[rwkeyes]

Problem and result

Two statements in the front page implied the fork invented things that both versions share.

• MTP. The four-engine-decisions list sits under a heading that now carries the fork-vs-stock comparison, so
  "guess, then check" read as the fork's advantage. Both sides run the same engine version (0.1.59) and the same
  draft layer; the fork's only drafting additions are two settings the measurement does not use.
• The 28.8 GB table. It is not an engine feature at all — it is the model publisher's second GGUF file. Nothing on
  the page said so, which invites the reader to assume we created or modified it.

Related discussion

none — documentation only.

Approach

1. The section now opens with "Stock Strata v0.2.13 and this fork run the same engine version (0.1.59)", so the
   four bullets read as a description of the engine both versions share rather than of the fork.
2. After the bullets, a paragraph says it outright: the expert split, the table and guess-then-check are stock
   behaviour, the carry changes how none of them runs — and names the one PLE file the carry does touch
   (ple_parity, the table's self-test, changed to skip when the fixture is absent, not to compute differently).
3. It then lists what the fork actually adds around drafting — per-request spec=N and spec_window, with
   spec_window: 0 in the measured config and GuideLLM never sending spec — and points at the server-side
   difference that the comparison measures.
4. The PLE bullet and the glossary now state the table is the model's: publisher's second file, 160 × 320,001,536 at
   IQ4_NL, 28,800,138,240 bytes, read by stock and fork alike, not created or re-encoded by either.

Validation

Passed:

• The shared-code claim is checked, not asserted: git diff --name-only v0.2.13 3512b4cce511 contains no runtime
  PLE or draft-layer change; the PLE paths in it are src/kernels/ple_parity.cpp and its SYCL twin, and the diff is
  the fixture-missing SKIP (exit 77) plus a reworded string.
• The table claim is checked: both 0.2.13 arms' configs point --ple-gguf at the same publisher file, and the
  tensor geometry in that file closes to the quoted byte count.
• Links, anchors and rendering unchanged: no broken relative link, all #anchors resolve, GitHub's Markdown API
  parses 7 tables and 2 images.

Failed:

none observed.

Not run:

• No measurement or build; text only.

Limits

• The drafting additions are named but not measured — spec_window was off and spec was never sent, so the page
  still makes no speed claim about them.