VerityOS uses generative AI in a high-stakes reporting domain — so we govern it deliberately, by construction, and we say plainly where the edges are. Our AI governance aligns with Singapore’s IMDA Model AI Governance Framework and the international standards it draws on — the OECD AI Principles (2019, updated 2024) and the UNESCO Recommendation on the Ethics of AI (2021, 193 states). And this page isn’t just how we behave — it’s a product you get: every tier includes an AI Data Governance workspace to run the same discipline over your own AI systems.
Mapped to the Model AI Governance Framework for Generative AI (IMDA / AI Verify Foundation, 30 May 2024) and the human-oversight taxonomy of the Model AI Governance Framework, 2nd Edition (2020); aligned with the OECD Recommendation on AI (2019, updated 2024) and the UNESCO Recommendation on the Ethics of AI (2021).
“The AI reads. A person decides. The database remembers who.”
In VerityOS the AI is advisory, never authoritative. It reads a bill and proposes extracted figures with a confidence score on every field. No AI output is ever written to the emission vault without a named human explicitly approving it.
Server-side schema validation rejects malformed AI output before it can even reach a human. This is enforced in code and in the database — an append-only vault, and an approval step that records who approved each figure — not left to discipline.
Human retains full control; the AI only recommends.
Human monitors and can intervene.
No human oversight.
The Framework allocates responsibility across three layers. Being precise about which layer we occupy is the whole basis of speaking credibly: VerityOS does not claim to have “solved” model safety — that is Anthropic’s layer. VerityOS implements, by construction, the deployer controls the Framework asks for.
Trains and aligns the Claude models; owns model-level safety evaluation and alignment R&D.
Deploys a model it did not train, inside a narrow, high-stakes domain. Owns the deployer-side controls: how the model is used, output validation, human accountability, and transparency.
Managed database, authentication, storage, and serverless compute on AWS-backed infrastructure in the Singapore region. Owns infrastructure security controls; providers are named in our sub-processor register.
We map our controls to Singapore’s IMDA framework because it’s the standard our market is assessed against — but IMDA did not invent trustworthy AI in isolation. It draws on the same international standards the rest of the world uses. VerityOS aligns to the shared stack: IMDA + OECD + UNESCO. Same principles, three vantage points.
Sources: OECD Recommendation of the Council on Artificial Intelligence (OECD/LEGAL/0449, adopted May 2019, updated May 2024) — the first intergovernmental AI standard and the basis of the G20 AI Principles; UNESCO Recommendation on the Ethics of Artificial Intelligence (adopted November 2021 by 193 member states). The UN Global Digital Compact (2024) is the emerging global architecture these standards feed into.
The first intergovernmental standard on AI, and the basis of the G20 AI Principles. IMDA's framework draws on the same trustworthy-AI concepts.
The first global standard on AI ethics, adopted by all 193 UNESCO member states.
We map to IMDA’s nine dimensions because it’s the framework our market is assessed against — and because it operationalises the OECD and UNESCO principles above into concrete, testable controls. For each dimension: what VerityOS actually does, and where it lives in the product. The controls are enforced in code and in the database, not asserted in a policy PDF.
| Framework dimension | What VerityOS does |
|---|---|
| 1Accountability | Every emission figure records the human who entered and approved it. Every consequential action is written to an append-only audit log (actor, action, detail). Freemansland Creatives (a certified management consultancy) is the accountable deployer; Anthropic is the accountable model developer.Lives in: the vault schema, the audit log, the in-app AI Governance view. |
| 2Data | The model is only ever shown the client’s own primary evidence — bills, receipts, invoices the client uploads. No web-scraped or contentious third-party corpus, no cross-tenant data (row-level security enforces isolation). Client data is Singapore-hosted and not used to train the model.Lives in: the ingestion pipeline, RLS policies, the privacy notice, sub-processor disclosure. |
| 3Trusted Development & Deployment | An AI “food label” on every extraction: the model and prompt version used, intended use (structured extraction of activity data — not autonomous decisions, not financial advice), known limitations (can misread a smudged scan or ambiguous unit), and mitigations (confidence scores, surfaced warnings, server-side validation, human approval).Lives in: the extraction record, the review queue, the in-app AI Governance view. |
| 4Incident Reporting | Structured error logging captures every failed extraction. Corrections are first-class: a wrong figure is fixed with a visible reversal entry, never a silent edit, so the incident and its remediation are both on the permanent record. A responsible-disclosure path is published and monitored.Lives in: structured error logs, the reversal mechanism, the responsible-disclosure policy. |
| 5Testing & Assurance | The product’s entire reason for existing is third-party assurability. The assurance pack exports every figure with its source document, extraction record, pinned factor, hash chain, and approver — so an external assurer can independently re-verify each number.Lives in: the assurance pack export. |
| 6Security | Security by design: all AI output is schema-validated server-side before it can touch the database — which is also the primary defence against prompt injection (a malicious instruction hidden in a document cannot write an unvalidated figure to the vault). Plus row-level security, no service-role key in the app, distributed rate limiting on the AI endpoint, magic-byte file sniffing, per-file hashing. Follows OWASP Top 10 + OWASP LLM Top 10.Lives in: server-side output validation, database-level isolation, rate limiting, the security policy. |
| 7Content Provenance | Our strongest dimension. The append-only, hash-chained evidence vault is cryptographic provenance applied to every AI-touched number. Each entry stores the hash of the entry before it; tamper anywhere and the chain breaks, visibly. At report finalization the chain-head fingerprint is handed to the client and independently timestamped (RFC 3161) — so provenance holds against everyone, including the platform operator. Every figure traces to its exact source document (SHA-256 content hash), the extraction that produced it, and the human who approved it.Lives in: the vault hash chain, the report anchor, the evidence library, the assurance pack. |
| 8Safety & Alignment R&D | Primarily the model developer’s layer (Anthropic). Our deployer-side contribution is domain constraint — narrow structured extraction with validation and human approval sharply limits the blast radius of any model error. We do not overclaim research we don’t do. |
| 9AI for Public Good | Makes assurance-grade ISSB / IFRS S2 reporting affordable for SMEs and their consultants who otherwise cannot access it — directly aligned with IMDA’s SME digitalisation and GenAI-for-SME goals. The product exists to help Singapore SMEs meet the FY2029 assurance mandate credibly. |
ISO/IEC 42001:2023 is the world’s first certifiable AI Management System (AIMS) standard — adopted in Singapore as SS ISO/IEC 42001:2024, with an SAC accreditation programme live since February 2025. It is the AI analogue of ISO 27001. This is an early-mover window: VerityOS is architected to its control structure now, ahead of the market.
ISO/IEC 42001:2023 Information technology — Artificial intelligence — Management system; adopted locally as SS ISO/IEC 42001:2024. Singapore Accreditation Council (SAC) accreditation programme for AIMS certification bodies live since 17 February 2025. Data-governance backbone follows ISO/IEC 27001 / 27701 practices and PDPA compliance — Singapore’s Data Protection Trustmark (SS 714:2025). Note: SS 714 is a personal-data protection standard, distinct from the AI-management standard — the two are not the same thing.
These are the AI-specific control objectives an auditor cares most about — and VerityOS meets their substance today, in code and in the database. Each row: the standard’s objective, and the real VerityOS control that satisfies it.
| Annex A objective | Our real control | Status |
|---|---|---|
| A.7Data for AI systems | Our strongest control. The append-only, hash-chained evidence vault, SHA-256 source-document provenance, pinned versioned emission factors, client-only (no cross-tenant) data, Singapore hosting, and no-training posture ARE AI data governance — quality, provenance and preparation, by construction. | Strong |
| A.6AI system lifecycle | Server-side schema validation, per-field confidence scoring, and an append-only audit log (audit_events / app_errors) cover design → validation → deployment → monitoring and event logging across the AI lifecycle. | Met |
| A.8Information to interested parties | The in-app AI transparency label (model, prompt version, intended use, known limits, mitigations) plus this /ai-governance page disclose capabilities, limits and intended use to the people who rely on the output. | Met |
| A.9Use of AI systems | Responsible use is enforced by design: narrow structured extraction only, no autonomous action, and a named-human approval gate before any figure reaches the vault — the misuse surface is deliberately small. | Met |
| A.10Third-party & customer | Responsibility is allocated explicitly: Anthropic is the model developer; VerityOS is the deployer; certified cloud providers supply the infrastructure — all named in our sub-processor disclosure. The shared-responsibility model above is this control. | Met |
Everything above describes how we govern our own AI. The product goes further: every client workspace carries an AI Data Governance module that operationalises ISO/IEC 42001 for your AI systems — run on the same evidence engine as your emissions numbers. One governed pipeline, two reports: an ISSB-aligned sustainability disclosure, and an AI-governance conformity record.
Overall conformance percentage with progress by control objective — plus the governance evidence captured on every AI extraction: model, prompt version, confidence, warnings, and the named human approver.
Every AI system in the organisation on the record — purpose, owner, risk context. The inventory an ISO/IEC 42001 audit starts from.
The full 65-control catalogue — 27 management-system clauses plus 38 Annex A controls (A.2–A.10) — tracked per control with status, owner, and linked evidence.
Structured impact assessments for consequential AI systems, stored as first-class records beside the emission evidence — not a spreadsheet on someone's laptop.
Where it lives: in the app, under each client workspace → Governance (Overview · Systems · Controls · Impact) — grounded entirely in real vault data. The workspace produces the conformance record; certification itself remains a paid audit by an SAC-accredited body. We say audit-ready, never “certified”.
Speaking credibly means naming the edges. These are the limits we state up front rather than imply away.
The clearest way to see how VerityOS governs its AI is to watch one bill move through it — extraction with confidence scores, the human approval gate, and the hash-chained vault entry that records who approved what.