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.
WHITEPAPER
Apr 2026
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.
WHITEPAPER
Mar 2026
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.
WHITEPAPER
Mar 2026
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.
WHITEPAPER
Mar 2026
WP-04
Wormhole Training — Encoder-Decoder Attention Supervision
Encoder attention patterns supervise decoder LoRA training. Cross-domain transfer measurable with domain-dependent decay.
WHITEPAPER
Mar 2026
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.
WHITEPAPER
Feb 2026
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.
WHITEPAPER
Feb 2026
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.
WHITEPAPER
Jan 2026
§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.
PAPER
Apr 2026
§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.
FIELD NOTE
Apr 2026
§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.
PAPER
Apr 2026
§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.