane

Start by asking what the system can do. A dashboard that displays a proposed size is different from a control that can reject an order. A notification service may not be able to cancel an outstanding instruction or close existing exposure. Name the account, supported instruments, decision owner and permitted actions. Describe which control runs before submission and which reconciles state after execution.

Keep the policy document, software implementation, deployment and subscriber route as distinct evidence objects. An endpoint inventory or research status labelled live does not prove that a particular account receives that version. For a cloud-owned production control, request the deployed revision, active route and durable execution receipt. This guide describes an inspection method, not a test of an unnamed vendor's private account. No assumed access to its broker or production infrastructure is claimed.

Inspect the input contract and its freshness

Trace account cash, marked equity, positions, pending orders and instrument specifications to identifiable sources. Save both observation time and receipt time where available. An account snapshot before a partial fill cannot be treated as the state after that fill. A last price from one venue may not be the executable or valuation price for the account's instrument.

Define what happens when an input is missing, stale, conflicting or unsupported. The exact fail-closed or bounded-action policy requires authorisation and context; do not silently assume zero exposure. A worksheet should clear obsolete results after invalid input. An operating control needs a documented recovery path and observable error status. Verify units, currency, minimum increments and funding rules before checking any limit percentage. Missing identity cannot be repaired by giving a numeric field a plausible default.

Follow one fictional request through the decision

Consider an invented cash-funded long-share request with starting cash 10,000, entry 50, stop 48, extra per-share cost 0.10 and declared stop budget 100. Forty-seven whole shares have planned loss 98.70 and conservative entry funding plus all entered extra costs of 2,354.70. A request for forty-eight fails the same budget even though both funding amounts fit inside the starting cash.

Save requested and accepted quantities separately, together with the reason for rejection or reduction. Check existing holdings and outstanding orders independently; this one-request arithmetic does not establish the joint account position. A passing synthetic request verifies a branch of the implementation. It is not an actual broker submission or evidence that the policy's budget suits a person. Preserve the fixture outside market-performance reports so an accounting test cannot become a fabricated operating receipt.

Exercise failures as well as the clean path

A useful test matrix includes a blank point value, unsupported increment, zero or negative price, stale account snapshot, duplicate instruction and partial-fill update. Define the expected outcome before running each fixture. Test a cancellation request followed by a fill arriving before its acknowledgement. Confirm the state journal can represent both events rather than deleting the fill because a cancel button was clicked.

Inspect retry and idempotency behaviour so one instruction cannot become several orders after a timeout. An unavailable price should have a named response, not a reused earlier green result. Test recovery with the original request identifier and retained state. These are software correctness checks, not sampled losses or independent strategy-validation observations. A good test outcome demonstrates that declared failure behaviour is implemented; it does not establish how often real-world failures occur.

Reconcile reports against actual orders and positions

An accepted instruction needs a later disposition: rejected by the broker, pending, partially filled, filled, cancelled or unresolved. Reconcile cumulative executed quantity and open quantity against the authoritative account reports. A requested reduction does not reduce exposure until the corresponding execution is established. Keep transaction identifiers and timing so a reviewer can identify mismatches without trusting the dashboard summary.

Separate alert delivery from order execution. A delivered risk notification does not prove that someone acted, while a failed notification should not erase an order already enforced by a separate control. For deployed services, normal operation must not depend on a personal computer, desktop tunnel or closed local agent session. Ask for cloud ownership and durable artifact evidence where that operating claim is made. Local research can test implementation, but cannot substitute for production-route proof.

Produce a layered verdict instead of one trust badge

Report input coverage, rule correctness, failure behaviour, deployment identity, route identity and reconciliation separately. Link a pass or unresolved finding to the relevant artifact, version and date. A tested calculation may coexist with an unverified deployment; a verified deployment may still have an incomplete input contract. Keep those distinctions visible so the buyer understands the practical consequence of each gap.

The downloadable record captures the proposed decision and the evidence needed to trace it. Use the synthetic size example as a repeatable fixture, then request actual operating receipts before claiming that a live system is enforcing controls. Do not infer a vendor's recovery procedures or regulatory compliance from its marketing terms. This page makes no certified-control or account-safety claim. An exact unresolved request is the appropriate output when the evidence required for a stronger verdict has not been supplied.

Review checkpoints

Evidence requests, not completed study results
ObjectPreserveDo not infer
AuthorityDisplay, rejection and execution powersA warning can cancel an order
DecisionFresh inputs and requested/accepted quantitySynthetic success is a live broker receipt
OperationCloud revision, route and reconciliationResearch LIVE proves deployed enforcement

Download blank review record (CSV)

Headers only. No trades, subscriber outcomes or operating receipts are supplied.

Source documents

Documentation establishes the stated mechanism, not tested trading performance. Checked 10 October 2026; a later version can differ.

Detailed research

Joint gap risk and shared exposure

Compare planned stop loss with disclosed hypothetical account shocks without pretending to estimate correlation, tail probability or a guaranteed maximum.

A correctness fixture and a financial test differ

The new reference uses Decimal to reproduce fixed cash rows independently of the browser implementation. Agreement checks arithmetic, not financial adequacy or strategy advantage. All original review checkpoints remain. A stronger system finding needs actual order states, marks, quantities, costs, funding terms and policy versions. A model-validation sample needs its own independent periods and source availability. A precise local fixture should not be counted as observed market evidence or used to reopen an exit tournament after evaluation.

A filled source request

In an invented review the quantity arithmetic is supported but actual exit availability and financing are absent. The finding is correct planned calculation, unverified execution and unresolved account terms. No score is created from those partial states. The workbook may show an assumed gap loss larger than the planned stop reference while giving no frequency for it. The source request names the missing account and policy records instead of manufacturing fills or an aggregation formula to complete the story.