Alpha software — use with appropriate caution
IRDME is in active development. Features may change, results may contain errors, and the platform should not be used as a sole basis for clinical, financial, or safety-critical decisions. All structural findings require independent expert validation before any real-world application.
About IRDME
IRDME (Item-Relation-Dimension-Multilayer Engine) is a structural analysis platform. It takes any system you can describe as items connected by typed relations — a protein network, a codebase, a financial market, an organism — and finds which items are structurally important, how layers of relationships compare to each other, and where the topology of a system diverges from what its declared architecture suggests.
The core scientific finding underlying the platform is the Functional Proximity Law: hub importance scores persist more strongly between layers encoding functionally similar relationships than between layers encoding different ones. This pattern has been confirmed across 31 pre-registered experiments: 13 canonical domains (10 confirmed, 3 denied; covering molecular biology, neuroscience, computer systems, ecology, linguistics, and AI architecture) plus 18 external validations on independently-authored datasets (15 confirmed, 1 denied with named boundary condition; extending to particle physics, computational geometry, medicine/epistemology, COBOL legacy software, and cross-species connectomics across 600 million years of evolution). The full scientific record is on arXiv:2604.23639.
What you can do with it
Paste a GitHub URL. IRDME extracts the repo, builds import/structural/co-change layers, and returns hub shadows (hidden maintenance risks), chameleons (silent attack surfaces), and universal hubs (load-bearing modules). No code reading required.
Compare two codebases structurally — e.g. WordPress PHP → Next.js TypeScript. Per-module risk class: safe_transfer, refactor_opportunity, migration_hazard. Flask → Express isomorphism confirmed at sim=1.000 on the universal hub.
IRDME identified TP53 as the top hub across all 4 molecular layers of the p53 network — from topology alone, without reading any biology. TNF-α (the adalimumab target, Humira) was found as the top activates hub in the cytokine cascade the same way.
Compare two independent datasets describing the same system. Hub-agreement nodes are structurally trusted — their centrality is confirmed by two separate sources with different curation methodologies. Hub shadows (hub in one source only) are boundary cases: discovery candidates or data quality gaps. A SHA-256 certificate records the comparison for reproducibility. Meaningful when both sources cover the same entity set.
Model a project plan as a multilayer graph: Gantt (schedule dependencies) and Risk (failure propagation paths). IRDME computes which tasks are universal hubs across both layers — these are the true critical path items. Hub shadows (high Gantt rank, low Risk rank or vice versa) expose tasks where project timelines and risk models diverge.
Merge two independent datasets into a unified multilayer graph with a named inter-system layer. IRDME resolves node-ID conflicts automatically via optional per-dataset prefixes, then computes which nodes are universal hubs across the merged structure — bridging entities that matter in both systems. Useful for cross-organisational network analysis, biological pathway integration, or multi-source knowledge graphs.
Build multi-step analysis chains. Each step derives a subgraph from the previous result (top hubs, divergent nodes, custom). Convergence is tracked automatically — a Jaccard novelty guardrail blocks redundant steps before they waste time.
AIG was identified as the top betweenness hub in the 2008 financial network from topology alone — before reading any FCIC report. The same structural signal (hub in derivatives, mid-tier in credit) is computable on any public market data.
Service model
IRDME is a closed-source service. The analysis engine, graph algorithms, and hypothesis verification system are proprietary and not publicly available. The science — the law statement, the evidence table, the pre-registered hypotheses — is published openly on arXiv and in the Paper page. What you interact with on this platform is a service interface to that engine, not the engine itself.
Pre-registration records (hypotheses committed before analysis runs) are public at github.com/vladi160/preregistrations.
Pre-registration does not guarantee trustworthiness
Pre-registration creates a timestamped public record of hypotheses, but it cannot prove that no private analysis was run before committing. Git commit timestamps are not cryptographic proof of first analysis. Pre-registration here functions as a logging and accountability mechanism — it makes the intended predictions visible and verifiable after the fact — not as a seal of scientific integrity. Treat the pre-registration record as useful context, not as independent verification.
How to interpret results
Results are structural observations, not ground truth
IRDME identifies topologically important items from graph structure alone — without reading domain content. This is a feature (it is objective) and a limitation (it has no domain knowledge). AIG was identified as a systemic risk hub by topology, not by reading balance sheets. TP53 was identified as the #1 protein hub by graph structure, not by biology. Topology is a signal. Expert interpretation is required to turn it into a decision.
CONFIRMED / DENIED verdicts apply to hypotheses, not to domains
When a hypothesis is CONFIRMED, it means the data matched the pre-registered prediction. It does not mean the result is universally true or that the law applies everywhere. When a hypothesis is DENIED, it is a finding, not a failure — each denial names a boundary condition that refines when the law applies.
The engine can produce incorrect results
As alpha software, IRDME may produce errors, unexpected output, or statistically underpowered results when graphs are small (n < 10), layer definitions are ambiguous, or edge weights are missing. Check the r value, the sample size, and the p-value before drawing conclusions. When in doubt, run the null model comparison (irdme null in the CLI) to confirm the pattern is not random.
No warranty for high-stakes applications
Results from this platform must not be used as the sole basis for medical, clinical, financial, legal, or safety-critical decisions. IRDME provides structural signals. Domain experts, additional validation, and appropriate professional judgment are always required.
Scientific integrity
Every experiment on this platform that makes a directional prediction uses pre-registration: hypotheses are committed to a public GitHub repository with a SHA-256 hash before any analysis is run. This creates a public record of the prediction with a git timestamp. Pre-registration is a logging and history mechanism — it establishes what was predicted and when, and makes the full history of predictions (including denied ones) publicly visible.
The arXiv paper was published April 26, 2026. Post-publication experiments have extended the confirmed count and identified two new named boundary conditions: BC_RADIAL (structural degeneracy in radial architectures) and BC_INVERSION (fan-out with leaf clustering). A separate active research program — the M_MED series — applies IRDME to pharmaceutical evidence chains, modeling each as three epistemic layers (justifies, selects_endpoints, cites_as_support). Eight pre-registered experiments across vaccines, statins, antidepressants, opioids, Alzheimer's, schizophrenia, epilepsy, and Parkinson's disease have produced a structural decomposition framework (EM/ETD/MSC axes) and two independently verified citation anomaly detections identified from topology alone without reading the papers. All experiments are separately pre-registered before each run.
Project status
Paper published: arXiv:2604.23639 · April 26, 2026 · Zenodo DOI 10.5281/zenodo.19647148