[rwkeyes]

Problem and result

## Why it is fast described the engine. The page's job in that slot is to make the fork's case against the release it
was ported to, so the section is renamed "Why this fork is faster than stock Strata" and rebuilt as an argument:
what is not faster, what is, by how much, why, the control that isolates the cause, how to reproduce it, and what it
does not show.

Related discussion

none — documentation only.

Approach

1. Lead with the negative result. Both trees build engine 0.1.59; the eight-leg engine battery is within 1.1 %. The
   reader learns the claim rests on the server before seeing any gain, which is what makes the rest credible.
2. Then the gain, with the chart, the absolute plus reciprocal table, and the work finished in the fixed window.
3. Then the cause, as a contrast between what each server does at a fresh prompt.
4. Then a control: on one 0.2.12 build, conv_route: shadow (classify and log, never switch) lands on stock's
   number, 934 ms at one stream, and conv_route: on lands on 727 ms. Same build, same card, same workload — the
   switch, not the build, moves the first token. Cited to DIRIGO.md, which carries both arms.
5. Then reproduction: the two-arm script, the per-level JSON, and make-chart.py, which recomputes every figure on
   the page from those files.
6. Then the limits, then one closing paragraph for the three engine decisions both versions share, since those are
   stock's and are explained in docs/HOW_IT_WORKS.md.
7. The drafting table moved to What this fork changes, under "the same drafter, more control"; the extraction and
   suite results moved there too, as "two measured differences that are not speed".

Validation

Passed:

• The control's numbers are the ones DIRIGO.md carries for the 0.2.12 pair (934 / 3 004 / 7 583 / 9 807 ms shadow
  against 727 / 2 640 / 6 114 / 9 758 ms on); the 934 vs 727 ms quoted is the one-stream pair, and the section says it
  is a different build from the 0.2.13 comparison.
• The comparison's numbers are unchanged and still trace to dirigo-results/0.2.13-port/.
• Links and anchors: the old #against-stock-v0213-on-the-same-card is gone from the page (0 occurrences); the new
  heading anchor resolves from Start here and from What this fork changes; no broken relative links.
• Rendering: GitHub's Markdown API gives 8 tables and 2 images; the lookup=0|1 cell stays a single cell.

Failed:

none observed.

Not run:

• No new measurement: the control is the 0.2.12 data already in the repo, the comparison is the 0.2.13 run.

Limits

• The control comes from a different build and a different day than the headline comparison; it is labelled as such
  and is there to show the mechanism, not to add to the percentages.
• One pass per arm; the drift bound is stated next to the numbers.