We found a hidden geometry inside every Transformer.
We built the tools to open the black box.
Lossless realignment. Deterministic 2× compression without calibration data. Spectral fingerprinting for supply chain integrity.
Canonical Basis Realignment
Lossless affine transformation rotating Transformer weights into an interpretable coordinate system where each axis is independently measurable and manipulable.
Rich Club & Bipolar Oscillator
Positive-pole and negative-pole axes in rhythmic alternation across layers. Content-invariant structural signature — the model breathes regardless of input.
Knowledge Injection & Transplant
Cross-model knowledge transfer via correlation structure mapping. Teacher pole cohesion injected into student at empty axes. Weight modification persists through all layers.
Anti-Hallucination Gate
POS/NEG pole ratio at output layer as hallucination detector. POS dominance signals fabrication; NEG dominance signals filtered output. Spectral equalizer amplifies the filtering pole in real time.
Homeostasis Discovery
The residual stream erases any intermediate perturbation within two layers. Architectural stability enforced by normalization, attention statistics, and FFN operating ranges.
Spectral Collapse
Singular value magnitudes are redundant across layers. The model operates on geometry, not intensity. The spectrum can be reshuffled — output unchanged.
Cross-Architecture Validation
U-alignment emerges on any architecture using RMSNorm - FFN type (SwiGLU or GELU MLP), LayerNorm. Tested on decoder-only and encoder-only; encoder-decoder pending.
LayerNorm→RMSNorm Bridge
Replace LayerNorm with RMSNorm on Vision Transformers — canonical basis becomes available. Lossless realignment verified. Enables per-axis precision, spectral fingerprinting, and compression on ViT without retraining.
Strategic Resonance
Targeted weight perturbation creates a pressure differential amplified by model dynamics. Output-level energy redistribution captures the gain without any training.
Deterministic Geometric Quantization
Int8 quantization from weight geometry alone — no calibration data, no inference pass. Block-aligned scales capture local structure. Matches fp16 perplexity on tested architectures.
Block-Scales Geometry
GGUF format fingerprint originates from per-block quantization scales, not packed weight nibbles. Scales randomized → fingerprint collapses. Scales preserved → fingerprint intact.
LASER Effect
Low-rank Actually Surpasses Exact Reconstruction. During training with rank constraint, the compressed model can exceed the uncompressed original. Validated across multiple scales.