Practical notes on Home-Cage Monitoring, FAIR metadata, behavioral data analysis, and preclinical research infrastructure — written from real project experience.
Article 50 was not written as a research-integrity rule. But three decisions underneath it — disclose clearly at first exposure, scale the obligation to what the machine actually did, and mark it machine-readably — are a better starting point for scientific AI disclosure than the yes-or-no checkbox. The one correction the transplant needs: Article 50 asks whether an audience can recognise synthetic content, while science has to ask where the model touched the research process. A proportionality ladder from language editing to agentic provenance, and why only checkable disclosure has ever moved behaviour.
Animal harms are justified only by the knowledge a study yields — so biased, unreported, or non-reusable data leave that harm partly unjustified after the fact. Is that a quirk of the mouse lab or a general property of research on animals? A cross-field synthesis of some fifty papers, from preclinical oncology to precision livestock and wildlife biologging, says the debt travels: reporting compliance of 23%, data withheld from 93% of on-request promises, welfare that degrades the data it produces, and no mature framework for animal-data governance anywhere. The repayment mechanism is the same in every setting — FAIR metadata and reuse, which is how Reduction is actually delivered across studies.
I am finishing a short book — The Ethical Debt: Why Wasting Animal Data Is Wasting Animal Lives — and this note previews its argument and its ten chapters. The one idea it is built on: data welfare is animal welfare. An animal experiment is a bargain across time, and when the resulting data become unusable the bargain breaks after the fact, falling forward onto the next animals. The preview walks from how data die, through the 3Rs in the data age, to the repayment mechanisms — FAIR metadata, virtual control groups, credible NAMs, and a WellFAIR charter — grounded throughout in the reproducibility and data-sharing evidence.
ARRIVE and PREPARE are often mentioned as if interchangeable — they are not. ARRIVE is a reporting checklist that governs the write-up after a study ends; PREPARE is a planning checklist for everything that should happen before it begins, including facility evaluation, staff training and husbandry that reporting guidelines never touch. This note sets the two side by side, uses Norecopa's aviation-checklist argument for why reporting alone cannot fix a study, and places both on the open-science "pathway to better research" — from PREPARE through to PUBLISH. It closes on what planning means for FAIR data: the metadata that makes a dataset reusable is decided up front, not scraped together at submission.
A paper list tells you what was measured with a Digital Ventilated Cage — not how. So we read the methods sections and turned them into a connected pipeline: one raw 12-electrode signal forking into activity, rhythm, place, and the cage environment, each with its recipe, thresholds, and the cage configurations it is valid for (ALI for any cage; trajectory metrics single-housed only). It lands on the real constraint — not the maths, but getting the raw data and its metadata into a usable, FAIR form. Built by a Neuronautix agent team from the Tecniplast catalogue, as a working demo of the knowledge base as a second brain.
Beyond the 3Rs lies a fourth, too-often-optional obligation: to care for the data. A mouse can be housed and monitored ethically and still be wasted if its data are incomplete, unannotated, or lost after one paper. When data cannot be reused, animals may be used again — that is the ethical debt, and it is paid in lives. Grounded in the WellFAIR paper, ARRIVE/PREPARE, MNMS, the reproducibility literature, and the road to virtual control groups.
A referenced explainer of the metrics derived from the Tecniplast Digital Ventilated Cage — Activation Density, the Rest Disturbance Index, circadian and spatial-preference metrics, the Bedding Status Index, the Urination Index, and GYM500 wheel running. For each: what it is, how it comes from the capacitance signal, and what it indexes biologically.
A 2026 multicentric study shows machine learning on continuous home-cage locomotion flags distress 3–6 days before human checks. What AI adds to HCM research — earlier endpoints, digital biomarkers, multi-animal phenotyping, severity scoring — and where it augments rather than replaces human expertise.
A repository-level manifest of epistemic provenance — who and what (human or AI) produced each claim, and how confident it is — the macro companion to per-claim trust markup.
Recap of the COST TEATIME WellFAIR webinar with Benoit Petit-Demoulière: FAIR data as the most under-used lever for the 3Rs, demos of FAIR3R.fr and Metadatapp, and the path from born-FAIR capture to AI agents and virtual control groups.
A lightweight root manifest that lets a repository state its own FAIR posture — honestly, with gaps marked — and point to codemeta, CITATION.cff and RO-Crate.
A proposal for two lightweight root files that give any repository a readable, machine-checkable FAIR posture and an explicit epistemic provenance declaration — addressing both the findability gap and the credibility gap in AI-assisted scientific publishing.
FAIR metadata management is not just a technical requirement — it is part of the animal welfare obligation. Lost or non-reusable animal-derived data represents an ethical failure in preclinical research.
A 2050 vision of federated, interoperable, FAIR-by-default Home-Cage Monitoring — unified data streams, shared ontologies, open classifier infrastructure, and the 3Rs made structural — paired with an honest six-point sense-check of the limits that will still be standing.
FAIR metadata, treated as a New Approach Methodology in its own right, is the substrate that turns isolated preclinical data into reusable evidence and credible virtual control groups. Companion to the SCAND-LAS 2026 talk.
A knowledge graph is not an endpoint. The operational sequel to the May 18 KG-architecture note — distinguishing graph DB / KG / ontology / GraphRAG, separating evidence-ontology-application layers, calibrating GraphRAG, treating link prediction as ranked hypothesis, and running four experiments before any custom model.
Data without context is not evidence. A narrative tracing the library analogy through FAIR principles, WellFAIR ethics, and the metadata requirements that make virtual control groups scientifically valid.
AI-ready preclinical data needs human review at the points where ambiguity matters: context of use, ontology mapping, protocol deviations, uncertainty, and claim interpretation.
Long context is not understanding. A reference-backed argument for replacing universal-LLM-reader workflows with metadata, ontologies, graphs, validation, and targeted human review.
A metadata agent for NAM evidence should start from a schema, not a summary prompt. Its job is to extract candidate fields, validate them, and route uncertainty to experts.
Post-hoc curation can repair labels, but it cannot recreate missing protocol context, provenance, endpoint definitions, controls, or assay deviations. NAM platforms need metadata capture at source.
Fifteen years of behavioral neuroscience at EPFL and CNRS/INSERM Montpellier, an EPNA-awarded Science Advances paper, and a working consultancy that turns preclinical data problems into open infrastructure — a short introduction to the person behind Neuronautix and how to hire him.
FAIRSCAPE and BioCompute are useful patterns for NAM provenance because they record datasets, software, parameters, runtime context, personnel, and workflow inputs and outputs in machine-readable form.
How neuronautix.com uses eight specialized agents, a self-updating knowledge base, and a structured publishing pipeline to make each note smarter than the last — with no database and no build step.
LinkML is useful for preclinical metadata because it turns a single schema into validation rules, JSON-LD contexts, documentation, APIs, Python classes, and semantic mappings.
NAM platform qualification should not start after the assay works. Metadata is part of the evidence: context of use, provenance, controls, endpoint definitions, and validation records determine whether a result can be reviewed or reused.
A practical build plan for NAM Evidence Commons: schema-first metadata capture, validation reports, provenance packages, evidence graphs, and agent-assisted curation.
FDA's Elsa and HALO trajectory changes the practical question for NAMs: not whether the assay is novel, but whether the data is structured enough to review, compare, and reuse.
FDA now accepts NAMs in INDs. The bottleneck is not generating the data — it is making it structured, traceable, and interpretable enough for a reviewer to act on. Five concrete ways data management closes the gap between experiment and regulatory confidence.
Organoids, organ-on-chip, and in silico models each have their own data silos. NAMO is a LinkML-based ontology from the Monarch Initiative that provides a single structured framework for all three — with a five-dimensional validation concordance model and deep integration with UBERON, Cell Ontology, and ChEBI.
First notes in preparation — check back soon, or get in touch to suggest a topic.
Have a topic in mind?
Suggest a method note or case study
If you are working on a Home-Cage Monitoring implementation, FAIR metadata strategy, or behavioral data analysis challenge and want it covered, contact Neuronautix directly.
