SOS1 is the top structural DAMPER in KRAS-mutant lung cancer (z_dr=+4.18). Pre-registered prediction: SOS1 should be more CRISPR-essential in KRAS-mutant cell lines. Result: DENIED. KRAS-WT cells are actually more dependent on SOS1 (d=+0.52, p=0.29). But KRAS itself is strongly essential in KRAS-MT (d=-1.34, p=0.0002). Structural DAMPERs predict pharmacological vulnerability. Structural ANCHORs predict genetic essentiality. These are two different axes, and the distinction matters for drug discovery.
In EGFR-mutant LUAD: EGFR is a structural ANCHOR (z=-2.05), KRAS is near-zero. In KRAS-mutant LUAD (n=155, EGFR-WT only): EGFR returns to z=+0.012 -- nearly zero. The driver gene controls which protein becomes the network's organizing center. When you swap the driver, the structural roles swap with it. SOS1 is revealed as a RAS-activated DAMPER (z=+4.18 in KRAS-MT vs -0.19 in KRAS-WT), not a pan-LUAD signal.
EGFR is a structural DAMPER in colorectal cancer (z=+2.41) and a structural ANCHOR in EGFR-mutant lung cancer (z=-2.05). Complete structural inversion. Yet EGFR inhibitors kill cells equally well in both cancers (d=-0.028, p=0.91, 5/5 drugs). This pre-registered result establishes a new principle: two opposite structural positions can produce the same pharmacological output when Driver Addiction Override takes effect.
We have been running a pre-registered series on a single question: can a protein's structural position in a cancer signaling network predict whether a drug targeting it will kill cancer cells? Thirteen experiments across six cancer types, two assay platforms, and two network scales later -- the answer is more nuanced, and more structured, than we expected.
Four pre-registered experiments on a 16-protein NSCLC signaling network. ERK1 is the biggest structural coupling bottleneck -- but DepMap CRISPR screens show MYC is the most essential undrugged node. Structural coherence and cancer survival are orthogonal axes (Spearman r = -0.722, p = 0.0024). The drug development landscape targets nodes in the wrong structural tier.
A full account of the T_MODEL experiment series: we started looking for topology to predict attention head importance, found that activation magnitude (OAN) is a strong gradient-free predictor for binary sentiment classifiers, confirmed that the standard Taylor gradient criterion fails across all three models tested, and discovered that OAN's boundary condition -- where it works and where it doesn't -- may be more informative than the positive result itself.
M_MED8 pre-registered experiment on Parkinson's disease as a competition rule test: dopamine depletion in substantia nigra is one of the best-confirmed mechanisms in neurology, yet UPDRS is an institutionally entrenched symptom scale structurally analogous to PANSS in schizophrenia. Which axis wins? Neither wins cleanly. The Pearson-Spearman inversion identifies two distinct institutional distortion operators: Rank-Preserving Amplification (RPA, Alzheimer's) and Magnitude-Preserving Inversion (MPI, Parkinson's). Eight experiments, three independent axes, two independently verified anomaly detections.
M_MED7 pre-registered experiment on the epilepsy evidence chain. Three mechanistically incompatible frameworks (sodium channel / channelopathy, GABA/glutamate network, autoimmune synaptopathy) coexist without paradigm islands -- the graph is connected. The seizure frequency endpoint is entrenched through convergent selection from all three frameworks. The unexpected structural finding: multi-framework convergent endpoint selection produces stronger r(justifies, selects_endpoints) = +0.490 than single-founder circular chains like antidepressants (+0.41). A new structural class is identified: convergent institutional endpoint lock-in.
M_MED6 pre-registered experiment: the statin cardiovascular evidence chain replicates the healthy topology of M_MED2 vaccines. r(justifies, selects_endpoints) = +0.676, p = 0.030. Healthy class n = 1 to n = 2. The spectrum from healthy to distorted pharmaceutical evidence chains is now anchored at both ends with statistically significant cases. No citation anomaly detected -- supporting the hypothesis that citation anomalies are specific to circular chains.
M_MED5 pre-registered experiment: the Alzheimer's amyloid cascade hypothesis dominates both justification and citation layers (circularity confirmed). Topology identified Lesne 2006 as a structural citation anomaly -- high citation rank, zero justification edges -- without reading the paper. The connected/disconnected contrast with M_MED4 (schizophrenia) is now a comparative structural observation, not a construction artifact.
IRDME analysis of the schizophrenia evidence chain found that the dopamine and NMDA hypotheses have zero structural contact across all three epistemic layers. They are not competing in the same evidence space - they are parallel research islands that have never formally engaged each other. Pre-registered experiment M_MED4_v1.
Across 146 cross-layer analyses in 33 pre-registered experiments spanning 8 domains, we found two structural classes that appear in every domain: Universal Hubs (same node tops multiple layers) and Layer Specialists (different nodes lead each layer). Universal hubs are the infrastructure: TP53, transformer, W boson, VFS, app, thalamus. Layer specialists are the disruptors: LLaMA vs transformer, CHEK1 vs TP53, IRQ vs VFS, A[15] vs A[0]. Both classes are detectable from graph topology without reading any domain content. A third class, Circular Hubs, appears specifically in pharmaceutical evidence chains (monoamine_hypothesis, pain_undertreated_claim).
We reanalyzed all 45 cross-layer hub correlation results across 33 pre-registered IRDME experiments to compare Spearman rho and Pearson r. Overall: Pearson is marginally higher on average (0.510 vs 0.463). But the GAP between them is not noise -- it is a structural diagnostic. Large positive gaps (Spearman >> Pearson) appear consistently in cross-species and cross-domain settings where rank ordering is preserved but magnitudes differ. Large negative gaps (Pearson >> Spearman) appear in within-domain settings where absolute hub centrality scores carry real structural signal. The gap tells you what kind of cross-layer relationship you are measuring.
We ran FPL cross-layer analysis, structural betweenness, and state correlation on ISCAS85 c6288 (16x16-bit Wallace tree multiplier, 2448 nodes). All three identified different nodes as most important. Betweenness: node 6207 (internal final-adder gate). State correlation: node 256 (A[15], MSB -- betweenness=0.00). Structural degree: node 1 (tied with 31 others). Node 256 has state_correlation degree=81 but betweenness=0 -- invisible to single-layer structural analysis. FPL holds strongly (r=0.5004, p=0.002, z=+23.22 above null). 1022/2448 nodes (42%) are cross-layer divergers. Directly answers the critique of H_LOGIC_BLACKBOX_v1. Pre-registered. 4/4 confirmed.
We ran IRDME on a 4-bit ripple carry adder with all gate labels replaced by neutral IDs. 4/4 hypotheses confirmed: the carry-chain gates (G_12=cout_1, G_20=cout_0, G_21=cout_2) were the top structural hubs. Honest scope: a ripple-carry adder has a linear carry chain that is a trivially recoverable structural backbone -- any centrality method finds it. This experiment validates the IRDME pipeline on hardware and establishes the starting point, but does not demonstrate a novel inference capability. The harder test (c6288 32-bit multiplier, FPL vs betweenness baseline) is next.
We mapped the opioid prescribing epidemic evidence chain (1980s-2010s) as a multilayer graph and pre-registered four structural hypotheses. Result: 3/4 CONFIRMED, 1/4 PARTIAL. pain_undertreated_claim is rank #1 in BOTH the justification layer and the citation layer -- the same structural circularity pattern found in antidepressants. porter_jick_letter (the 1980 NEJM 5-sentence letter misused as addiction-safety evidence) was identified as a citation anomaly by topology alone: high citation rank, peripheral justification rank. FPL gradient collapses in circular chains (r=0.14) vs healthy vaccine chain (r=0.75). Pre-registered.
We mapped the childhood vaccine evidence chain as a multilayer graph and compared its structural topology to the antidepressant circularity finding (M_MED1). The vaccine chain confirmed 4/4 hypotheses: the founding mechanism (adaptive immunity) is the justification hub, but empirical outcomes (clinical guidelines) dominate the citation layer. The two are different nodes -- structurally healthy. FPL gradient r=0.75 p=0.014 (statistically significant at n=8). Pre-registered. Contrast: in the antidepressant chain, the same founding assumption was top hub in all three layers.
After confirming FPL in the Standard Model (M_PHYSICS_1, 5/5), we pre-registered structural rank-diverger predictions for all 17 particles: photon and gluon rank top-3 in force coupling but bottom-2 in mass proximity (rank gap=13-14). Null model: all layer pairs significant (p<=0.030). Confirmed 7/10 pre-registered hypotheses across two experiments. One informative denial: higgs tied at degree=3 with bottom, top, and z_boson -- revealing ranking sensitivity to degree ties.
A pre-registered experiment on a 4-bit ripple carry adder: IRDME, given only the physical wiring topology and exhaustive simulation state correlations, correctly recovers the carry propagation chain as the structural hub -- without any symbolic circuit description. Pearson r = 0.5117 (p = 0.004, large effect). The carry bits are persistent betweenness hubs across both layers (cross-layer betweenness r = 0.7714).
MathComp was run as a pre-registered formal-corpus follow-up with corrected h2 schema fields. Result: h1 DENIED (near-zero cross-layer persistence), h2 PARTIAL, h3 CONFIRMED (role divergence), indicating a boundary-style structural regime rather than a clean persistence replication.
The Coq Corelib replication (n=17) produced a positive directional result consistent with Lean/mathlib4: degree r=0.2875 (positive, p=0.2874), with cross-layer analysis showing moderate persistence (r=0.4913) and betweenness persistence r=0.5091. This is an underpowered but important independent formal-systems replication.
The ISCAS85 c432 replication (n=196) confirmed the pre-registered betweenness-persistence hypothesis with r=0.4255 and p=0.002, while degree persistence remained weak (r=0.2138). This extends the digital logic result from the 4-bit adder to a standard benchmark circuit.
X9 v2 reran the prebiotic-chemistry program on a published external autocatalytic network from Xavier et al. 2020 via the CatReNet example-9 dataset. The result was 4/4 CONFIRMED, which upgrades X9 from a model-based signal to an externally replicated one.
Three new IRDME results now point in the same direction: structure is portable across substrates. A worm predicts a fly, topology extracts logic from Lean4, and prebiotic chemistry already shows layered structural organization. For a startup, that means dataset certification, structural priors, and generative compression become products, not only papers.
X9 tested origin-of-life prebiotic chemistry as a three-layer graph: bond chemistry, autocatalytic membership, and replication co-occurrence. The pre-registered run confirmed 3/4 hypotheses, including the FPL rank ordering in this domain.
In Lean4 mathlib4, declared proof dependencies and behavioral proof coupling show strong hub persistence (r=0.777, p=0.004). The result supports a stronger claim: topology does not just describe formal systems, it exposes their operational logic.
We trained a generative model on a 302-neuron worm brain. Then we pointed it at a 3036-neuron fly brain it had never seen. It correctly predicted which neurons would be hubs -- with cosine similarity above 0.94. Pre-registered, confirmed.
How to use one experiment's hub output to seed a comparative analysis. Introduces the --seed-from and --compare-to flags, the role-vector cosine metric, and an exploratory C. elegans generative compression result (avg_seed_cosine=0.845, avg_rank_cosine=0.888). Not a confirmatory pre-registered result - a methodological contribution.
Pre-registered cross-species FPL replication in the Drosophila larval connectome (n=2952 neurons, Winding et al. 2023): h1 CONFIRMED Pearson r=0.363 / Spearman rho=0.663, p=0.002. Hub rank order conserved across 600 million years of evolution. Spearman >> Pearson discrepancy is the primary finding. 25/31 pre-registered, p=0.000439.
F2_cobol_legacy_v1: IRDME on a 14-program COBOL banking application finds r(PERFORM↔COPY)=0.807 (p=0.002) vs r(PERFORM↔data_field_sharing)=0.119 - the Functional Proximity Law holds in 50-year-old procedural mainframe code. Two structurally dormant programs appear as rank #2-#3 hubs in data_field_sharing despite zero PERFORM calls. Rank gap=11 each. Multilayer rank divergence identifies components that are operationally peripheral in execution topology yet central in shared-state topology.
IRDME applied to an 8-node epistemic graph of antidepressant research finds that the monoamine hypothesis is the #1 hub in BOTH the justification layer AND the citation-as-support layer. The founding assumption and the primary evidence are the same node. This is the structural signature of a self-referential evidence loop — detectable from topology alone, without reading a single paper.
The photon is the second most connected particle in the force-coupling layer of the Standard Model — and has exactly zero connections in the decay-channel layer. The first pre-registered IRDME experiment in particle physics: 5/5 hypotheses confirmed. W boson is the universal decay hub. Photon is the first physics domain hub shadow.
A planetary gear assembly modeled as a CSG graph: r(structural ↔ CSG-operations) = 0.668, p = 0.042, Δr = 0.935. Pre-registered before running. Computational geometry via continuous CSG operations is now a confirmed domain. BC3b boundary defined.
cobra (the Go CLI framework behind kubectl, Hugo, and dozens of major projects) produces r = −0.862 (p = 0.014). The declared hub is the least behaviorally active node. This is BC_INVERSION: fan-out with leaf clustering — and it is structurally distinct from BC_RADIAL.
docopt in Go, Java, and Rust all denied the Functional Proximity Law via the same mechanism. When a single hub fans out to all leaves, the behavioral layer's hub-rank vector approaches a constant — and Pearson r collapses toward zero. Named mechanism: BC_RADIAL.
A structural audit of the 2008 financial crisis network: AIG was mid-tier in declared counterparty contracts but the dominant hub in derivatives flow. The gap between declared and behavioral rank is a textbook hub shadow — visible from topology alone, without reading a single balance sheet.
F1 pre-registered result: CHEK1 and PARP1 are the top divergers between functional proximity and genetic interaction layers in the p53/DDR network — correctly predicted from topology. The Pearson r is too low to call CONFIRMED, but the structural signal is real and the hub-identity predictions are precise.
F12 pre-registered result: the Functional Proximity Law holds across the full C. elegans 302-neuron connectome (n=279 with synapses). 13 of 15 command interneurons from the smaller M_EXT2 study appear in the top-20 hubs — hub identity preserved at 20:1 scale compression.
BC3 denied the Functional Proximity Law in mathematics. F9 confirmed it — using a different layer pair on a different dataset in the same domain. The comparison isolates exactly which structural feature caused the denial.
F6 pre-registered result: Transformer is the top hub by citation lineage and architectural inheritance, but LLaMA leads current benchmarks. Two layers, same 20 models — structurally divergent hubs reveal the gap between historical influence and contemporary relevance.
A pattern confirmed across 16 different domains: the nodes most likely to cause catastrophic failure are structurally invisible to the models designed to find them. We call this the hub shadow.
The tasks most likely to fail your project don't appear on the critical path. IRDME's Project Risk Topology Report finds them by comparing declared schedule dependencies against resource sharing and risk coupling.
We pre-registered 5 structural hypotheses about the p53 protein interaction network before running any analysis - then certified agreement between a curated BioGRID/DepMap dataset and STRING v12.0. All 5 confirmed. TP53, MDM2, ATM, BRCA1, and CHK2 are structurally robust across independent sources. CHEK1 and PCNA are boundary cases - hub in STRING only.
Godot scores r = 0.925. PostgreSQL scores r = -0.120. Both are large, mature, well-maintained open-source codebases. The difference is not code quality — it is governance and commit discipline.
We ran the structural hub analyzer on 18 major GitHub repositories and stored every result publicly. 10 of 18 confirmed the IRDME structural law (r >= 0.3). The pattern splits cleanly by software category.
A plain-language explanation of the Functional Proximity Law — why hub nodes in a network tend to stay important across multiple types of relationships, and what this means for science.
Science is only credible when predictions precede results. Here is the exact pre-registration protocol used for every Functional Proximity Law experiment — SHA-256 hashes, UTC timestamps, and public commits.
The Functional Proximity Law is denied for the 2008 financial crisis — r = 0.042, p = 0.824 between derivatives counterparty and direct credit exposure. This is not a failure. It is the finding. The denial has a named mechanism: institutional regime mismatch.
post.php ranks #19 in import graph centrality but #4 in co-change centrality — a rank gap of 15. It is structurally peripheral, behaviourally central. Every migration that ignores this distinction will break.
The C. elegans connectome experiment uses topology alone to identify PVCL and PVCR as primary command interneurons. r = 0.7774, p = 0.004. The layer definitions and node rankings come from White et al. 1986 — fully independent of IRDME. This is true external validation.