EvoXplain detects when a model's explanations are pipeline artefacts rather than mechanistic reality — providing falsifiable evidence for model risk teams in pharma, biotech, financial services, and AI regulation.
Modern ML pipelines exhibit mechanistic non-identifiability: equally accurate models reached by retraining the same pipeline with different seeds, splits, or hyperparameters can rely on disjoint feature sets to make their decisions. Averaged explanations across reruns produce a ghost — a vector that resembles no mechanism the model actually uses, yet is what regulators, clinicians, and auditors are typically shown.
EvoXplain treats the attribution-vector manifold as a topology with reproducible basins that are pipeline-dependent and biologically meaningful. The battery is engineered to kill the multiplicity hypothesis under every plausible null — randomness, sampling noise, model instability, hyperparameter drift, attribution-method choice. Multiplicity that survives is real.
Across a 100×100 design (100 splits × 100 seeds, seeds 800–899) on TCGA pan-cancer and TCGA-BRCA Luminal-vs-Basal, the falsification battery confirmed that explanation multiplicity persists under every null — and that the resulting basins encode distinct biological pathways, not noise.
EvoXplain is designed for organisations where ML decisions are subject to scientific scrutiny, regulatory audit, or legal challenge — and where averaged or single-seed explanations are insufficient evidence.
Detect when biomarker attributions, drug-target rankings, or patient-stratification features are pipeline artefacts before they enter a study design. Distinguish reproducible biological signal from explanation drift across retraining runs.
Provide model risk teams with falsifiable evidence that explanation features are mechanistically grounded — not artefacts of the training pipeline. Integrates into model validation, challenger reviews, and adverse-action documentation.
Equip auditors and conformity assessment bodies with a reproducible test for whether claimed explanations of high-risk AI systems are robust under retraining — addressing the gap between transparency obligations and post-hoc XAI fragility.
EvoXplain is released under AGPL-3.0 and protected by a UK provisional patent. The codebase, falsification battery, and reproducibility scripts are open. Commercial deployment, dual-licensing, and bespoke integration are handled through direct enquiry.
The framework, preprint, and falsification harness are publicly available. Issues, replications, and academic collaborations are welcome.
arxiv.org/abs/2512.22240 →Integration into regulated ML pipelines, dual-licensing for commercial use, model-risk consulting, and conformity-assessment partnerships.
contact@evoxplain.com →