Evidence

Measured performance. Honest comparison.

Every number was measured on local consumer hardware - mostly a Dell Inspiron 3582 running Zorin OS. No cloud instances. No GPU acceleration. No cherry-picking. Last verified July 29, 2026.

Methodology

  • Hardware: Dell Inspiron 3582 (Intel Celeron N4000, 4GB RAM, eMMC storage) and a consumer laptop. No server hardware. No GPU.
  • Software: Python 3.12, Rust 1.78+, SQLite 3.45+, Zorin OS / Linux. All benchmarks run with default settings.
  • Reproducibility: Every benchmark has a reproducible command. Clone the repo, activate the venv, run the command.
  • Honesty: We publish negative results too. The STRATA full-repo scan takes 21 seconds. The Fragment cold-start is 349ms.

Cognitive Performance

MetricValueContextHardware
Skill retrieval latency<1msvs 200ms-2s for LangGraph/AutoGen LLM planningDell Inspiron 3582, Zorin OS
Cross-domain routing accuracy100%Trading to Governance to Analysis, no LLM reasoning passDell Inspiron 3582
Concurrent skills loaded50928ms bootstrapping overheadDell Inspiron 3582
Memory retrieval (406K memories)<5msSQLite + Rust FFI, full scanDell Inspiron 3582
Reproduce: Skill retrieval latency
python scripts/system_capability_test.py --suite routing
Reproduce: Cross-domain routing accuracy
python scripts/system_capability_test.py --suite routing
Reproduce: Concurrent skills loaded
python scripts/system_capability_test.py --suite hotload
Reproduce: Memory retrieval (406K memories)
python scripts/system_capability_test.py --suite memory

Vector & Text Search

MetricValueContextHardware
HNSW vector search0.26ms16,219 embeddings, disk-persisted indexConsumer laptop
FTS5 full-text search (1K memories)2.6ms medianPhrase-first ranking, SQLite FTS5Consumer laptop
Session record per turn0.5msSQLite store, ~3ms total overhead per conversation turnConsumer laptop
Session recall (20 turns)0.6msTag-filtered search + client-side sortConsumer laptop
Reproduce: HNSW vector search
python -m pytest tests/unit/test_hnsw_index.py -v
Reproduce: FTS5 full-text search (1K memories)
python -m pytest tests/unit/test_sqlite_backend.py -k fts -v
Reproduce: Session record per turn
python -m pytest tests/unit/test_session_recorder.py -v

SIMD / Rust Acceleration

MetricValueContextHardware
Rust AVX2 GEMV (256x256)563us12.5x speedup over Python scalar (5.6ms)Consumer laptop, AVX2
Ternary kernel (12-layer model)67ms/tokenvs 873ms/token Python - real-time vs noticeable lagConsumer laptop, AVX2
PredictiveCoder (dim=128)19x speedupRust PyO3 vs Python scalarConsumer laptop
Batch cosine similarityAVX2 acceleratedN-parallel vector comparisons in single SIMD passConsumer laptop, AVX2
Reproduce: Rust AVX2 GEMV (256x256)
cargo bench -p wm-neuro --bench ternary
Reproduce: PredictiveCoder (dim=128)
cargo bench -p wm-neuro --bench predictive

Infrastructure & Dispatch

MetricValueContextHardware
MCP tool dispatch (median)29-33ms3-10x faster than typical MCP implementations (29-33ms vs 100-300ms)Consumer laptop
MCP tool dispatch (P95)36-55ms95th percentile latencyConsumer laptop
MCP tool dispatch (P99)38-86ms99th percentile latencyConsumer laptop
Success rate100%Across all tool callsConsumer laptop
Memory per call0-0.18MBHeap allocation per dispatchConsumer laptop
Throughput29.38 req/sSustained requests per secondConsumer laptop
Homeostatic loop overhead0.35msFull loop including physical checksConsumer laptop
Rate limiter pre-check452K ops/sRust EventRing, zero allocation hot pathConsumer laptop
Reproduce: MCP tool dispatch (median)
python scripts/benchmark_gauntlet.py --suite mcp

Probabilistic Data Structures

MetricValueContextHardware
HLL cardinality tracking508KB / 100K memories1.2% error rate, O(1) space - runs forever without degradationN/A (mathematical)
Count-Min SketchSub-1KBFrequency estimation with bounded errorN/A (mathematical)
Thompson sampling banditBeta posterior updatesContextual tool selection - agent gets better with use without retrainingN/A (algorithmic)

Scale & Stress Tests

MetricValueContextHardware
Memories indexed406,836SQLite + Rust FFI, sub-5ms retrieval maintainedDell Inspiron 3582
Events processed (Live Era)33,297voice_expressed, memory_created, oracle_cast, pattern_detectedDell Inspiron 3582, Zorin OS
Memories (v15.8 era)111,6652.2 million associationsDell Inspiron 3582
Current memories (v24.0)89,500+14 galaxies, HNSW + FTS5 dual searchConsumer laptop
Test suite8,268 passing56 skipped, 0 failing, ~120s runtimeConsumer laptop
Reproduce: Test suite
cd core && python -m pytest tests/ --ignore=tests/archive_v14 --ignore=tests/archive_v11 --ignore=tests/archive --ignore=tests/archive_polyglot --ignore=tests/legacy --ignore=tests/adhoc --ignore=tests/verify -q --timeout=30

Known performance gaps (honest)

  • STRATA full-repo scan: 21.2s. Scanning ~720 Python files with 80+ checkers. The survey mode at 3s is the practical quick check.
  • Fragment cold-start: 349ms. First query builds the BM25 index from disk. Warm-cache benchmark needed.
  • No AVX-512 support. Only AVX2 is used. AVX-512 with cache tiling could achieve 92% of theoretical peak.
  • No speculative decoding. The inference router cascades by escalation, not parallel draft+verify.
  • Streaming inference stubs. The Rust streaming inference layer has compute_layer stubs that return input unchanged.