Docs › site-editorial-plan

Editorial plan — from record to authority

The operator, 2026-08-11: the site should become “less unformed data dump, more accessible authority with detailed metrics, thoughtful analysis and well-written highlights” — and should let them see what’s going on, because the volume of testing has outrun anyone’s ability to track it in their head. The second point orders the work: orientation first.

The four layers

0. Now — the orientation page, and Phase 1’s deliverable. What is running at this moment, what landed in the last day/week, what is queued, and what decisions are waiting on a human. Generated from the records (run-meta, claims by date, candidate gate changes, git log) so it cannot drift from reality. This is the page the operator opens with coffee.

1. Front — the landing page. Three to five current highlights in the house style (below), a one-paragraph “start here” stating the site’s premise: home-lab measurements with wall-metered energy, visible retractions, and provenance on every number — the things published benchmarks don’t carry.

2. Reference — per-model and per-lever pages, the authority core. Each model page: verdict paragraph (written, not generated), throughput-vs-depth chart, Wh-per-correct, guard status, gate history, provenance badges. Charts are build-time static SVG from run records; no client JS. The design doc’s §3 topology/node-page vision folds in here.

3. Archive — claims, runs, configs, energy as they exist: the citable substrate, one click beneath every number above.

House style for a highlight

Four sentences, strictly: (1) the finding, plainly; (2-3) why it matters here; (4) the caveat that keeps it honest. Every number links to its claim. No first person in the finding; process history lives in methodology-lessons, not in highlights.

Standing rules

  • A statement without a claim link does not ship.
  • Confidence and provenance render as badges, not prose hedges.
  • Superseded/retracted content stays reachable but never load-bearing.
  • The Now page regenerates on every build; reference pages regenerate their DATA but their verdict prose is authored and reviewed.
  • Samples before mass production: one model page + one essay reviewed by the operator before any bulk generation.

Phases

  1. Now page + this plan (immediately useful to the operator).
  2. Samples: 122B reference page + “The KV quantisation saga” essay → operator review.
  3. Style locked → generate remaining model/lever pages; landing page rewrite.
  4. Charts pass; essays 2-3 (“How we un-fooled ourselves”, “Vulkan vs ROCm, resolved”).