Per Ardua
Research Program
Structural Compression Theory — 28 papers across neural network geometry, multi-agent coordination, organizational physics, market mechanisms, and applied analysis
Activation Geometry
Neural network training dynamics and model composition through activation-space analysis
Training Trajectory Structure: Regime Detection, Cross-Seed Convergence, and Speculative Weight Prediction
Consolidated paper. Three-regime taxonomy via activation fingerprints generalizes across 124M-7B. Cross-seed convergence: CV 0.41-2.43%, phase boundaries synchronized to +/-50 steps at most scales. Linear prediction achieves 60-90% strict acceptance at 7B in stable regimes. Universal momentum catastrophe: 100-10,000x loss inflation at all scales. Supersedes Leap+Verify and Ensemble Collapse.
Constellation-Indexed Model Composition
Dynamic specialist composition via activation fingerprints. Generalist-space indexing: 20.6% to 98.1% win rate. Stochastic resonance at sigma=0.020.
The Shape of the Problem: Domain-Invariant Structural Signatures
INLP erases domain signal to near-chance while shape classification holds at 95.6%+. Clean double dissociation of domain and structure.
Capability Manifold Surveillance: Topological Detection of Model Distillation
Activation fingerprint geometry as security primitive. AUC 1.000 on systematic attacks at 1.5B+. Economic deterrence via temporal accumulation and coverage-clustering tradeoff.
GenAI Is Socially Awkward: RLHF Instruction Tuning Damages Social Cognition at Small Scale by Suppressing Pragmatic Inference
RLHF hurts social cognition at 7B (-18 points) but helps at 72B (+6). The deficit is processing mode: compliance training suppresses fuzzy pattern matching social reasoning requires.
The Geometry of Failure: Why Direction-Space Interventions Cannot Achieve Domain Selectivity in Neural Networks
Consolidated paper. Shaped noise breaks 100% of repetition loops but cross-domain selectivity fails. Layer-resolved selectivity (peak 0.57 at layer 10), spectral isotropy (INLP/random amplification ratio 0.991), the concentration barrier theorem (any k directions capture at most k/d_eff of activation variance), and information content (1-2 domain-specific bits against ~15 total) converge: the limit is geometric. Classification accuracy does not imply intervention precision; the forward pass is an isotropic amplifier that provides no directional leverage.
Causal Basis Discovery for Domain-Selective Noise Injection
INLP is the only basis with positive selectivity (+0.618). All causal bases (patching, contrastive, gradient) are anti-selective. Classification-intervention dissociation established.
The Inter-Instance Compression Barrier: Domain-Specific Information Loss at the Natural Language Interface
NL interface is a uniform lossy channel, dropping ~80% of activation information at layer 10 proportionally across domain-specific and domain-agnostic subspaces. The concentration barrier is the sole constraint.
The Dissociation of Geometry and Function: Why Activation-Level State Transfer Cannot Compete with Text
Consolidated paper. Four experiments on activation-level state transfer converge: geometric similarity does not imply functional equivalence. Text preserves 63.7% of reasoning trajectory; activation injection adds 1.2%. Priming selection (activation-guided input) closes 48.9% of the expert gap, outperforming injection by 5.4x. Domain reversal: text preserves legal best, activations preserve science best. Supersedes Projection Transmission, Bandwidth-Fidelity, Continuation Perplexity, and Ensemble Gravity.
Context Management in Autoregressive Language Models
Consolidated paper. Nine prompt-structure experiments on Qwen 2.5-7B, with key results replicated on Mistral-7B. A reset instruction before domain priming cuts cross-entropy 39%. 15 tokens of priming capture 98.8% of the benefit. Meta-commentary hurts. Mode switching, not garbage collection. Chain degradation saturates logarithmically by hop 5. Cross-domain handoffs outperform within-domain.
Universal Entanglement in Transformer Activation Space
Consolidated paper. SVD directions can be concept-pure for classification yet carry all concepts in activations (minimum cross-concept damage 38.8%). EI > 1.0 at the terminal layer of all four models tested. Random Gaussian projections to 448+ dimensions reproduce learned entanglement intensity (EI=1.50 in Qwen-7B); PCA to 28 dimensions reduces it below 0.1. Superlinear amplification: triple EI is about 2x mean pairwise.
The Entanglement Theorem: Structural Concept Coupling as a Geometric Consequence of High-Dimensional Encoding
Proves via concentration of measure that, for a given informative subspace rank, every informative direction carries all encoded concepts. Specialist bound: a geometric argument for compositional architecture. Alignment implication: surgical concept editing is geometrically limited.
Entanglement Under Fine-Tuning: Scale-Dependent Disentanglement, Crosstalk-Guided Companion Selection, and the Creative Writing Falsification
Consolidated paper. Code+NL fine-tuning cuts entanglement intensity 32% at 7B, 93% at 14B, and 100% at 32B (Qwen-2.5-Coder; EI = 0.000 in all 8 seeds at 32B). At 7B the response is architecture-dependent (CodeLlama rises 31%). A math companion beats an NL companion on HumanEval+ (48.2% vs 36.6%, p = 0.016), but the lowest-crosstalk companion, creative writing, falsifies naive crosstalk minimization. CSR is the best fine-tuned condition (49.4%).
Bridge Theory
Unifying frameworks connecting neural and organizational phenomena
Organizational Theory
Structural forces, information dynamics, and governance in organizations
Multi-Agent AI: Substrate-Independent Dysfunction
Token governance overhead scales with communication links, not agent capability: fully connected governance grows from 25.5% at n=3 to 69.4% at n=15, with CV=0.0000 across capability levels.
Ambient Structure Discovery via Stigmergic Mesh
22,500 Monte Carlo runs. Coordination failure TPR 0.82 at FPR<0.15. Patent filed USPTO #63/981,369.
Emergence: A Programming Paradigm for Constraint-Shaped Agency
WF1-WF4 well-formedness definitions. Q-dynamics contraction. Tessera as existence proof.
Contraction Dynamics: Constraint Typing and Gaussian Fixed-Point Convergence
Pure mathematics. Gaussian Contraction Theorem. Constraint type system. Five open problems.
The Strategic Rate-Distortion-Perception Tradeoff
Info-theoretic bridge. RLHF mapping: sycophancy as generative distortion. Conformity pressure accounts for 22-72% of receiver distortion across representative parameter regimes.
Applied Analysis
Empirical investigations of organizational mechanisms
Tournament-Based Performance Evaluation and Systematic Misallocation
Agent-based simulation of 994 engineers in 142 teams. Forced ranking misclassifies 32% under random team assignment and 53% under realistic conditions; the paper proposes sparse cross-team Bradley-Terry comparison as an alternative. A $10M pool is misallocated by $1.6M–$2.7M per cycle.
Structural Immunity: Platform Dominance and Corporate Impunity
Six-filter multiplicative pipeline calibrated against CFPB data: 0.75% baseline survival; about 0.002% (2 of 80,000 harmed consumers) in consumer finance.
The Inversion of Doubt: Skepticism Inverts Epistemic Warrant
Cross-domain metascience synthesis. Three structural mechanisms. Composite threshold from RR divergence, calibration gap, and falsifiability timescale.
The Topology of Influence: Latent Structure and Resilience in Online Populations
1.4B Reddit comments, 44,837 subreddits. 12-dimensional latent space. Influence wave propagation (ρ=0.27). Simpson's paradox in leadership (18.5x per-dimension, 1.0x global). Topology predicts resilience (ρ=0.56).
Decentralized Drone Swarm Coordination via Broadcast-as-Shared-State and Hierarchical Bisection
One primitive (broadcast-as-shared-state with PCA-tree-keyed local computation), four provable layers (assignment, recovery, priority, localization). 1.43% gap from Hungarian at N=10,000. 14× fewer messages than CBBA at 3.0% gap. Composition theorem: the four layers run on one broadcast without interference. Witness-alarm byzantine defense. v1.1 adds comms-layer empirical validation. v1.2 (§9.4) extends across drift handling, online replanning, and Bayesian SAR with 2220 runs across centralized SAROPS-class PF (Allen Leeway 85 object classes), Bandyopadhyay-Chung consensus filter T-sweep, and a voting baseline that empirically refutes the “voting alone is sufficient” framing. Asymmetric-deaf cliff at n/2 confirms voting-theory threshold in the SAR setting. Phase A literature audit positions the work in canonical decentralized-Bayesian-SAR (Bourgault 2003), consensus-on-PDF (Bandyopadhyay-Chung 2014/2018), and connectivity-preservation (Minelli 2020 RAS) lineages. Bootstrap CIs throughout.
Entropy-Bounded Decomposition for Time Series Forecasting
Layered analytical decomposition (Fourier, trend, shocks, AR, wavelets) with a maximum-entropy stopping criterion. Beats TimesFM by 65-73% on periodic ETT data (ETTh1, ETTm1), loses on ETTh2 and ETTm2; residual autocorrelation predicts the split. No training, no GPU.
Welfare-Optimal Matching in Hiring Markets via Min-Cost Flow
MCF-Tiered: 95.3% vs greedy 24.5% company satisfaction. Adaptive β=0.7σ(S). 3 datasets, scorer-independent. p<10⁻²⁰.
No Concerns: Conjunctive Hiring and the Dilution of Exceptional-Skill Selection
Selectivity is not aim. With k mandatory skill thresholds, enrichment for exceptional ability on any one skill equals the inverse of its pass rate and falls toward 1 as k grows: a ten-box process rejecting 90% of applicants lifts top-1% specialists from 1.00% to 1.26%. Once skills are pass/fail, no rule beats all-pass. Whether it matters depends on the payoff structure of the work.
Three Headlines from One Equation: A Replication and Sensitivity Analysis of “The AI Layoff Trap” (Falk and Tsoukalas, 2026)
Replicates the model and verifies all ten propositions, then shows the policy conclusion is a parameter choice: baseline inputs give catastrophe and an automation tax, standard-economy inputs give stability or under-automation, optimistic historically grounded inputs prescribe a subsidy. Three parameterizations, three headlines, one equation.
Academic Monograph
The unified theoretical framework
Structural Compression Theory: A Unified Information-Theoretic Account of Organizational Dysfunction, Creativity, and Substrate-Independent Selection Dynamics
Academic monograph unifying the consolidated paper corpus into a single formal framework. Proves that compression under selection produces systematic drift from reality toward internal fit — and that the mechanism is substrate-independent across cognition, organizations, and AI. Three theorems, one lemma, five sufficient conditions, and the inseparability corollary: dysfunction and creativity share a single channel, distinguished only by selection regime. 22 chapters across five parts, validated across 500 computational configurations with zero counterexamples. Derives the Hallucination Corollary: you cannot eliminate hallucination without eliminating generalization.
In Progress
Active research awaiting formal publication