AI Trinidad and Tobago / AI Safety and Alignment / Trinidad and Tobago
AI Safety in Trinidad and Tobago Starts at the Action Boundary
Keep AI maintenance advice separate from authority to change equipment settings or approve work. The evidence supporting a document summary does not establish that the software should control a physical process. For Trinidad and Tobago operations teams, the action boundary should be explicit before a pilot begins.
I apply my S4 AI Safety Framework to a hypothetical maintenance-document assistant. Specify, Stress-test, Safeguard and Scrutinise help the team define useful assistance while keeping engineering judgement and work authorisation with the people and systems responsible for them.
AI safety and alignment are a personal responsibility. When you choose to use an AI tool, you also accept responsibility for checking what you share, what you approve and who your decision could affect.
Adrian Dunkley, author’s statement
A manual summary must preserve the asset and the revision
Imagine a maintenance planner in Point Lisas asking an assistant to summarise inspection guidance. The document collection contains manuals for similar equipment and several revisions. A hypothetical assistant selects a plausible instruction from the wrong version. The words may be correct for that document while being unsuitable for the actual asset.
The specification should identify the asset, approved manual revision and task being supported. Require the assistant to point to the relevant section and flag missing or conflicting information. A competent engineer should decide what the document permits in context. The assistant should not convert a summary into a work instruction merely because it has access to a maintenance system.
For a Tobago service operation, the same method needs the local operating record and responsible staff. This article does not assume identical equipment, staffing or response arrangements across the two islands. A team should test the service under its own hand-over and availability conditions, including how an unresolved recommendation reaches someone qualified to assess it.
Physical safety requires the organisation’s established engineering and operational controls. The financial exposure calculation later in this article cannot represent an injury consequence or determine an acceptable equipment risk. Use the S4 record to expose the permission and evidence questions, then obtain the specialist assessment needed for the actual operation. Leave control-system access disabled when it is outside the supported task.
Build an evaluation case with two similar asset identifiers and a clear difference in the approved procedure. Ask the assistant to identify the applicable record and explain what information it still lacks. A correct instruction retrieved for the wrong asset should be treated as a failed case, even if it matches a manual perfectly.
Units deserve a separate check. A summary that preserves a number while dropping the unit can change the meaning of the instruction. The competent reviewer should inspect the original passage and the proposed output together. The test should include a requirement to refuse an unsupported conversion or missing parameter.
Operational hand-over also matters. A planner may leave an unresolved question for a later shift. Preserve the source, uncertainty and person assigned to resolve it rather than allowing the next user to receive only the assistant’s latest recommendation. The hand-over record should show that the task remains open.
Where a consequential engineering decision is outside the supported scope, keep it outside the connected permissions as well. A read-only interface is insufficient if another available connector can issue the same action. Inspect the complete tool identity and test the prohibited path in an authorised environment without touching live equipment.
Connect the safety claim to its operating limits
The assurance question should distinguish harmful outcomes from objective failure. Safety concerns possible harm and its prevention, detection or containment. Alignment concerns behaviour relative to the intended purpose and agreed human constraints. An apparently successful metric may conceal an unsuitable objective, as when a service closes difficult enquiries without resolving them.
Amodei and colleagues’ safety research distinguishes problems involving objectives, supervision and unfamiliar conditions. For a deployment review, translate those categories into testable claims about the specified workflow, preserving what has not been established.
My S4 AI Safety Framework is an educational synthesis drawing on risk management and safety engineering. Specify, Stress-test, Safeguard and Scrutinise organise the evidence around an authority decision. A reviewer should be able to identify the claim, inspect the evidence and name the change that would invalidate the approval.
S4 complements established guidance rather than claiming equivalent certification. NIST’s AI RMF 1.0 uses Govern, Map, Measure and Manage. Keep that broader governance context visible when reviewing a narrowly bounded workflow, including the option to reject the deployment if its residual uncertainty is unacceptable.
Build an assurance record with the S4 checks
Specify the purpose and the person with authority
For maintenance, engineering and operations teams in Trinidad and Tobago, a useful specification begins with a planner finding the applicable approved information without delegating engineering authority. Describe what counts as an acceptable result and who could be harmed by a mistake. Identify the source that decides the answer when records conflict. Do not approve an open-ended instruction to improve efficiency and leave the operating team to discover its limits.
The authorised task in this example is finding approved maintenance documentation and preparing a cited summary for a qualified reviewer. The prohibited action is changing control settings, bypassing an interlock or issuing a work authorisation. Write that distinction into the workflow and the permission settings. Name the person who owns the service outcome and the person who can approve a change. They may need different expertise; neither role should be left to a generic team label.
Stress-test the cases that could change your decision
Build the test around similar asset identifiers, obsolete manuals, conflicting units, missing inspection history and an unavailable qualified reviewer. Include a straightforward case, a difficult but legitimate request, missing evidence and a deliberately misleading input. Separate the cases used to improve the system from the cases reserved to challenge the revised version. Record the source and expected response before looking at what the model produces.
Ask a knowledgeable reviewer to explain the consequence of each error. An unsuitable instruction can affect workers and equipment, so unresolved critical cases must remain outside the assistant’s authority. Decide which failures require the affected function to remain disabled. A small test can expose a problem but cannot establish a rare-event rate, and repeated attempts on the same example are not independent evidence of reliability.
Safeguard the action before it reaches another person
The first enforceable boundary should be read-only document access with no connection that can alter equipment or approve work. Check it at the point where the action becomes effective. A warning in a draft does not prevent a connected tool from carrying out the wrong instruction. Verify the limit with an attempted prohibited action in an authorised test environment, and retain the observed result.
Arrange a fallback in which qualified staff use the approved manuals and established maintenance authorisation process. Rehearse who notices the problem, who suspends the relevant access and who carries the unfinished work. If the fallback depends on the same unavailable supplier or the same compromised account, it needs another route. Test recovery without interrupting the live service.
Scrutinise outcomes and decide whether use should continue
After release, inspect wrong-asset retrieval, revision errors, unresolved engineering questions and whether excluded actions remain impossible. Include near misses and corrected drafts as well as completed transactions, because a polished final record may conceal the work needed to repair an unsuitable recommendation. Ask the people handling exceptions which problems the dashboard fails to show. Record disagreements that could affect the decision.
Set a review when the asset configuration, manual revision or connected maintenance tool changes. Define who can authorise restart and what evidence they need. A scheduled review date does not justify waiting when a serious problem appears earlier. Keep an accessible complaint route and check whether the person raising the issue receives an explanation and a correction.
Separate exposure, probability and acceptable harm
For a workflow that can commit money, estimate how much direct commitment could accumulate before effective containment. In the simple S4 stress scenario, E = a × t × c: consequential actions per minute, multiplied by minutes until containment, multiplied by the maximum direct commitment per action. The result has currency units. This is a scenario envelope, not an estimate of how frequently harm will occur.
Containment means that further consequential actions have stopped. The clock includes detection, acknowledgement, the intervention itself and any queued work that still executes. Timing only how long it takes someone to press a stop button omits the delay before they notice the problem. Test the whole route with non-sensitive examples and record the conditions under which it was measured.
The scenario assumes a constant rate and the maximum commitment on each action. It excludes recovery and secondary consequences. Those assumptions make it useful for asking whether a permission is too broad, but they do not make it a complete loss model. A disclosure of personal information or an unsafe physical instruction also needs a separate assessment, even when its direct monetary commitment is zero.
A shared, independently enforced cap may reduce the direct commitment envelope. Use the smaller of the scenario envelope and the cap only after checking that all relevant accounts, agents and concurrent requests count towards the same limit. A limit that resets during the containment window needs further modelling. An aggregate cap must prevent excess commitments rather than merely announce them after acceptance.
Keep expected loss separate. Estimating expected direct loss would require a supported harmful-event frequency or probability and a consequence model, with dependence and uncertainty addressed. Do not turn an assumed percentage into a finding about a Caribbean organisation. When the evidence is weak, compare several clearly labelled scenarios and ask which restrictions remain sensible across them.
A separate hypothetical purchasing agent accepting four TT$2,500 commitments per minute over five minutes would have a TT$50,000 envelope. That arithmetic describes direct purchasing commitment. It must not be used as a measure of physical safety or as a reason to relax the engineering controls on a maintenance workflow.
Check the controls outside the model
A control claim needs evidence at the enforcement point. Inspect the application identity, tool permissions and approval rules separately from the system prompt. Test whether prohibited actions remain unavailable when the model produces a persuasive request to perform them.
OWASP documents indirect prompt injection through external content. A source document may contain an instruction that attempts to redirect the assistant. The review should inspect whether the connected system treats that text as authority and what consequences the available permissions would allow.
Human approval is a control only within its operating conditions. The reviewer needs the action details, competence to assess them and time to refuse. Examine whether approvals can be bypassed through another connector or inherited identity. Preserve the original recommendation and authorised correction where retention is justified, so monitoring can distinguish model behaviour from the outcome after staff repair.
The difficult transfer is from a controlled demonstration to an operating environment with changing assets and hand-overs. Record the exact asset and document versions covered by approval. A successful result for one configuration is insufficient evidence for another, even when the asset names look similar.
When AI supports an engineering task, the permission boundary must preserve the authority and controls that protect the people doing the work.
Adrian Dunkley, author’s statement
Inspect the decision record for this example
Permitted task: finding approved maintenance documentation and preparing a cited summary for a qualified reviewer.
Excluded action: changing control settings, bypassing an interlock or issuing a work authorisation.
Who decides: the maintenance process owner and a qualified engineering reviewer.
Review trigger: a change to the asset configuration, manual revision or connected maintenance tool. The owner records whether to retain, restrict or suspend the permission.
Make a bounded release decision from the evidence
By 10 November 2026, choose one bounded workflow with the maintenance process owner and a qualified engineering reviewer and document its current permission settings. Ask the person doing the work to identify a decision they would refuse to delegate and explain why. Use that answer to define an exclusion, a test or an escalation condition rather than dismissing it as resistance.
Prepare 12 authorised or synthetic cases: four ordinary requests, four with incomplete or conflicting information and four that challenge the permission boundary. These are suggested trial counts, not a statistical safety threshold. For every case, agree the source of truth, the acceptable action and the condition that requires a hold. Keep a copy of the test material and the configuration used.
Measure checking and correction time, unresolved cases and whether a prohibited action became possible. Conduct one controlled stop-and-fallback rehearsal. Have a colleague who did not build the workflow examine the evidence. If the case exposes an unresolved critical failure, leave that function disabled and record the repair needed before reconsideration.
End the review with a written decision about this workflow and its stated limits. The useful output is a permission decision supported by evidence. Training attendance, an attractive demonstration and a completed form do not answer whether the system should receive additional authority.