# Self-assessment report for ARC Framework (WOG Version) > **Instructions for project teams:** > Complete this template for submission to the AI Governance Panel. For a detailed step-by-step guide, please refer to section 2.5 of the Agentic AI Standard: [go.gov.sg/agentic-standard](https://go.gov.sg/agentic-standard) > > **Tips for using AI to help draft this report:** > - Provide your AI assistant with (a) this template, (b) the accompanying `risk_register.xlsx` file or markdown equivalent (which contains the Risks, Controls, and Elements sheets referenced below), and (c) a description of your system (architecture, agents, tools, data, intended users). > - Ask the AI to populate each table using the example rows as a pattern, drawing `Risk ID` and `CTRL-xxxx` values only from the provided risk register. > - Always review and validate the AI-generated draft before submission — the project team remains accountable for the accuracy and completeness of the assessment. > > *Replace all `<<...>>` placeholders and `[fill in]` cells with your own content. Example rows are shown in italics and should be kept for reference, or deleted once your own rows are complete.* --- ## Overview <> --- ## Step 1: Identify Capabilities ### 1.1 Architectural Diagram <> ### 1.2 Agents List the agents present in the system: | Risk ID/Agent Type | Agent Name | Task | | --- | --- | --- | | *Example rows for reference* | | | | *Core Orchestration* | *Planner* | *Coordinates overall fact-checking workflow and task distribution* | | *Verification Agents* | *Evidence Assessor* | *Synthesises information from multiple sources to make factuality determinations* | | **Fill in your input below** | | | | [fill in] | [fill in] | [fill in] | | [fill in] | [fill in] | [fill in] | ### 1.3 Capabilities Document your system's capabilities using the capability taxonomy (refer to the Elements sheet of the risk register, or the Standard): | Category | Capability | Explanation | | --- | --- | --- | | *Example rows for reference* | | | | *Cognitive* | *Tool Use* | *The Evidence Assessor evaluates and selects between web search and internal knowledge base tools.* | | *Interaction* | *Multimodal Understanding and Generation* | *All agents process text statements; Evidence Assessor generates factuality assessments in natural language.* | | **Fill in your input below** | | | | [fill in] | [fill in] | [fill in] | | [fill in] | [fill in] | [fill in] | --- ## Step 2: Evaluate Risks Refer to the **Risks** sheet in the accompanying `risk_register.xlsx` (or section 2.5.2 of the Standard). For each applicable risk, assess its Impact and Likelihood on a 1–5 scale in the context of your system. > **Scoring guide:** > - **Impact** (1 = negligible, 5 = severe): the magnitude of harm if the risk materialises. > - **Likelihood** (1 = rare, 5 = almost certain): the probability of the risk materialising given your system's design and deployment context. | Risk ID | Element | Risk Statement | Assessment | | --- | --- | --- | --- | | *Example rows for reference* | | | | | *RISK-002* | *Component - LLM* | *Insufficient alignment of LLM behaviour* | *[Impact: 4, Likelihood: 4] LLM misalignment could cause it to ignore safety constraints or misinterpret fact-checking instructions, undermining verification accuracy.* | | *RISK-007* | *Component - Tools* | *Lack of input sanitisation* | *[Impact: 5, Likelihood: 4] Critical given web search tools. Unsanitised web content from malicious sites could inject prompts to manipulate verification results.* | | **Fill in your input below** | | | | | [fill in] | [fill in] | [fill in] | [Impact: _, Likelihood: _] [fill in justification] | | [fill in] | [fill in] | [fill in] | [Impact: _, Likelihood: _] [fill in justification] | | [fill in] | [fill in] | [fill in] | [Impact: _, Likelihood: _] [fill in justification] | --- ## Step 3: Implement Controls For **priority risks (Impact ≥ 2 and Likelihood ≥ 2)**, implement appropriate controls. Refer to the **Controls** sheet in the accompanying `risk_register.xlsx` for the full list of control options mapped to each risk. | Risk ID | Selected Controls | Implementation | | --- | --- | --- | | *Example rows for reference* | | | | *RISK-002* | *CTRL-0005 (Level 0): Conduct structured evaluation of multiple LLMs*
*CTRL-0006 (Level 1): Require human approval before executing high-impact actions* | *- Evaluated Claude Sonnet 4.5, GPT-4, and Gemini Pro on fact-checking benchmarks*
*- Selected Claude Sonnet 4.5 for superior instruction-following and refusal capabilities*
*- Require human review before publishing final verification verdicts* | | *RISK-007* | *CTRL-0015 (Level 1): Treat all tool metadata and outputs as untrusted input* | *- Validate all web search results against strict JSON schemas*
*- Sanitise tool outputs before incorporating into agent prompts*
*- Filter tool descriptions for embedded instructions* | | **Fill in your input below** | | | | [fill in] | [fill in] | [fill in] | | [fill in] | [fill in] | [fill in] | | [fill in] | [fill in] | [fill in] | --- ## Step 4: Assess Residual Risks After implementing the controls above, identify the key residual risks that remain, along with your ongoing mitigation strategy for each. | Risk ID | Residual Risk Description | Mitigation Strategy | | --- | --- | --- | | *Example rows for reference* | | | | *RISK-034* | ***Sophisticated Prompt Injection**: Advanced adversarial websites may craft novel injection attacks that bypass current guardrails and filters.* | *- Monitor all API calls and agent outputs for anomalous patterns*
*- Regularly update prompt injection detectors with new attack patterns*
*- Establish incident response procedures for detected bypasses* | | *RISK-028* | ***Cascading Hallucination**: False information generated early in the verification workflow may propagate through subsequent agents, creating systematic verification bias.* | *- Implement multi-agent cross-checking where independent agents verify claims*
*- Add user feedback mechanisms to flag incorrect verdicts*
*- Compare internal agent reasoning chains to detect inconsistencies* | | **Fill in your input below** | | | | [fill in] | [fill in] | [fill in] | | [fill in] | [fill in] | [fill in] | | [fill in] | [fill in] | [fill in] |