# Learning index performance: measurement method

Overskill publishes this first-party case study about its own Rails platform's
signed-out `/learn` page. It is not a benchmark of a generated app.

## Recorded experiment

- Date: September 8, 2026, 23:23:58–23:27:49 UTC.
- Browser: HeadlessChrome 152.0.0.0; viewport 390 × 844; device-pixel ratio 1.
- Emulation: Slow 4G network; 4× CPU throttling on one desktop host.
- Six loads: baseline runs 1–3, then candidate runs 1–3. Each used a fresh
  browser context. The run order was sequential, not randomized or interleaved.
- Both variants used the same warmed, guarded local Rails test-runtime
  configuration, configured Terser JavaScript compression and Rack gzip. The
  runtime restarted between variants. The production-only tracking pixel was absent.
- Content and required controls were retained; the selected JavaScript and
  stylesheet changed. Signed-in visitors retained the full application assets.

## CSV fields and calculations

[Download the six-row CSV](learning-performance-observations-20260908.csv).

- `variant`: `baseline` means full application assets; `candidate` means focused
  learning assets. These are historical versions of the same page.
- `run`: order within that variant, starting at 1.
- `js_css_encoded_body_bytes`: sum of the two selected resources' browser
  ResourceTiming `encodedBodySize` values. These gzip HTTP response bodies
  total 1,888,949 bytes for baseline and 122,104 for candidate in every run.
  This excludes response headers and all other resources. Their separately
  recorded `transferSize` totals were 1,889,549 and 122,704 bytes respectively;
  neither measure is the whole page. Standalone gzip files can differ.
- `lcp_ms`: trace-derived navigation-to-LCP duration, rounded to milliseconds.
  The LCP element was the same `P.learn-lead` text paragraph in all six traces.
- `cls_rounded_2dp`: recorded main-frame layout-shift sum rounded to two decimal
  places. Every run was 0.03; the unrounded sum was 0.0253585166, not zero.

Body-byte reduction: `(1,888,949 - 122,104) / 1,888,949 × 100 = 93.5358763%`,
shown as 93.5%. Median LCP is the middle observation after sorting each group
of three: 5,428 ms and 2,852 ms. The charts show every LCP point and start
their numeric scales at zero. MB means decimal megabytes (1,000,000 bytes).

All six rows were reconciled with the saved resource receipts and raw trace
navigation/LCP events. The public extract omits local filesystem, profile,
request and session details; raw traces remain internal. No new measurement
was made to prepare this article.

## Separate behavior review

The 24 style/geometry comparisons covered the index and five guides at 390 and
1440 px wide, 1000 px high, in light and dark themes. They swapped stylesheets
on the same DOM with animations frozen. Checked properties showed no
differences, no horizontal overflow and no unloaded images. These are not 24
performance runs or exhaustive visual/accessibility proof. Separate browser
checks covered menu, theme persistence, brief/copy, guide contents, and
navigation across the full-asset boundary.

## Limits and reuse

Fresh browser contexts and emulation are not physical-phone field data.
The sample is small and sequential. Production artifact hashes and byte lengths
differed from the local artifacts. The initial September 9 rollout covered the
index and five guides; later articles were not part of this experiment.

No field Core Web Vitals, INP, generated-app, ranking, AI citation, conversion
or registration improvement is established. Reuse the method and record your
own result; do not adopt this percentage as a target or guarantee. The original
local harness used a now-removed worktree and is not a public runnable recipe.

[Read the case study](/learn/lighter-learning-pages).