Measured Results.
Reproducible Methods.

Training Efficiency Proof

This is the strongest confirmed training benchmark in the public stack: layered SCBE supervision improved loss at fixed compute without increasing model size or training time.

Headline Results

Detection Configuration Sweep

Eight configurations tested on the same 91-attack corpus + 15-sample clean calibration set. Each configuration adds or modifies detection signals.

Key finding: E4 (semantic + triple-weight remainder) provides the best balance: 85.7% detection with 0% false positives. Adding null-space features (E5 variant) pushes to 100% detection but the holdout false positive rate needs calibration.

Null Space Ablation

Incremental feature test: does adding null-space absence detection improve results?

Key finding: Null-space features are maximally diagnostic (100% detection) but the threshold is too aggressive on held-out benign data. The absence signal is real -- it just needs proper calibration on a larger holdout set.

Attack Corpus (10 Classes, 91 Attacks)

Purpose-built adversarial corpus covering the full OWASP LLM Top 10 + SCBE-specific attack vectors.

Reproduce: python -c "from tests.adversarial.attack_corpus import get_all_attacks; print(len(get_all_attacks()))"

Industry Comparison

SCBE detection vs ProtectAI (industry-standard prompt injection detector) on the same 91-attack corpus.

Compliance Spectrum

Where SCBE-AETHERMOORE sits across the public-to-classified compliance landscape.

Null Space Tongue Signatures

Attack types identified by what is ABSENT in their tongue profile, not just what is present. The pattern of empty dimensions is a unique fingerprint per attack class.

Key insight: Attacks that use narrow language (encoding, exfiltration) leave characteristic holes. Direct overrides that try to sound legitimate across all domains activate all 6 tongues -- which is itself suspicious because normal text rarely fills all six.

Reproduce: python scripts/benchmark/unified_triangulation.py (see null_space_analysis section)

Sacred Tongues Tokenizer Verification

Bijective encoding verification across all 6 Sacred Tongues (1,536 total tokens).

Reproduce: python -m pytest tests/crypto/test_sacred_tongues.py -v (45 tests, all passing)

Reproduce Everything

# Clone and install
git clone https://github.com/issdandavis/scbe-aethermoore-demo
cd scbe-aethermoore-demo
pip install numpy

# Run benchmarks
python scripts/benchmark/spectral_sweep_benchmark.py      # 8-config sweep
python scripts/benchmark/null_space_ablation.py            # Null space A/B/C
python scripts/benchmark/unified_triangulation.py          # Combined + null patterns
python scripts/benchmark/scbe_vs_industry.py               # vs ProtectAI

# Run adversarial test suite
python -m pytest tests/adversarial/ -v                     # 91 attacks, 10 classes

# Run Sacred Tongues verification
python -m pytest tests/crypto/test_sacred_tongues.py -v    # 45 bijectivity tests

# Run pump tests
python -m pytest tests/test_polly_pump.py -v               # 3 orientation tests
        

All benchmark scripts output JSON to artifacts/benchmark/. Results are deterministic (no randomness in detection logic).

Known Limitations

• Keyword-based tongue scorer: Current tongue profiling uses 80+ keywords across 6 academic domains. A 10-word query may only hit 2-3 keywords. This is a ceiling that needs replacement with a learned projector.
• Null space FP rate: Null-space features push detection to 100% but the holdout false positive rate is 100% -- the threshold is too aggressive. Needs calibration on a larger benign holdout set.
• Single-point d_H equivalence: Hyperbolic distance from origin is a monotonic transform of Euclidean norm (disproven Feb 5, 2026). Use phase+distance or trajectory curvature instead.
• Corpus size: 91 attacks is small. Industry benchmarks (HarmBench, StrongREJECT) use 1000+. Expanding the corpus would improve confidence.
• No real-world deployment data: All results are on synthetic benchmarks. Production deployment would reveal edge cases not covered by the current corpus.

SCBE-AETHERMOORE · Issac Davis · Patent Pending USPTO #63/961,403

GitHub · HuggingFace · PyPI · npm