Skip to content

Energy and traffic enrichment experiment

The audit has separate adapters for GreenFrame resource-model estimates, co2-runner CPU-package counters, and a CrUX lookup. These answer different questions and must retain their measurement boundaries. Sitespeed.io's hardware-energy path has been inspected but has not been executed. See methodology and sources.

Fresh GreenFrame estimates

The fresh report supplement captured six pages on November 4, 2026, from 22:38:59 to 22:41:40 UTC. All 18 repetitions completed successfully. Each row summarizes three independent browser scenarios; the range is observed variation, not a confidence interval.

Page Client-device median Client-device range Network median
Guide homepage 12.00 mWh 10.38–13.85 mWh 0.76 mWh
Guide: manage images 11.15 mWh 10.52–14.50 mWh 2.80 mWh
Guide: new in Wagtail 8.1 10.72 mWh 10.67–11.04 mWh 4.19 mWh
Wagtail.org homepage 12.08 mWh 11.23–12.58 mWh 11.65 mWh
Wagtail.org: Wagtail 8.1 blog 16.41 mWh 15.48–18.00 mWh 25.26 mWh
Wagtail.org: sustainability 13.84 mWh 10.66–14.96 mWh 7.51 mWh

The guide manifest and Wagtail.org manifest retain exact Wh values, individual repetitions, component estimates, timestamps and artifact digests. Their adjacent page directories publish raw Docker counters, full-precision model output and the executed scenario/bridge. Each bundle includes plan.json and reproduction.json; sibling energy.json files are exact manifest copies for report rendering. These later captures supplement the scored reports without changing their original evidence or judgments.

The boundary is modeled browser-device CPU, memory, disk and screen energy. Network energy is modeled separately; remote servers are excluded. The unchanged GreenFrame 2.0.1 model uses a synthetic 45 W CPU and 14 W screen, with CPU usage normalized by available host CPUs. The screen estimate dominates these client-device results. This is neither measured electricity nor total-device energy, and it must not be added to SWDM or extrapolated directly to annual site impact. The results establish a reproducible sample, not a causal efficiency ranking.

The shared Linux workstation had an Intel Core i5-1145G7 and eight logical CPUs. Docker 29.9.2 ran the pinned mcr.microsoft.com/playwright:v1.30.0-focal image, digest sha256:9e6eddb08b8566076ef778007859096bc90f8984236061c086eaede2fb60430b, with Chromium 110.0.5481.38. GreenFrame supplies a fixed Chrome 89 user-agent string and --disable-web-security; these are experimental limits, not a current-browser compatibility claim. Temperature, power state and ambient activity were uncontrolled; local tests and campaigns were paused during collection.

Each fresh browser context used 1,280 × 800, navigated through load, then waited three seconds without consent interaction or scrolling. GreenFrame starts its accounting window after browser startup and a two-second warm-up, and ends after browser close. Modeled counter spans were approximately 3–5 seconds. Repetitions shared a container, so this does not establish cold host caches. The narrow energy sample covers three pages per site, fewer than the browser audit sample.

Earlier CPU-package measurements

Target homepage Valid repetitions Median CPU-package energy Observed range
Guide 3 71.91 J 68.76–72.35 J
Wagtail.org 3 77.86 J 76.65–78.17 J

The energy artifact records each run, counter sample count and span, session duration, source digests, and first-sample sensitivity. These are sequential, round-robin, headed Firefox sessions at 1,280 × 720: navigate through load, wait three seconds, close. No consent interaction or scroll is included. The sessions lasted about 5.7–6.2 seconds including startup/shutdown. The Chromium transfer protocol uses 1,280 × 800 and different instrumentation; the figures are not interchangeable phase measurements.

Linux RAPL permissions and the CPU-package event were already available. No system settings were changed. The shared workstation's temperature, power state, and unrelated activity were not experimentally controlled; no audit tests or model campaigns were deliberately run alongside these measurements. There is no idle subtraction, matched performance baseline, or causal site comparison.

The experiment used co2-runner revision cf665f5 with uncommitted changes, including the CPU-package selection fix. The linked revision identifies the base source; the recorded file hashes identify the modified implementation. The adapter records file hashes in addition to the commit. It leaves the checkout untouched and isolates profiles/HARs under ignored .audit-work/energy/. The published artifact excludes those raw files.

The validator accepts one top-level CPU-package series and rejects missing, empty, malformed, nonfinite, negative, or unordered samples. It does not sum overlapping domains. A valid all-zero series is flagged and excluded from the summary. The first sample is included to match co2-runner; excluding it is recorded separately as a sensitivity calculation.

Executed CrUX lookup

The pinned mirror's September 2026 dataset matched both exact HTTPS origins:

Origin CrUX rank bucket Annual pageview scenario
https://guide.wagtail.org 10,000,000 approximately 245,000
https://wagtail.org 5,000,000 approximately 573,000

The traffic artifact contains exact mapping values, dataset month, mirror commit, compressed-file digests, searched buckets, and mapping digest. These are scenarios based on a different population and period, not observed site traffic. The initial experiment stopped at pageview scenarios; the annual footprint extension now combines them with an explicitly assumed sampled-page mix. The guide's scenario must not be replaced by Wagtail.org's bucket or combined as if origins were interchangeable.

Reproduce

# Uses the supplied checkout and its installed Node/Firefox dependencies.
uv run --locked --script audit/energy.py --output .audit-work/new-cpu-package-energy.json
# Reads public monthly dumps; no CrUX or BigQuery key required.
uv run --locked --script audit/traffic.py --month 202609

The Python entry points have executable shebangs, inline requirements, and script lockfiles. audit/rank-pageviews.csv freezes the supplied mapping. New measurements have different timestamps and may differ numerically. Keep Harbor/model work separate from hardware-energy collection.

For the fresh GreenFrame protocol, install CLI 2.0.1 into an isolated local directory and pull the pinned runner image. The captured CLI archive SHA-256 is fbfbbfd6ba034b2dd7ed35f6a412f32873f2ab0d3d9dbe06a5b590ab0fc1effc; the stable download URL can change, so verify both version and digest. Use the published plans and new output directories:

docker pull mcr.microsoft.com/playwright:v1.30.0-focal
uv run --locked --script audit/greenframe.py \
  --plan docs/reports/assets/report-refinement/guide/energy/plan.json \
  --output-dir .audit-work/new-guide-energy \
  --mode run --binary .audit-work/tools/greenframe/bin/greenframe
uv run --locked --script audit/greenframe.py \
  --plan docs/reports/assets/report-refinement/wagtail/energy/plan.json \
  --output-dir .audit-work/new-wagtail-energy \
  --mode run --binary .audit-work/tools/greenframe/bin/greenframe

New agent collections record energy readiness by default. workflow collect --energy run --greenframe-binary <local-binary> attempts the accepted sample after browser collection; --energy skip records an explicit opt-out. Missing prerequisites or invalid counters retain null metrics and a reason. No hardware permissions or global settings are changed.

Next validation

Run a matched, repeated sitespeed.io/co2-runner comparison on a controlled fixture, including first-sample handling and journey duration. Calibrate traffic scenarios against owner analytics with matching origin and period. Both are tracked in the backlog; neither uncertainty is hidden by adding a score.