Evaluation, verification and oversight for autonomous AI systems
AI agents are starting to do real work — and mostly, we take their word for it. I design experiments that reveal how these systems actually behave, and build open-source tools that check their answers against evidence.
agents & multi-agent systems·evaluation·evidence & verification·privacy·healthcare AI
ORCID
Google Scholar
GitHubOpen to full-time applied AI roles · evaluation, forward-deployed & applied AI engineering · London, Oxford or UK-remote · Get in touch →
01Flagship research · Observer Zero
What happens when AI agents become scientists — inside a world whose ground truth we control?
Observer Zero is a persistent artificial society for studying how autonomous LLM agents gather evidence, interpret it, revise their beliefs and reach collective conclusions. Its agents inhabit Meridian — a world with fictional physics — where they run experiments, exchange letters and hold explicit, probabilistic beliefs about laws they cannot see. Because the simulator holds perfect ground truth, every belief can be scored against reality: the design separates what a society had the evidence to know from what it actually came to believe.
When the laws of their world secretly changed, the agents gathered evidence enough for a blind statistical detector to find the change — and almost none of them ever believed it.
Eight agents measure a world they can never see into directly. Observer Zero sits outside it, holding the ground truth — so what each agent could know can be separated from what it came to believe.
When the Laws Changed
Can autonomous agents recognise that a physical law of their world has changed?
Across 150 universes, 90–100% of agents transiently noticed something was wrong — yet 0 of 40 gravity-shift worlds reached a strict law-change conclusion. The anomalies were real; the interpretation reached for faulty instruments instead.
Does Society Help?
Can an artificial scientific society overcome the individual failure through social interaction?
Pre-registered, 85 universes, societies of two and eight. Left alone, grounded societies never spoke; seeding one communicative agent spread unsupported claims, not scrutiny. The combined two-study paper is in peer review, with the model published in the CoMSES library.
The Eureka Threshold
What would it take for an AI scientist to discover that its world is simulated?
Pre-registered and frozen before the first run: 170 universes across four model families, where one instrument's readings genuinely came from outside the world. No agent ever formed that hypothesis — 0 of 38 treatment runs — while agents licensed to consider it endorsed it before the manipulation had produced any evidence. Belief tracked prompting, not evidence. The paper is in peer review.
02Selected systems
From research questions to working systems.
Strong claims about AI behaviour need trustworthy measurement. One question runs through all three of these systems — how do we know an AI is doing what we think it is? — and each answers it in a different domain. Every one ships with its own evaluation.
An open-source verification framework for evidence-grounded AI. Deterministic grounding and extraction checks — no LLM judge anywhere in the loop — so verdicts are reproducible, auditable, and cheap enough to run on every answer or every commit. Four production systems run on it in CI, where it has caught silent parse failures, a citation-parser inconsistency, and a simplifier dropping an antibiotic dose from a discharge summary.
Redacta
How can agents work with sensitive information without exposing it unnecessarily?
A privacy boundary for AI and agent workflows. Redacta pseudonymises patient identifiers — names, NHS numbers, dates — into labelled tokens, and its MCP server keeps the map that reverses them outside the model’s context entirely. It then checks the model’s output for identifiers that found their way back in, so the boundary is verified rather than assumed. Benchmarked under adversarial attack by its own published evaluation harness, and deployable from a free iPhone app to a self-hosted Kubernetes cluster.
Patiently AI
Can generative AI simplify medical information without losing meaning?
A deployed, MHRA-registered Class I medical device that turns clinical letters, results and discharge summaries into plain language — in 18 languages, on web, iOS and Android. Validated in a three-phase, peer-reviewed study (Frontiers in Digital Health, 2026) — computational readability analysis, expert review by 15 healthcare professionals who rated outputs 4.49/5 for accuracy, and a 54-patient survey. Five national awards, in real-world use.
03The wider lab
Experiments, tools and detours.
The same method at smaller scale: build it, measure it, publish the measurement. A progression runs through the lab — from applications, to infrastructure, to evaluation, to studying the behaviour of the systems themselves.
RefCheckr
Automated claim and reference verification for medical and scientific content — with its model choices benchmarked in the open.
→InfrastructureGroundwork
Deployment-readiness harnesses for document QA and structured extraction — a CI gate that catches fabrication before users see it.
→Retrieval · MCP DirectoryPubCrawl
Real literature for AI assistants: PubMed, Europe PMC, clinical trials and drug labels over MCP — no hallucinated citations.
→Methodology · MCP serverAI Water & Energy Calculator
Why published estimates of AI water use differ 100× — and what yours comes to. The same methodology runs as an open-source MCP server with no model in any tool.
→Fine-tuning · validationTime-Machine Chess
Era-specific engines fine-tuned on 150 years of historical games — and validated against the history they learned from.
→LabsAll experiments
HushMap, PosterLens, StudyDiff, BiomarkerFinder, TrialGen, PlaceboGPT and other prototypes and detours live in the Labs.
04About
Who’s behind this.
I build AI systems — then ask harder questions about whether we should trust them.
Nick Lamb is an independent AI researcher and engineer with an unusual route in: twenty years of regulated medical communications (PhD, CMPP), a field where every claim must trace to evidence. That standard now runs through everything on this page — agents whose beliefs are scored against hidden ground truth, verification that never takes a model’s word for it, and a registered medical device validated before it reached patients.
PharmaTools.AI began as a collection of healthcare AI tools, and healthcare remains its proving ground: few domains punish an ungrounded claim faster. The work now spans empirical AI research, autonomous agents, evaluation and privacy infrastructure — much of it open source.
Interested in the same questions?
How autonomous AI systems reason, use evidence — and fail. If you’re working on related problems, or want this kind of discipline applied to your own AI systems, I’d be glad to hear from you.
