Context embedding lab
Dual-space retrieval, tongue compression, and 21D state comparison. Best next step is replacing synthetic embeddings with a real encoder baseline.
This section is the public proof surface for SCBE. It collects benchmark summaries, replication notes, mathematical framing, and experimental tracks so technical readers can inspect claims without forcing the homepage to carry the whole argument.
The board below is the readable summary. The repo remains the audit surface. Use both together.
On 56 held-out attacks and 32 hard negatives, the opt-in DeBERTa gate reached 93% recall with a 34% benign false-positive rate. The recall lift comes from the classifier, not the geometry; model-only hits escalate for review.
Publication-safe scores, rules and forum Research Watch, experiments, npm packages, repositories, and claim boundaries.
Useful for catching regressions against project-authored attacks. It measures corpus fit, not generalization.
Promising, but still synthetic until a real encoder baseline is swapped in.
Useful as explainability instrumentation even before they become headline claims.
Explains the benchmark split and the real decision path through L3, L7, L12, and L13.
Articles, discussions, and Polly AI assistant. Connected to Medium, GitHub Discussions, and the SCBE research domain.
Full spec sheets for every detection, response, and counter-attack module. All priced by H(d,R) = R^(d²).
Build outward in tracks: public summaries first, charts second, then live explainers when the underlying method is stable enough.
Dual-space retrieval, tongue compression, and 21D state comparison. Best next step is replacing synthetic embeddings with a real encoder baseline.
A-to-Z trajectories instead of A-to-B endpoints. Good for showing where embeddings oscillate, reverse, and settle.
Frequency, amplitude, coherence, spin, tongue dominance, and settling as six views on the same underlying signal.
Link people from summaries into the real documents rather than trying to make the landing page carry the full weight.
Two-layer intelligence model, deterministic control shell, and hyperbolic permission space.
The six sacred tongues, phi-weighting, and semantic decomposition.
Good bridge document between symbolic language, signal view, and experimental output framing.
A public architecture note collecting the harmonic wall, Sacred Tongues, phi-oscillator, and the design rules that tie the math to the story.