[rwkeyes]

Problem and result

The MTP bullet read as if guess-then-check were the fork's advantage, and the paragraph that was meant to correct that
named only two of the fork's three drafting settings. Checking the two servers line by line also turned up one item I
had implied was ours and isn't.

• Both arms of the 0.2.13 comparison started with --spec 6 --spec-min-p 0.5 and sent no per-request override, so
  drafting contributes nothing to the measured gain. That is now on the page, with the settings.
• The real difference is control over the drafter, not a different drafter, and it is now a four-row table.
• spec_min_p per request is stock behaviour: upstream v0.2.13's server already forwards
  strata_tune{pcie_frac, spec_min_p, prefill_pipe, aux_cpus}. The fork's additions are spec (depth, clamped to
  2…startup), lookup (prompt lookup on/off), the spec_window policy, and the TUNE / POST /v1/tune channel that
  changes settings globally without a restart.

Related discussion

none — documentation only.

Approach

Replaced the paragraph under the three engine decisions with: the shared-behaviour statement including the exact
drafting settings both arms used, a four-row table of stock vs fork around drafting, and one line saying none of it was
switched on for the numbers below (spec_window 0, GuideLLM sends no strata_tune). The glossary's MTP entry now
carries the same three additions and the note that spec_min_p per request is on both sides.

Validation

Passed:

• Both servers read, not assumed. grep -n 'strata_tune' serve/server.py in the two 0.2.13 trees: the stock tree
  forwards pcie_frac, spec_min_p, prefill_pipe, aux_cpus; the fork adds spec (with the comment "clamps it to
  2..--spec … a deeper window than the engine was started with is refused there") and lookup. spec_window exists
  only in the fork (set_spec_window, 0 = off).
• The measured settings come from the two config files the arms ran: identical --spec 6 --spec-min-p 0.5 --mtp … --mtp-window 8192.
• Rendering: the new table renders as 3 columns × 5 rows, and the lookup=0|1 cell stays one cell (escaped pipe).
  No broken relative links; all anchors resolve.

Failed:

none observed.

Not run:

• No measurement of the tuning controls themselves; the page still makes no speed claim about them.

Limits

• The table describes what each server accepts. It does not claim a measured benefit for any of the fork's four
  additions, and says that spec_window was off for the comparison.