[rwkeyes]

Problem and result

## Why it is fast answered a different question than its heading asks. It explained the engine's design and then
spent a six-row table on what upstream's past releases gained (0.1.13, 0.2.0, 0.2.4, 0.2.7) — a reader comparing
this fork against the current release learns nothing from those rows. The actual comparison lived 180 lines further
down, under What this fork changes.

After: the section states the fork's gain against stock v0.2.13 directly, with the chart and the table, right
after the four engine decisions that explain why the engine is fast at all. The past-release table is replaced by a
link to the history table in docs/DETAILS.md.

Related discussion

none — documentation only.

Approach

1. Design first, then the gain. Four bullets on how the engine fits a 24,576-expert model on one PC, one line of
   absolute speed for this card (1,404 tok/s prompt at 4K, 80.8 tok/s decode), then the comparison as a subsection.
2. ### Against stock v0.2.13, on the same card carries the claim and its evidence: the short-version paragraph,
   the chart, the TTFT table with absolute and reciprocal columns, one sentence for work finished in the window, the
   contrast between the two servers at a fresh prompt, the 24-token extraction and the green suite, three sentences
   on what the numbers do not claim, and the Measured-on block.
3. One copy of the claim. What this fork changes keeps the list of levers and a three-line pointer back up,
   instead of a second, longer copy of the same tables.
4. No stale history. "How each step of the engine's speed was won is the history table in docs/DETAILS.md; it
   is not repeated here."

Validation

Passed:

• Numbers unchanged and re-checked against dirigo-results/0.2.13-port/: 739/2,622/6,688/9,641 ms against
  949/3,163/7,500/9,659; absolute gaps 210/541/812/18 ms; reciprocals 28.4/20.6/12.1 %; window work 22/22/22/23
  against 21/20/19/21 and 62.6/62.6/62.6/65.4 against 59.7/56.9/54.0/59.7.
• Anchors and links: every #anchor resolves against its heading (including the new
  #against-stock-v0213-on-the-same-card); no broken relative links. The chart is referenced exactly once.
• Rendering: GitHub's Markdown API gives 7 tables, 2 images and both h3 subsections; no ragged table rows.

Failed:

none observed.

Not run:

• No new measurement, no build. The data is the 2026-11-09 run.

Limits

• One pass per arm; the page says so and scopes the claim to unique-prompt traffic.