The Thesis: Who Authorised This Action?
A Singapore portfolio manager's agent may be able to send an order to a broker in another ASEAN market. That capability leaves the investment committee with a harder question: who permitted this particular instruction, for which portfolio and instrument, under what limits, and did that permission still hold when the broker acted? SNN.NET's EIA-025 distinguishes the identity of an agent, delegation, authorisation at the point of action, execution and accountability. SNN.SG extends that inquiry to a market with several decision owners. It examines what each can demonstrate when a proposed transaction crosses institutions and borders.[20][19]
Start by identifying the action, not the system. Reading an inventory, generating a proposed order, instructing a broker and reallocating an executed fill can be four separate permissions. A general objective such as managing liquidity cannot resolve which of those operations the institution assigned to the agent. When a delegated task spans several tools, the evidence should allow the receiving owner to distinguish the authority to propose an action from the authority to change the institution's position.[20][11]
The crucial interval starts when an agent forms an instruction and ends when an independent receiving system acts on it. During that interval a mandate can be narrowed, an instrument can become ineligible, a person can withdraw an approval or a retry can recreate an order already accepted. A system that checks only at login may miss those changes. A receiver that checks only the signature may establish where the instruction came from while remaining unable to establish who had the power to issue it. If the receiver can check only part of the mandate, the missing part should remain explicit in the decision record.[11][12][18]
A valid instruction can also depend on facts outside the agent. The administrator might have suspended a portfolio, the issuer might have changed an instrument's terms, or the broker might be operating under a different version of the eligible-security list. A timestamp on the message records an event, but cannot reveal which facts the receiving control checked. For that reason the regional case preserves both a decision-time snapshot and what was learned later, instead of silently backdating a correction.[11][13]
For capital providers, the consequences can extend beyond technical access. A stale instruction may create an exposure the manager did not intend, a settlement obligation the operations team must investigate, or a statement the issuer or investor should qualify. Those possibilities are research questions, not evidence of a breach, loss, invalid trade or higher funding cost. A narrow, well documented decision to pause or refuse can be as useful as a successful action. The relevant evidence explains the scope of the instruction, the receiver's own decision and what actually happened thereafter.[14][16]
The article uses three levels of confidence. The cited institutions establish only the positions or practices within their own records. The scenario illustrates where evidence can fail to travel. The proposed comparison would measure whether better records improve reconstruction. None of those levels gives the editorial desk authority over a regulated decision. A negative finding may be that a record authenticates the sender while the institution's delegation remains unknown, even if the trade result can be observed.[14][20]
Global Evidence: Technical Design and Institutional Accountability
NIST's 27 October 2026 technical blog proposes ten design tenets. It distinguishes agent identities from human accounts, calls for delegation to be constrained and revocable, and separates the policy that authorises an action from the model that proposes it. It also proposes independently preserved, tamper-evident histories. These are initial design proposals. NIST's separate zero-trust architecture explains why access should not follow implicitly from network location or ownership. Neither document supplies a legal mandate or says that every action must receive a fresh human approval.[11][12]
One practical difference is who writes the audit record. If the proposing model can also edit the decision record, a reviewer may be unable to distinguish the originally checked facts from a later explanation. Independently held logs and a recorded policy version can make that distinction inspectable. A separate control need not be a new company or a central registry; in the proposed design it is a control whose decision and evidence are protected from unilateral alteration by the agent.[11][12]
The Financial Stability Board's final report of 15 October 2026 sets out twelve non-binding sound practices for financial institutions. Its concerns include board oversight, materiality, documentation, lifecycle controls and external dependencies. IOSCO's March 2025 consultation report adds a capital-markets lens by examining uses and risks across market participants, including issues for investor protection and market integrity. The UK ICO's call for evidence opened on 8 October 2026 and was scheduled to close on 20 November; it asks about accountability and data protection, among other matters. These three records have different remits and procedural status. They do not amount to one international rule on who can bind a fund.[14][16][15]
The three institutional signals also use different units of analysis. The FSB looks at financial institutions and governance across the lifecycle, IOSCO examines market uses and risks, and the ICO consultation asks how existing data-protection issues arise in agentic settings. Use the relevant document for the relevant inquiry. A survey response, a proposed tenet, a final non-binding report and a future-effective supervisory guideline cannot all populate one field labelled approved or compliant.[14][16][15][1]
There are useful but narrower technical tools. OAuth token exchange can describe subject, actor, audience and scope; the receiving institution still needs to know whether those values correspond to an authorised business action. A verifiable credential can make aspects of provenance and integrity checkable; a verifier still needs to assess the issuer, the assertion and its use. A sound control therefore keeps authentication, institutional delegation, runtime decision and actual execution separately observable. The suggested regional test asks whether that separation improves decisions in practice. It is not an assurance conclusion or an extension of the global methodology.[18][13][20]
A receiver may accept a reference to a signed delegation from a known issuer for one purpose and still require another control for execution. As an example, an access token limited to one account need not express an order ceiling or restriction on onward delegation. A credible design either carries those limits in attributable claims or obtains them from an authorised policy source. It also specifies what happens when neither route supplies a current answer.[18][13][11]
Singapore: Supervisory Expectations and Mandate Boundaries
On 7 October 2026 MAS issued its Guidelines on Artificial Intelligence Risk Management for Financial Institutions. They apply across financial institutions and forms of AI and set supervisory expectations proportionate to the scale and nature of use. MAS gives 7 October 2027 as the effective date. The immediate decision for a Singapore institution is how to map what an agent can do to its AI governance, use-case materiality and controls before that date. The Guidelines do not supply the trading limit, mandate holder or approval for an individual order; those come from the applicable arrangements and competent owners.[1]
An implementation inventory should identify the decisions at risk and their operators, including where a component only recommends an action and where an interface can make it effective. For a financial institution, a material use case should be connected to its actual risk owner and reviewed when powers change. Adding a send-order function to a previously advisory tool is a useful example: its business mandate and fallback procedure need attention even if the underlying model and user interface have not changed.[1][2]
IMDA's agentic AI framework gives organisations a practical way to bound use cases, define consequential human checkpoints and manage agents over their lifecycle. Its official summary identifies a May 2026 update covering third-party agents, multi-agent systems and automation bias. For a Singapore desk, a useful separation is between permission to gather research, permission to draft a ticket and permission to transmit a binding instruction. A limit change, an unfamiliar counterparty or a high-impact exception should be evaluated under the actual risk policy. This edition proposes review thresholds for a trial; it does not turn IMDA guidance into a universal checklist or a prohibition on bounded autonomy.[2][11]
A real checkpoint is assessed by what the reviewer receives and can change. The proposed order, policy version, affected portfolio and clear refusal or escalation path should be available before approval. A generic confirmation shown at login cannot establish that the reviewer considered the subsequent size, venue and expiry. An operator's ability to stop the action should be tested when the principal is unavailable and when the original platform experiences a disruption.[2][11][14]
Outsourcing and privacy are separate enquiries. MAS Notice 658 concerns outsourced relevant services of banks in Singapore, with its own scope and effective date. The PDPC advisory guidelines address use of personal data in AI recommendation and decision systems. A regulated entity should establish which role it occupies, whether the service and data are in scope, and what the applicable arrangements allow. A vendor's ability to read a portfolio file does not answer whether its agent may route a trade, retain a log abroad, or reuse customer data. A Singapore parent approval cannot answer those questions on behalf of every entity in the chain.[3][4]
The Singapore desk can frame separate vendor questions before onboarding: who operates the gateway, which subcontractors process which records, what evidence can be retrieved after a service disruption, and who can revoke permissions at the destination? These are editorial diligence questions. They do not presume that a particular contract is outsourced relevant service, or that an advisory source has resolved a lawful basis. A contract that allows audit access may still need a practical retrieval test.[3][4][14]

ASEAN: Common Records, Different Decision Authorities
The 2024 ASEAN AI governance guide is a voluntary regional reference, and the 2025 expanded guide adds generative AI considerations. They give teams a language for controls and human involvement, while national requirements and sector instruments retain their own scope. A group can therefore use a common template to identify the entity, system and receiving purpose without giving each answer the same legal meaning. A single policy marked approved at group level can otherwise hide a difference between the entity that supplies data, the entity that directs an order and the entity allowed to execute it.[6][7]
One proposed cross-border file uses a common core with local addenda. The core identifies the proposed action, sender, receiving entity, time and reference. The addenda hold the locally applicable policy, competent owner, version, access restrictions and unresolved questions. A common field can carry different answers. If a party cannot disclose its entire policy, it can provide a bounded assertion of what it checked, provided the receiver records the resulting limits on reliance.[6][7][13]
In Malaysia, the Securities Commission's revised Guidelines on Technology Risk Management took effect on 19 August 2024 and address technology and service-provider risks for the capital-market entities within their scope. In Indonesia, OJK's 29 April 2025 banking AI governance publication concerns responsible use in Indonesian banking, including advanced AI. A Malaysian broker and an Indonesian bank may both receive data for a Singapore-managed transaction, but each must determine its own duties and whether it can rely on the evidence offered. OJK's banking scope does not resolve the scope of a non-bank service provider; a Malaysian technology control does not validate the Indonesian bank's decision.[10][9]
Consider an operator that outsources order preparation to a shared vendor in one market, routes through a second market, and records assets with a third. A shared brand or system name does not show which licensed legal entity performs each function. The receiving file should identify the responsible counterparty and preserve who can issue an instruction, reject it, correct it or make a separate local determination. The record should not substitute the group's location for the legal entity's actual perimeter.[10][9][3]
The joint ASEAN–EU guide to model clauses provides a vocabulary for documenting international transfers of personal data. It does not determine whether the actual data, parties and purpose in an ASEAN transaction fit any particular transfer arrangement. A desk should separate transaction authority from access to customer, employee or account information. For example, a broker might need an account reference and a scoped authority assertion without receiving an internal committee minute containing personal details. A competent local owner decides what may be provided, on which basis and with what retention. Unavailable material remains unavailable; the receiver cannot record it as verified.[8][4][13]
A data hand-off can be deliberately narrower than a trade hand-off. The parties may exchange event identifiers and timestamps while making the underlying record available only to a designated reviewer. Whether this is sufficient depends on the actual data, transfer basis and recipient's decision. If reviewing a protected object requires authorisation not yet obtained, the correct description may be that the event was reported but the material record was inaccessible for this purpose.[8][4]
Capital Markets: A Bond Order After a Limit Change
Consider a hypothetical Singapore fund that invests in ASEAN infrastructure debt. Its manager allows an agent to prepare, but not independently approve, an order for a Malaysian ringgit bond through a Malaysian intermediary. An Indonesian bank handles a distinct cash or custody leg through its own arrangements. Suppose the investment committee caps the order and limits the eligible instrument, account and time window. At 10:01 the manager authorises the proposed size; at 10:03 a risk officer lowers the remaining limit. At 10:04 the agent retries after an ambiguous response. All times and facts are invented to test controls; this is not a named fund, transaction or market incident.[1][10][9]
Suppose the first transmission carried the original cap but the broker could not confirm whether it reached the venue. The lower cap at 10:03 changes the manager's decision to retry, but does not retrospectively cancel a fill. The broker may have recorded an earlier receipt and the manager may only receive the result after the retry. Reconciliation therefore requires the proposal, receipt, execution and observation times, and should make explicit which party could know which fact at each point.[11][16]
Before routing, the Singapore manager must determine which mandate revision and remaining exposure apply. The broker must decide under its own rules whether to accept the instruction, and its acknowledgement may differ from a fill. If the earlier attempt has filled in part, the retry requires reconciliation and may need a new decision. A valid credential and a message timestamp can coexist with an expired mandate or the wrong fund sleeve. An instruction blocked by the local broker must remain distinguishable from an executed trade with a later compliance concern. If a fill did occur, settlement and custodian entries are separate evidence; their existence does not decide how the manager should handle an exception.[11][18][16]
Consider three distinct endings. A timely refusal can prevent the duplicate while leaving the initial order unresolved. A partial fill can change available limits and require review of the unfilled balance. Two fills may require an exception process even if both are reflected in the custodian's ledger. The example does not determine cancellation rights, best execution, impairment or liability. Those depend on the instrument, applicable rules, agreements and decisions that the hypothetical does not supply.[10][5][16]
Project Guardian describes fixed-income tokenisation work and objectives around custody and settlement; Project Mandala demonstrates policy checks in a proof of concept. Those records motivate the interface design, but neither establishes an autonomous trading deployment or grants permission to rely on a foreign check. Whether a bond is tokenised or book-entry, the investment team should distinguish recommendation, order, fill, allocation, cash or securities movement and final position. The risk channels include duplicate exposure, unplanned liquidity use, servicing errors and correlated dependence on one gateway. No return, default, instrument eligibility, legal enforceability, settlement finality for the scenario or cost-of-capital effect is inferred.[5][17][14]
Capital-market diligence can use the same separation before investing in a manager or infrastructure provider. Ask which receiving decisions depend on the same vendor, how much exposure is able to accumulate before a stop, and whether a participant can reconcile independently when its normal feed fails. Documenting those questions can help an investment committee articulate risk. Any impact on valuation, liquidity buffers or spreads remains a separate judgement based on observed frequency, severity and contractual terms.[14][16][17]

Implementation: Preserve Decisions and Outcomes Separately
The decision file starts with a mandate that the business owner can explain without access to model internals. It should identify the delegating role, legal entity, account, permitted action, instrument and counterparty perimeter, quantitative caps, duration and revocation path. The technical control translates only the relevant parts into credentials and policy. At the receiving boundary, bind the check to the exact instruction, freshness of authority and the route used. An inability to resolve the current policy, owner or remaining limit should trigger a documented hold or escalation under the agreed rules. A model's confidence score is not a substitute for any of those records.[11][12][18]
A prospective control should distinguish a revoked mandate from a stale local copy and a destination that cannot determine status. Test all three when creating the authority record. For a quantitative cap, record the unit, currency, whether the amount is gross or net, exposure already consumed and the policy used to calculate it. A receiving control cannot interpret a nominal cap correctly if the decision depends on partial fills, currency conversion or another order not included in the reported amount.[11][18][1]
At each hand-off the reviewer needs the attributable decision and the evidence of what followed. A unique order reference and retry key help relate attempted actions; match them against broker acknowledgements, partial fills, cancellations and custodian entries before deriving exposure. Store UTC and the declared local zone where time matters, and distinguish when a permission changed from when the receiver learned of it. If a revocation reached one venue but not another, preserve both observations and the effective policy decision at each destination. A log should retain accepted, denied, escalated and unknown results and support a later reconstruction, with approved access, rather than claiming that a missing result means no order occurred.[11][14][5]
Use a narrow, practical reconciliation path for records that disagree. Operations can preserve the unmatched order, ask the broker for the authoritative state, link the result to the correct sleeve and document any amendment without erasing the original. When different systems encode cancelled or rejected differently, preserve the source values and the reason for a mapping. Otherwise the translation can manufacture a clean record and remove the exact exception an auditor or risk owner needs to investigate.[5][17][14]
The archive should be governed independently of the acting agent. A reviewer may use reference IDs, digests and authorised retrieval without copying all underlying personal or contractual data into every recipient system. A digest can help demonstrate integrity relative to an identified object; it does not prove that the object was truthful or sufficient for the intended decision. The same caution applies to evidence of market impact: an unmatched position may need investigation, but cannot establish financial harm without the relevant positions, terms and analysis. Preserve the original decision, revised evidence, disclosure boundary and the person responsible for reconsidering dependent conclusions.[13][4][8][3]
A protected record should be retrievable for authorised review without making the full archive available to every party. Set the minimum representation that can travel and a procedure for resolving a disputed representation. Keep an access decision and disclosure purpose alongside the event where relevant. When an earlier assessment changes, preserve its original scope and notify the appropriate receiving owner through the actual agreed process. Do not treat the newer evidence as though it were known at the earlier decision.[13][4][8]
SNN.SG Perspective: Four Actions, Twelve Evidence Objects
SNN.SG proposes four timed actions. Before enabling live order transmission, the Singapore mandate owner and local intermediaries document exactly which decisions are theirs and test the agreed limit and refusal paths. At each material instruction, the policy owner verifies current mandate and decision scope independently of the model and escalates uncertainty to the designated person. At execution and settlement, operations reconciles orders, fills, cancellations and positions across the receiving systems, including unmatched messages. After a material incident or the first review cycle, an independently appointed reviewer reconstructs a successful case, a rejected case and an unresolved case, and records whether each owner revisited its decision.[11][1][10][9]
Define the role of the independent reviewer before the trial. Give that reviewer access to the same authorised evidence on both paths and do not allow the executing agent to select the examples or grade them. The review order can be varied to reduce learning effects. Log each disputed reference judgement, the competent decision owner and the reason the evidence did or did not suffice. A case may legitimately support a limited answer rather than a verdict about the whole transaction.[14][2][20]
The trial retains twelve objects: the legal-entity, account and accountable-role map; the signed mandate and its scope; the agent identity and delegation chain; the current policy and limit snapshot; the exact order proposal and idempotency reference; the event-time authorisation or refusal record; the approval, escalation and revocation trail; the broker acknowledgement, fill and cancellation sequence; the custody and cash reconciliation; the jurisdiction and data-access schedule; the independently retained event and integrity log; and the correction and downstream reconsideration record. These are objects to locate in existing lawful records where possible. Their existence or completeness does not convert a historical decision into a current mandate.[18][13][4][8]
The event-time permission record should distinguish who made a business decision and which automated policy enforced it. The order proposal and idempotency reference should be capable of matching a rejected attempt as well as an accepted one. The dependency record should locate each use affected by a later correction, such as a risk limit report or a position statement. Separate owners should be able to reconstruct their own part of the sequence without treating the other institution's decision as their own.[11][18][13]
Use a proposed set of thirty decision episodes, chosen before scoring and including stale mandates, partial fills, replayed instructions, missing receipts and legitimate refusals. In a sandbox or historical replay with equal lawful access, vary whether reviewers start with the current process or the linked file. Have a separate reference panel preserve dissent and justified unknowns. Compare false permissions, false refusals, duplicate exposures, missed revocations, unexplained settlement exceptions, time to reconstruct and custodian effort. A safer result requires fewer material errors without unacceptable delay or overblocking. Thirty is a design suggestion, not a completed study, an assurance sample or a claim about any firm.[14][16][17]
Predefine what counts as an error before comparing approaches. Examples include approval based on an expired mandate, an avoidable false block, a duplicate that survived reconciliation and a confident assertion built on an unavailable source. Compare case by case before aggregating; a fast decision on a low-risk case does not offset a serious failure on a high-impact one. Report burden for each participant rather than averaging away local costs that a smaller broker or custodian cannot absorb.[14][16][6]
What to Watch: Reperformance and Evidence of Impact
The next observable change may be a policy limit update, revocation across two endpoints, a broker's rejection, a partial execution, a corrected position or a rule clarification. Each should be attached to the specific decision it can change. The October MAS Guidelines have a future effective date; the ICO inquiry has a separate procedural timetable. A new official record would justify an editorial update only for the proposition it actually changes. An independently recorded refusal is useful even if no transaction followed; an accepted instruction alone says little about what the account ultimately holds.[1][15][11]
Two useful records to watch are a revocation issued by the mandate owner and a time-stamped acknowledgement from each affected endpoint. They may not arrive together. Compare the scope and time of each and preserve instructions received in between. For a regulatory change, retain the document and its issue, effective and entity-scope fields, so an updated supervisory expectation does not get applied retrospectively to an unrelated foreign entity or become evidence of an approved individual transaction.[1][10][9][11]
A claim that the design reduces losses, improves liquidity or lowers the cost of capital would require evidence beyond a replay: comparable exposures and incidents, costs, time periods, risk and plausible alternative explanations. One common vendor could also create a correlated failure or an access barrier if smaller counterparties cannot produce a prescribed format. Fewer successful orders might indicate appropriate refusal or needless friction. The proposal weakens if it raises delays, conceals disagreement, displaces working controls or merely makes unauthorised execution easier to narrate.[14][16][6]
A second stage could examine financial effects only if the first stage shows that the method improves the underlying decisions. Review comparable portfolios or incidents, exposure and time, and account for changes in market conditions or trading activity. We would also look for adoption costs and concentration effects. If a simpler existing control gives the same reconstruction without requiring more data to move across borders, the preferred action may be to keep or improve that control.[14][16][8]
The evidence chain contains eighteen distinct external institutional records: five Singapore records, three ASEAN regional or joint records, two member-state records and eight international records. The .NET canonical thesis and SNN.SG EIS-027 are editorial lineage and excluded. A report and its landing page count once; the record describes what was actually read and the scope of each source. No actual transaction, mandate, local decision, delegated authority, control effectiveness or independent external validation is established. The question that matters is whether a competent receiver can explain why this specific action was permitted, refused or left unresolved at the time.[20][19]
The editorial conclusion is narrow enough to be revised. Better separation of records may make a disputed instruction easier to evaluate; it does not grant authority to the agent or to the person reconstructing the record. A reasoned refusal, an unresolved record and an executed transaction all belong in the history. The quality test is whether the organisation can point to the relevant owner, current mandate, scope and evidence for each outcome, including the cases in which those links cannot be established.[20][19]

