§ PUBLICATIONS · WHITEPAPERS AND RESEARCH OUTPUTS

The record.

Every whitepaper, study session, and field note produced inside the GAUGE proof programme. Negative results and architectural collapses are published with the same prominence as wins.

WP-08
Attention-Mass Singularities as Domain-Algebraic Signal
Decoder attention singularities characterize algebraic domains. Sharpness ordering, non-overlapping entropy, and Fisher information dominance across five trained domains.
WP-07
Adapter Sensitivity Analysis
100 configs x 5 seeds Sobol sensitivity. Adapter levels (S1=0.291) and training samples (S1=0.223) dominate; bottom-heavy rank beats top-heavy; flywheel is multiplicative.
WP-06
GEO-SEO Independence and Readability Dominance
GEO and SEO optimization are independent (r=0.110). Readability dominates generative engine output positioning across all tested configurations.
WP-05
Gap Detection via Wormhole Transit Matrices
Wormhole-based gap detection achieves F1=0.9999 versus random baseline F1=0.317. Attention-mass concentration as a knowledge-gap signal.
WP-04
Wormhole Training — Encoder-Decoder Attention Supervision
Encoder attention patterns supervise decoder LoRA training. Cross-domain transfer measurable with domain-dependent decay.
WP-03
Cross-Domain Transfer in Multi-Adapter Architectures
Measuring how knowledge transfers between domain adapters. Transfer is measurable but decays as a function of algebraic distance between domains.
WP-02
Multi-Layer Agreement for Semantic Detection
MLA improves over single-layer semantic detection. Agreement across transformer layers provides a stronger signal than any individual layer.
WP-01
Semantic Inflation in Transformer Representations
Norm growth documented across 15 models. Last-layer representations dominate STS-B correlation, consistent with semantic inflation hypothesis.
§016
JEPA on S4 -- Dense Solves, Block-Diagonal Does Not
Dense JEPA recovers S4 multiplication at 100% across three seeds. BD4 predictor sits at 4.4% -- the frozen-random floor.
§015
The EML Paradox -- Better LM, Worse Symbolic Regressor
EML trees can in principle compute any elementary function. Gradient descent recovers sin(x) in 0 of 12 configurations.
§014
Frozen Probes on BD4 Representations
BD4 frozen-probe accuracy is 3.4x that of random initialization, confirming that block-diagonal structure learns useful representations even when composition fails.
§013
Power Analysis -- Every Comparison Underpowered
Ten benchmark comparisons at n=100; ten of ten underpowered. MMLU at measured effect size demands n=1,790 for 80% power.
12entries · newest first · Refleqt Labs, 2026