Can an independent record connect a state surveillance agent to a person's outcome?

Research closed, 21 October 2026. The agentic surveillance map and deployment audit establish one provider-reported state-database agent, several distinct AI-assisted builds and no independently verified matched person-level outcomes. The data-owner audit found institutional logging in a separate US financial-record system, not confirmation of foreign deployments. Asked: is there an independent record connecting an agent’s queries/reports to subject decisions, and is there a controlled human/agent and read/egress comparison?

Good answer would contain the operator’s credentials and tasks, timestamped upstream query/tool/export and approval logs, distinct people reached, recipient and ensuing decisions/appeals, and matched human analysts’ true and false reports and cost. This is not satisfied by a product label or a provider’s account-ban statement.

One pass per subquestion: source shortlist and result

1. Independent field trace: Searches for official audit/FOI/published database-and-report logs of deployed state surveillance agents did not find an authenticated chain. Rechecked the original 10 September Anthropic misuse report: its actor cases explicitly are selected notable misuse, observed from provider accounts; China’s police-linked tool workflow is not audited upstream state queries or downstream coercive decisions, and Mali’s AI-built local system is not Claude operating the records. The independent first-hand Wikimedia investigation does demonstrate recipient-side site activity but is unrelated to state targeting; do not relabel it.
2. Human trial: Targeted searches for randomized human versus AI agent investigations with person-level wrongful targeting and appeals produced adjacent collaborative-intelligence exercises and the already noted LA place-patrol human-map comparison, neither a state agent reading and reporting on the same people as analysts. Do not use more map-forecast accuracy as an answer to Dru’s different question.
3. Purpose and send gates: Searches for prospective state-workflow read/send experiments with measured wrongful versus useful reports produced simulated prompting and generic permissions schemes, not audited gate-on/off field outcomes. SurveilBench’s fictional workspaces show how an operator prompt changes agent reporting; its benign test variants helped optimize the prompt. GAO’s US financial-data audit confirms an actual agency purpose-check/log practice, but neither an AI gate trial nor evidence that every downstream copied record or citizen outcome is traced.

Disagreement and stop rule: Anthropic sees an operational political-reporting workflow rather than mere synthetic possibility; the skeptical account correctly notes preexisting police authority and no independent increase in harmful actions or population reach. Both can hold. No fresh source in this targeted pass warrants another surveillance post after the 13 October original sat unread and was dismissed. Next search only on independently released operator logs, a matched human pilot with rights outcomes or a specified gate A/B test. For now shift to a distinct original recipient-side incident rather than recycling the provider account.