These measurements describe execution of the 0.4.0rc1 working build on one Windows machine on 23 September 2026. They are observations, not latency targets, physical controller speeds or a scalability claim. The raw files retain every sample and the environment. Ordinary background processes were not isolated.
Python 3.12.14, Windows 11, AMD64. Each case has eight assets and 180 or 181 output intervals. One warm-up is excluded; 25 timed samples include simulate_continuity(...).to_dict(). Import time, JSON encoding, HTTP and rendering are excluded. P95 uses the nearest rank, not an interpolated percentile.
| Synthetic preset | Median | P95 | Maximum |
|---|---|---|---|
| Normal supply | 46.56 ms | 56.55 ms | 59.90 ms |
| Generator failure | 54.74 ms | 82.40 ms | 83.17 ms |
| Surviving path maintenance | 44.95 ms | 65.15 ms | 67.29 ms |
Raw native samples and exact input hashes.
python scripts/benchmark_continuity.py --runs 25 --preset normal --preset generator_failure --preset path_maintenanceThe ordinary generator_failure preset is measured here. The guided case disables charging and has its own input hash; do not interchange full run identities because their 100 kWh depletion numbers happen to match.
Chromium 153.0.8010.12, Windows x64, 1280 × 900, unthrottled loopback Vite preview. Three isolated contexts provide cold browser caches; the OS cache may remain warm. Stopwatch intervals include automation and assertion overhead. Bytes are decoded response bodies captured through first result, including worker requests, rather than compressed network-transfer sizes.
| Observation | Three samples or measured value |
|---|---|
| Navigation to visible prediction controls | 830 / 798 / 823 ms |
| Run click to visible guided result | 151 / 173 / 165 ms |
| First-result decoded payload | 462,718 bytes in each sample (about 452 KiB) |
| Python, cover image or video requested before first result | None |
| Generator-failure replay | 181 precomputed intervals at 120 ms per tick; 21.72 s programmed including stop tick; 21.89 s observed including automation |
Raw browser samples and per-resource bytes. The video and poster load on the separate evidence page. Python is an optional later download. Replay is an animation of an already computed result; it is neither simulation execution time nor real-time physical behavior. These local timings cannot predict mobile or internet load times.
python -m pip install -r requirements-studies.lock
python scripts/fetch_study_runtime.py
python scripts/prepare_browser_demo.py
npm --prefix apps/web run build:demo
npm --prefix apps/web run test:demo -- benchmark.spec.tsSet TWIN_PYTHON to the desired Python executable when it is not on PATH. Install the pinned frontend dependencies and Playwright browser first, as in the quickstart. The test writes a fresh guide-benchmark.json into its test-output directory; it does not replace this historical receipt.
| Environment | Verification performed by the configured workflow |
|---|---|
| Ubuntu / Python 3.12 | Core, API, installed wheel, three-case packet, native timing, loopback browser journeys |
| Ubuntu / Python 3.14 | Core, API, installed wheel, three-case packet, native timing |
| Windows / Python 3.12 | Core, API, installed wheel, three-case packet, native timing |
| Windows / Python 3.14 | Core, API, installed wheel, three-case packet, native timing |
| Chromium / Ubuntu | Browser/native equality, all 24 lesson default/challenge results, optional browser-Python verification, guided journey, reports, comparisons, responsive and accessibility checks |
Inspect the exact commit's CI runs. Each successful matrix run retains its own numerical-evidence-* packet and timing artifact for 90 days. The release receipt records the actual successful run IDs; this table describes coverage and is not an assertion that an untested commit passed. Timing differences across runners do not establish a speed ranking. SHA-256 integrity and numerical reproduction are distinct from an independent person's review.
The regression runs axe-core's WCAG 2 A/AA and 2.1 A/AA rules on the guide before and after calculation, the evidence hub, all twelve lesson themes, and advanced overview/topology/source-register states. Keyboard skip/focus, changing-result announcements, mobile fit, diagram text and reduced motion have targeted browser checks. Color alone does not carry scenario or status meaning. Automated checks are a focused engineering pass, not a complete accessibility certification or user study.