We build AI systems for the real world. Not demos. Not proofs of concept. Production infrastructure that runs when nobody is watching.
The empirical study of why AI products fail in deployment when they pass evaluation. Active lines of inquiry include confidence calibration at the interface layer; latency as a perceived-quality signal; the relationship between recovery affordance and user trust; and methodology for adversarial testing of composite model-and-interface systems.
Methodology for verification of AI-generated artefacts through structured review by multiple independent systems. Our framework, the AI Quality Trident, is being formalised for publication.
The technical study of where pretrained models diverge from useful behaviour in specific deployment conditions. We work on open-weights and frontier models, with attention to the gap between leaderboard performance and behaviour against real users.
Long-horizon autonomous AI and the cross-platform infrastructure required to operationalise it. Our open framework, IJFW, underpins this work and is in active development across eight AI coding agents.
Long-horizon research into the architecture of intelligence and the emergent properties that arise from it. Several lines of work in this area are powerful enough that we have not yet decided how, or whether, they should be released. Findings will appear in peer-reviewed venues when they are ready.
Ferrox Labs operates under a research-and-engineering discipline developed across twenty-six years of independent practice. Every artefact released by the laboratory, from a modified model checkpoint to a deployed tool, passes through the same verification process: pre-implementation acceptance criteria, multi-model adversarial review at every phase transition, and adversarial human testing under conditions designed to make the artefact fail.
We do not publish work we cannot defend in review.
We do not release tools we would not use ourselves.
We do not put our name on capability we have not stress-tested under adversarial conditions.
The discipline is documented in full in a forthcoming book on AI craft.
Original research, recurring benchmark series, and notes from the laboratory. We publish what we find. One benchmark report and two research reports are available now.
Better code than the frontier models wrote, at 1/29th of what they charge. 20 systems, 196 machine checks: the gated climb (the Anvil method, developed at Ferrox Labs) matched the best on correctness, led the profiled lanes on maintainability and measured runtime, and cost $0.0063 a task. The full harness and raw results are published for independent verification.
Read the report →Findings from twelve primary research sources on AI agent adoption, deployment, and ROI. Introduces the Agent Maturity Ladder, the Ferrox Labs methodology (Decompose, Cross-Audit, Validate), five validated workflow blueprints, security and legal readiness, and a 30-day deployment protocol.
Download the report →Independent benchmark across coding, agentic, reasoning, and practical categories with explicit methodology and contamination flags. Sources include Scale AI SEAL, Artificial Analysis, OpenAI, Anthropic, Google DeepMind, DeepSeek, Moonshot AI, BenchLM, HuggingFace, and official model cards.
Open the May 2026 edition →A forthcoming book documenting the engineering discipline that decides whether AI delivers in production. Built on twenty-six years of independent practice. Working title held back until release.
Framework references, working papers, and field notes from the laboratory. Released when defensible under independent review. No schedule stated, none implied.
Ferrox Labs is led by Sean Donahoe, originator of the empathy gap construct for AI quality, the Donahoe Loop development methodology, the AI Quality Trident multi-model verification system, and the Agent Maturity Ladder for diagnosing organisational readiness for AI deployment.