Benchmarks & Methodology
Glazyr's published performance metrics are derived from two distinct empirical benchmarks. This page documents our measurement methodology, hardware environment, and exact claim reconciliation.
What We Measure
We measure performance across two separate operational boundaries: data extraction speed and the full perception loop. We label these distinctly across all documentation.
Semantic Extraction (The “>100x” Claim)
Measures the elapsed time to obtain the page's full semantic tree (the structured DOM/AX tree an AI agent reasons over), starting from a fully loaded DOM state and ending with the populated data structure in agent process memory.
| System | Method | Latency Result |
|---|---|---|
| Playwright / CDP | Serialize tree to JSON & transmit over CDP WebSocket | ~400 ms |
| Glazyr Viz | Zero-copy read directly from shared memory (SHM) | ~3.0 ms |
* Isolated data-extraction measurement: excludes page load, network transit, and LLM inference.
Perception Loop (The “7.35ms” and “15–50x” Claims)
Measures one complete perception cycle: reading page state directly through the GPU compositor buffer, converted to structured state with deterministic (x, y) action coordinates ready for agent dispatch.
| System | Perception Mechanism | Step Latency |
|---|---|---|
| Traditional Agents (Playwright / CDP) | CDP serialization overhead + heuristic pixel coordinates | 120 ms – 400 ms+ |
| Glazyr Viz (Direct GPU Buffer) | Direct GPU buffer access + deterministic exact (x, y) targeting | 7.35 ms |
* The headline 7.35ms perception figure represents the median GPU buffer perception loop step.
Claim Reconciliation
Every performance claim and metric displayed across glazyr.com correlates to our benchmark tests as reconciled below:
| Public Claim | Location | Benchmark | Empirical Basis |
|---|---|---|---|
| 7.35 ms Perception | Homepage Hero, Technology | Benchmark B | Full perception step, direct GPU buffer access |
| ~3.0 ms Extraction | Homepage | Benchmark A | Semantic tree read directly from shared memory |
| Over 100x Speedup | Homepage | Benchmark A | ~400 ms CDP serialization vs ~3.0 ms SHM read |
| 15–50x Faster | Homepage, Technology | Benchmark B | 120–400 ms CDP range vs 7.35 ms GPU buffer |
| 120 ms – 400 ms+ | Comparison Table | A/B Baseline | Playwright / CDP per-step latency range |
| ~4.1 kb Payload / Step | Comparison Table | Baseline | Extracted text payload in traditional CDP stack |
Test Environment
To maintain scientific rigor and defensibility, our test environment parameters are declared below:
- Machine
- [FILL: CPU model / cores]
- GPU
- [FILL: model — GPU buffer path]
- RAM / OS
- [FILL: RAM size / OS distribution]
- Glazyr Build
- [FILL: version / commit]
- Chromium Build
- [FILL: Chromium version (identical across both systems)]
- Baseline Stack
- Playwright [FILL: version], CDP over WebSocket, headless
- Test Fixture
- [FILL: URL(s) or local fixture; DOM node count]
- Date Measured
- [FILL: Measurement date]
Measurement Controls
- Each benchmark is run [FILL: N] times following [FILL: warm-up count] warm-up cycles to allow JIT stabilization and cache warming.
- Reported figures represent the [FILL: median / p95] latency distribution across valid runs.
- Timing triggers begin at [FILL: exact start event] and conclude upon [FILL: exact end event], measured via [FILL: performance.now() / internal profiler].
- Both test subjects are evaluated sequentially on identical hardware, state, and runtime session back-to-back.
- Explicit Exclusions: DNS resolution, initial TCP handshake, full document DOM parsing, and remote LLM inference are strictly excluded to isolate pure browser perception and extraction latency.
Reproducibility
The terminal comparison referenced across the site is recorded directly on the benchmark environment described above. A standardized benchmark harness is accessible to enterprise evaluators upon request.
Request Benchmark Harness →What These Numbers Are Not
These metrics measure the perception and extraction layer specifically. They are not end-to-end task completion times (which depend on agent planning, LLM tokens, and target site response delays), nor do they represent non-standard browser configurations.