Frans Moore

Built on one premise: an agent’s report of its own work is a claim, not evidence

• Designed and operate the system through which a fleet of AI agents runs research, engineering and operations for my company every day, coordinating through shared memory and structured handoffs rather than through each other.
• Agents’ claims of “done” are checked, not trusted. In May 2026 a rule-based check on my automated dispatcher refused 98 of 143 AI-agent claims of success (69%), most after my own instructions dropped the required documents, a gap no agent flagged.
• Safeguards are measured before they are enforced. Each new gate runs in record-only mode first, so its false-block cost is known before it can stop work. One promising gate stayed off because the measurement showed it would have refused most completions.
• Safety constraints live in the tool layer, not the prompt, so no chain of agent reasoning can argue past them. The production runtime is fail-closed: if anything in it has been altered, it will not start.
• Audits itself: my own Critical audit found the system’s “completed” status was hiding failures, and I fixed it at the root. An internal red-team agent found a credential exposure; it was fixed and closed.

Public data and methods: github.com/ArcedeDev/bdbv-2026-lovs

• Built the first release solo in about a week with AI coding agents, then kept it current through 127 official situation reports from multiple publishers, reconciling conflicting sources into one cited picture.
• Every published figure traces to its source and the date the data describes. An evidence service for AI clients refuses to serve any number without a citation, so a model answering questions about the outbreak cannot state an unsourced figure.
• Pre-registered 87 public, dated, thresholded forecasts before outcomes were known (one block archived on Zenodo); 54 have resolved and 31 resolve on 1 Oct 2026. Scores are published against a base-rate baseline, misses beside hits, and I retired a model input when its pre-registered ranking came out inverted.
• Engaged UNICEF’s Eastern and Southern Africa regional leadership from May to August 2026; the Deputy Regional Director assigned regional staff to review the work in detail. Featured as an Apolitical Government AI Navigator case study and covered by Scrolla.Africa.

• Earth, LOVS and the agent-trust tools were built with AI coding agents working inside a verification-gated pipeline I designed: requirements are interrogated, assumptions checked against ground truth, and every change independently reviewed and red-teamed before it ships. The review stage has blocked merges and caught Critical bugs, including a data sync that silently dropped every row.

• Author of Composable Trust Infrastructure for the Agent Internet (Zenodo, CC BY 4.0), on verifiable agent identity, capability discovery and authorization. Core claim: every consequential action must prove who authorized it, what constraints applied, and what happened.
• Open Agent Trust Registry: 13 forks, 9 outside contributors, 11 registered issuers. AIR SDK (@arcede/air-sdk), an agent web-perception SDK with its own MCP server: 30 releases, 5,200+ npm downloads.

• 83 pre-registered tests and 87 logged corrections in a locked register, with power analysis before any null: underpowered results are reported with their minimum detectable effect, never as absence. Central finding: the failures traced to the observation model, not the biology.

AI, evidence and coordination systems for institutions making consequential decisions. A one-person company run with the agent fleet above; engagements with global health and biodiversity philanthropy, including an AI policy options map for a U.S. biodiversity foundation.

• Coordinated a three-year mixed-methods human health and biodiversity survey across South Kivu and Tanganyika provinces in a conflict-affected setting: household surveys, focus groups, entomological and fish sampling. The 2022 South Kivu wave alone covered 160 households on a 116-item instrument, 111 focus-group participants and 561 fish.
• Raised about $500,000 in grants and operating funds with the founder; managed organization and research finances; digitized provincial health-system administrative registers; published the program’s fisheries data to GBIF as originator.

Program evaluation and board governance; contributed to the 2022–2026 strategic plan and the MOU with South Sudan’s Ministry of Health.

Freelance software engineering, mostly video game development, with web development alongside. Received an Epic MegaGrants award.

Designed and ran an independent mixed-methods study of staff through the centre’s transition to Good Laboratory Practice certification (30 of 36 staff; Swahili instrument, translated and back-translated). Assisted malaria bed-net quality surveys. Two Best Poster awards. Also research assistant on NASA-funded Project FUSION (multiteam systems, LINC Lab) and a workaholism study (Healthy Work Lab), UGA.

Dissertation: a rapid evidence assessment of how immunization programs in low- and middle-income countries adapt during infectious disease crises (5,806 records screened, 19 studies included). Coursework in statistical methods for epidemiological and clinical research. Dissertation (PDF)

Honors thesis on job satisfaction at an East African medical research centre. Coursework included artificial intelligence, taught as CSCI 2670, Introduction to Theory of Computing; cognitive science ran through UGA’s Institute for Artificial Intelligence. Honors thesis (PDF)