v2.0 - Cold-Gated Re-ranking & Benchmarks

What's Changed

Features & Core Improvements

• Cold-Gated Re-ranking: Fixed hot-item stability issues from v1. The neural network now strictly scores cold items (count == 0), preserving MTF order for hot items in #4 @Elitsuv
• Noise Guard (min_evidence): Added access threshold gating to prevent singleton queries from causing cache thrashing in #4 @Elitsuv
• Benchmark Suite: Added 8.1M multi-regime benchmark suite testing stationary Zipf, workload drift, and noise degradation in #4 @Elitsuv
• Test Suite: Added 16 unit tests with 100% pass rate covering lookups, invariants, and edge cases in #4 @Elitsuv

Documentation

• README Overhaul: Added verified benchmark score tables, architecture overview, and test badges in #5 & #6
  @Elitsuv

Full Changelog: v0.1.1...v2.0

v0.1.1

v0.1.1 – Initial Prototype Release

What's Included

• Core adaptive logic (cyla/engine.py)
  • Tiny neural network scorer (4→12→1 MLP, 61 parameters)
  • Move-to-Front on early hits
  • Periodic NN-based prefix reordering (top 22%)

Known Limitations

• Early-phase overhead still high (negative gain in short runs)
• Learning convergence slow on current reward shaping
• Reorder cost noticeable on deep hits

Next Milestones (planned)

• Stronger reward shaping & lr tuning
• Pure MTF fallback (remove NN overhead)
• Real dataset tests (words, products, logs)
• Theoretical analysis (amortized cost)
• Visualization improvements

v0.1.0-beta – First Prototype

First – learning is happening but currently slower than naive linear scan.

Real numbers (5k items, 20k Zipf queries α=1.5):
Naive: ~140 steps/query
CYLA: ~230–250 steps/query
→ -67% efficiency (1.67× slower)

Included:
• engine.py (core logic)
• test.py (passing tests)
• basic benchmark

Planned next: Neural Network Implementation
Feel free to share your ideas for next release
@Elitsuv