[rwkeyes]

Problem and result

The front page and the first-run path were rewritten on 2026-11-13; a day of measurement since then has made four
statements in them false, imprecise, or a trap. All four are on pages a newcomer reads before anything else.

1. DIRIGO.md listed a limit that no longer exists — "the --conv-route-batch-after threshold was not swept" and
   the switch back was unmeasured. Both were measured on 2026-11-14 (§7): the switch back costs a conversation 1.9 s
   per re-open, so 4 stands, and sweeping it revealed the config key was never read (fixed in b906f0b5).
2. README.md gave no threshold — "runs of single-shot prompts" reads as "one request flips it". It is four.
3. The same table quoted −12…−22 % without the client count — that is 1–4 clients; at 8 it is nothing.
4. Both quick starts named data/expert-profile.bin without saying which model shape it fits. The profile is
   matched on the expert count, not the quantization: 48×512 for the full model, 48×256 for the pruned Coder — and
   the wrong one stops the engine with it is a profile for a different model, which reads like a corrupted pack.

Plus one dead link a reader could actually hit: docs/HOW_IT_WORKS.md pointed at ../README.md#how-does-it-work, a
heading the README rewrite removed.

Related discussion

• serve: honour conv_route_batch_after from the config, not only the flag #16 / [Docs] DIRIGO.md: record the one code change on the branch #22 — the code change and the invariant correction this follows from
• [Docs] Mixed traffic: keep batch_after at 4 (1 costs 1.9 s per conversation) #17 — mixed traffic, where the threshold was measured

Approach

Audited the ten front-door pages for links, anchors, and every claim touching batch_after, the cache switch, TTFT
numbers, --expert-profile and PLE; fixed only what was actually wrong.

Validation

Passed:

• The 512 figure is measured, not assumed: read from the IQ3_XXS GGUF's own metadata (48 layers,
  expert_count = 512), so "IQ2_/IQ3_/Q2_0 → expert-profile.bin" is grounded; the Coder packs' headers are 48×256.
• Every relative link and anchor in the ten front-door pages resolves after the changes (checker: punctuation
  stripped, spaces → hyphens, explicit <a name> honoured).
• The README renders: 8 tables, 60 rows, no ragged rows.

Failed:

• My first "fix" broke a working link. I changed MODELS.md#a-big-graphics-card-and-little-ram to #pick-by-memory
  because my slug function collapsed the double hyphen GitHub produces when a / is removed from
  "History / older measurements". The anchor resolves; the change is reverted, and the checker now uses GitHub's rule
  (each space → one hyphen, no collapsing).

Not run:

• No change to any number or measurement; nothing re-measured, the pages now agree with the inventory instead.

Limits

• The audit is limited to the front-door pages listed in the commit; bench/results/ and apps/host/docs/ still carry
  their historical dead links by design (they are evidence, not navigation).