{"agency":"Signal (demo · South Australia Police context)","accountable_official":"Accountable Official (demo)","policy":"Policy for the responsible use of AI in government (DTA, v2.0)","mandatory_from":"15 December 2026","use_cases":[{"use_case":"Crime trend analysis","risk_category":"limited","decisions":5,"data_sources":["SA Police Crime statistics (data.sa.gov.au) — FY2024-25 + FY2025-26 YTD — bundled snapshot fetched 2026-06-13 (live refresh in progress)","SA Police Crime statistics (data.sa.gov.au) — FY2024-25 + FY2025-26 YTD, live as of 2026-03"],"affected_groups":["Members of the public in the analysed regions","Communities subject to differential policing or reporting"],"benefits":"Surfaces trends and anomalies for 'Crime trend analysis' over aggregate, de-identified data, with every answer audit-logged for accountability.","risks":["Aggregate counts may be misread as rates, overstating differences between regions or groups.","An LLM-written narrative could misstate a figure or the direction of a trend.","Statistical anomalies could be over- or under-interpreted without context."],"mitigations":["Aggregates only — no personal information enters the system, so re-identification risk is minimal.","Every LLM narrative is checked by a deterministic faithfulness eval; unfaithful output is rejected and the deterministic template served (mean faithfulness 1.0, 0 fallbacks).","Statistically anomalous results are flagged for human review (100% of decisions in this use case); reviews and overrides are recorded against the decision.","Full audit trail is public (GET /decisions); every output carries a decision_id."],"human_oversight":"5 of 5 decisions in this use case were flagged for human review. Anomalous results require human review before action.","fairness_considerations":"Counts are not rates: differences across regions may reflect population, reporting or policing intensity rather than real offending. Outputs must not be used to rank places or people.","residual_risk":"Limited. With aggregates-only data, figure-faithful narratives and human review on anomalies, residual risk is low; the main residual risk is misinterpretation of counts as rates by downstream readers.","first_used":"2026-06-21T16:30:35.099832Z","last_used":"2026-06-21T16:31:04.194622Z"},{"use_case":"Multi-step crime trend analysis","risk_category":"limited","decisions":1,"data_sources":["SA Police Crime statistics (data.sa.gov.au) — FY2024-25 + FY2025-26 YTD, live as of 2026-03"],"affected_groups":["Members of the public in the analysed regions","Communities subject to differential policing or reporting"],"benefits":"Surfaces trends and anomalies for 'Multi-step crime trend analysis' over aggregate, de-identified data, with every answer audit-logged for accountability.","risks":["Aggregate counts may be misread as rates, overstating differences between regions or groups.","An LLM-written narrative could misstate a figure or the direction of a trend.","Statistical anomalies could be over- or under-interpreted without context."],"mitigations":["Aggregates only — no personal information enters the system, so re-identification risk is minimal.","Every LLM narrative is checked by a deterministic faithfulness eval; unfaithful output is rejected and the deterministic template served (mean faithfulness 1.0, 0 fallbacks).","Statistically anomalous results are flagged for human review (100% of decisions in this use case); reviews and overrides are recorded against the decision.","Full audit trail is public (GET /decisions); every output carries a decision_id."],"human_oversight":"1 of 1 decisions in this use case were flagged for human review. Anomalous results require human review before action.","fairness_considerations":"Counts are not rates: differences across regions may reflect population, reporting or policing intensity rather than real offending. Outputs must not be used to rank places or people.","residual_risk":"Limited. With aggregates-only data, figure-faithful narratives and human review on anomalies, residual risk is low; the main residual risk is misinterpretation of counts as rates by downstream readers.","first_used":"2026-06-21T16:31:04.196605Z","last_used":"2026-06-21T16:31:04.196605Z"}],"statement":"# AI use-case impact assessment\n\n**Agency:** Signal (demo · South Australia Police context)  ·  **Accountable official:** Accountable Official (demo)\n\n**Policy basis:** Policy for the responsible use of AI in government (DTA, v2.0); AI use-case impact assessments mandatory from 15 December 2026.\n\n## Crime trend analysis (limited risk)\n- **Decisions:** 5\n- **Affected groups:** Members of the public in the analysed regions, Communities subject to differential policing or reporting\n- **Benefits:** Surfaces trends and anomalies for 'Crime trend analysis' over aggregate, de-identified data, with every answer audit-logged for accountability.\n- **Risks:** Aggregate counts may be misread as rates, overstating differences between regions or groups.; An LLM-written narrative could misstate a figure or the direction of a trend.; Statistical anomalies could be over- or under-interpreted without context.\n- **Mitigations:** Aggregates only — no personal information enters the system, so re-identification risk is minimal.; Every LLM narrative is checked by a deterministic faithfulness eval; unfaithful output is rejected and the deterministic template served (mean faithfulness 1.0, 0 fallbacks).; Statistically anomalous results are flagged for human review (100% of decisions in this use case); reviews and overrides are recorded against the decision.; Full audit trail is public (GET /decisions); every output carries a decision_id.\n- **Human oversight:** 5 of 5 decisions in this use case were flagged for human review. Anomalous results require human review before action.\n- **Fairness:** Counts are not rates: differences across regions may reflect population, reporting or policing intensity rather than real offending. Outputs must not be used to rank places or people.\n- **Residual risk:** Limited. With aggregates-only data, figure-faithful narratives and human review on anomalies, residual risk is low; the main residual risk is misinterpretation of counts as rates by downstream readers.\n\n## Multi-step crime trend analysis (limited risk)\n- **Decisions:** 1\n- **Affected groups:** Members of the public in the analysed regions, Communities subject to differential policing or reporting\n- **Benefits:** Surfaces trends and anomalies for 'Multi-step crime trend analysis' over aggregate, de-identified data, with every answer audit-logged for accountability.\n- **Risks:** Aggregate counts may be misread as rates, overstating differences between regions or groups.; An LLM-written narrative could misstate a figure or the direction of a trend.; Statistical anomalies could be over- or under-interpreted without context.\n- **Mitigations:** Aggregates only — no personal information enters the system, so re-identification risk is minimal.; Every LLM narrative is checked by a deterministic faithfulness eval; unfaithful output is rejected and the deterministic template served (mean faithfulness 1.0, 0 fallbacks).; Statistically anomalous results are flagged for human review (100% of decisions in this use case); reviews and overrides are recorded against the decision.; Full audit trail is public (GET /decisions); every output carries a decision_id.\n- **Human oversight:** 1 of 1 decisions in this use case were flagged for human review. Anomalous results require human review before action.\n- **Fairness:** Counts are not rates: differences across regions may reflect population, reporting or policing intensity rather than real offending. Outputs must not be used to rank places or people.\n- **Residual risk:** Limited. With aggregates-only data, figure-faithful narratives and human review on anomalies, residual risk is low; the main residual risk is misinterpretation of counts as rates by downstream readers.\n\n"}