[rwkeyes]

Problem and result

The 22 % first-token figure is a single-user number, and the page let it read as a property of the fork. Someone
deciding whether to put this on a shared box would have found the caveat only in the eight-stream row and the "what it
does not show" list, after the chart.

Related discussion

none — documentation only.

Approach

1. A callout directly under the claim, above the chart: the advantage falls 22 % → 17 % → 11 % at one, two and four
   concurrent clients, is gone at eight (9,641 vs 9,659 ms), the reason (the engine serves one request at a time), and
   what to do instead (--batch/--slots, linked to docs/BATCHING.md), with 22 % described as the one-user figure.
2. The chart carries it too, in its footer: "Single-user figures: both arms ran one request at a time, so the
   advantage falls to zero at eight concurrent clients, where the two servers are queueing." The image is what gets
   screenshotted, so the qualifier travels with it. scripts/make-chart.py regenerated chart.svg; the three small
   sections were checked for clipping and overlap.
3. The eight-stream bullet in the "which row are you" list now points at the callout rather than restating it, so
   the point is made once at full strength.
4. The report adds the same scaling to its "does not show" list.

Validation

Passed:

• Percentages and milliseconds unchanged and traced to guidellm/*.json (22.1 / 17.1 / 10.8 / 0.2 %; 9,641 vs
  9,659 ms at eight).
• The reason is from the two configs (--batch 0) and docs/BATCHING.md ("By default Strata serves one request at
  a time").
• Chart: regenerated from the folder's JSON by scripts/make-chart.py; rendered at 1843×1612 and reviewed — three
  footer lines inside the canvas, no overlap with the throughput bars, and the eight-stream pair visibly equal.
• Links and rendering: docs/BATCHING.md resolves; GitHub's Markdown API renders the callout as a blockquote and
  the page keeps 8 tables / 2 images.

Failed:

none observed.

Not run:

• No multi-slot measurement; the callout sends the reader to the configuration instead of quoting a number for it.

Limits

• The scaling (22 → 17 → 11 → 0) is on this card, this model and one request at a time; with --batch 2–8 the shape
  would differ and is not claimed.