Submission for Kaggle Benchmarking Challenge

What I Benchmarked
Measured multi-turn agent context decay across 50 steps. Linear KV-cache and flat RAG suffer topic bleed, dropping recall from 97.2% to 60.8%.

Terminal
Figure 1: Benchmark harness with 14-Gate BBQ verification on Apple M4.

Models Tested

  1. XORAS 18,432-D Poincaré Memory
  2. Gemini 3.5 Pro
  3. Gemma-4 26B
  4. Flat Euclidean RAG

Findings
Decay Chart
Figure 2: Accuracy recall across 50 steps.

  • Flat RAG & Gemma-4: Accuracy drops to 60.8% with 48.6% state pollution.
  • Gemini 3.5 Pro: Degrades to 60.5%.
  • XORAS Poincaré Memory: Maintains 99.8% recall and 0.00% pollution.

Telemetry
Figure 3: POSIX SHM 3.16M ops/s throughput & Z3 SMT proofs.

My Benchmark