GovTech on 2 October outlined a “context layer” that sits between government datasets and civil-service AI assistants, encoding agency definitions — such as what counts as a resolved service case — as version-controlled code rather than static policy PDFs.
The approach, described as semantics-as-code, is GovTech’s answer to hallucinated statistics in chatbots deployed across ministries. Instead of letting a model invent SQL, the system injects pre-tested queries and official definitions before any database call, so an officer asking about backlog size gets the same numerator the management dashboard uses.
Why definitions drift without code
Government data lakes often label fields identically across agencies while meaning different things. A “closed” complaint at one ministry may include auto-closures after inactivity; another requires officer verification. Large language models trained on generic text cannot infer which rule applies. GovTech’s write-up walks through a hypothetical case where a naive query overcounts resolved tickets; the context layer forces the agency’s verified definition and query template.
Officials said Practice, GovTech’s engineering group, is preparing proofs of concept in selected public-sector domains to test accuracy, traceability and token savings from sending only relevant definitions to models. Targets cited include higher answer accuracy and lower token use by trimming irrelevant schema text.
Link to scam-fighting tech
The same GovTech ecosystem houses Open Government Products, which runs ScamShield and the experimental Scam Vaccine simulations reported this week. Context-layer discipline for internal data is separate from public scam drills, but both reflect a push to make automated systems auditable. Scam Vaccine sends simulated phishing to volunteers; the context layer ensures internal AI does not misstate enforcement statistics when ministers ask for briefings.
Singapore’s Smart Nation narrative increasingly hinges on whether AI tools respect policy boundaries. Encoding definitions as code allows changelogs when rules shift — a practical requirement as ministries roll out copilots ahead of budget debates. For technology watchers, the launch is less a product splash than infrastructure: the plumbing that decides whether a civil servant can trust the number on the screen.
Procurement and audit trails
Government AI pilots must survive audits by the Smart Nation and Digital Government Group and by ministry chief information officers. Version-controlled definitions give auditors a Git history when a statistic quoted in Parliament differs from a chatbot answer. That matters as MPs submit more written questions expecting machine-readable data pulls.
Vendors selling copilots to agencies will need to integrate with the context layer rather than bypass it. GovTech’s blog post positions the layer as shared infrastructure, which could reduce duplicated definition work across ministries that built separate data dictionaries during COVID-19 reporting.
Privacy considerations remain: injecting definitions does not anonymize personal data. Officers still need role-based access controls on underlying tables. The context layer governs meaning, not permission. Training civil servants on when to trust generated summaries versus source dashboards will remain a change-management task long after the code ships.