refactor: consolidate benchmark presentation into benchmarks/presentation/; widen charts, drop eyebrows

Author: noeltock

.gitignore

@@ -31,7 +31,8 @@ build/
 .clawpatch/
 
 # generated benchmark reports
-report/
+benchmarks/presentation/review.html
+benchmarks/presentation/scoreboard.html
 .playwright-mcp/
 
 # tuner: cached propose artifacts + captured regressions (smoke.seed.json + baselines/ stay committed)

README.md

@@ -235,7 +235,7 @@ output (Impeccable, Codex, Claude, and more) converts to native blocks, across s
 converters (the built-in rules, plus experimental LLM translators run via their CLIs).
 
 ```sh
-npm run bench          # score the suite; write report/review.html + report/scoreboard.html
+npm run bench          # score the suite; write benchmarks/presentation/review.html + benchmarks/presentation/scoreboard.html
 npm run bench:record   # also append a provenance-tagged run to benchmarks/results.jsonl
 ```
 

benchmarks/README.md

@@ -12,11 +12,11 @@ npm run bench
 ```
 
 This prints a scorecard (overall + per-producer) and writes two generated pages (gitignored,
-segregated from the committed suite) under `report/`:
+segregated from the committed suite) under `benchmarks/presentation/`:
 
-- **`report/review.html`** — per layout: the ideal end state beside *each producer's* render
+- **`benchmarks/presentation/review.html`** — per layout: the ideal end state beside *each producer's* render
   (the cross-producer comparison matrix).
-- **`report/scoreboard.html`** — scores over time, built from the results log.
+- **`benchmarks/presentation/scoreboard.html`** — scores over time, built from the results log.
 
 ```sh
 npm run bench:record   # same run, but appends one record to benchmarks/results.jsonl
@@ -250,7 +250,7 @@ Three additive extension points — none requires touching the core:
 2. **Engine** (a converter) — `scripts/engines/<name>.ts` exporting `convert()`.
 3. **Output target** (a block vocabulary, e.g. CoBlocks) — an adapter; directional, see `md/04` + `md/09`.
 
-Run `npm run bench` (prints the scorecard + writes `report/`); `npm run bench:record`
+Run `npm run bench` (prints the scorecard + writes `benchmarks/presentation/`); `npm run bench:record`
 appends a provenance-tagged run to `results.jsonl`. The golden rule (`md/11`): **be
 specific about the output (core blocks), neutral about the variable axes — never
 special-case one producer or one engine/model.** The `suiteHash` tells you when two runs

benchmarks/visuals/README.md benchmarks/presentation/README.mdbenchmarks/visuals/README.md renamed to benchmarks/presentation/README.md

@@ -1,16 +1,16 @@
-# Benchmark visuals
+# Benchmark presentation
 
 Public, screenshot-ready visualisations for the README and other collateral. One
-self-contained `index.html` + one lightweight `style.css` (Geist, no build, no
+self-contained `figures.html` + one lightweight `style.css` (Geist, no build, no
 dependencies).
 
 ## View
 
-Open `index.html` directly, or serve the folder:
+Open `figures.html` directly, or serve the folder:
 
 ```sh
-cd visuals && python3 -m http.server 8799
-# http://localhost:8799/index.html
+cd benchmarks/presentation && python3 -m http.server 8799
+# http://localhost:8799/figures.html
 ```
 
 Then screenshot each figure card (or the whole page) and drop the PNGs into the
@@ -25,10 +25,13 @@ main `README.md`.
 
 Figures 1–2 are data-driven; 3–4 are static diagrams.
 
+The generated `review.html` + `scoreboard.html` now live in this same folder
+(gitignored, regenerated by `npm run bench`).
+
 ## Refreshing the data
 
 The numbers are embedded as `MODELS` and `RUNS` constants in a `<script>` at the
-bottom of `index.html`, transcribed from `benchmarks/results.jsonl`. When new runs
+bottom of `figures.html`, transcribed from `benchmarks/results.jsonl`. When new runs
 land, update those two constants and re-screenshot — the charts scale to any number
 of runs automatically.
 

benchmarks/visuals/index.html benchmarks/presentation/figures.htmlbenchmarks/visuals/index.html renamed to benchmarks/presentation/figures.html

@@ -10,7 +10,6 @@
 <div class="wrap">
 
   <header class="page-head">
-    <p class="eyebrow">Block Runner · Conversion benchmark</p>
     <h1>How faithfully does HTML become native blocks?</h1>
     <p class="lede">Every conversion is scored against a fixed suite of design sections with a known ideal block
       tree. Higher means the output matched the intended <code style="font-family:var(--font-mono)">wp:*</code>
@@ -27,7 +26,6 @@ <h1>How faithfully does HTML become native blocks?</h1>
   <!-- HEADLINE: value by complexity -->
   <section class="figure">
     <div class="figure-head">
-      <p class="eyebrow">The result</p>
       <h2>Raw LLM vs Block Runner, by block complexity</h2>
       <p class="sub">The same conversion task scored five ways — two raw LLMs writing the markup themselves (no Block
         Runner), and three Block Runner paths: rules only, then a model (GPT-5.5 / Opus) proposing the structure
@@ -47,7 +45,6 @@ <h2>Raw LLM vs Block Runner, by block complexity</h2>
   <!-- DETAIL: per-block dots -->
   <section class="figure">
     <div class="figure-head">
-      <p class="eyebrow">The result · per block</p>
       <h2>Every block tested, as one dot</h2>
       <p class="sub">The same five paths and the same Simple / Complex split — but each dot is one block’s score,
         so you can see the spread the averages hide. The bar above each path is its mean.</p>
@@ -61,7 +58,6 @@ <h2>Every block tested, as one dot</h2>
   <!-- FIGURE E -->
   <section class="figure">
     <div class="figure-head">
-      <p class="eyebrow">How it works</p>
       <h2>The path from any HTML to clean native blocks</h2>
       <p class="sub">What runs when an end user pipes HTML through the CLI. One pass: translate the structure,
         prove it against real Gutenberg, fill in real media — and surface anything it couldn’t express natively.</p>
@@ -75,7 +71,6 @@ <h2>The path from any HTML to clean native blocks</h2>
   <!-- FIGURE F -->
   <section class="figure">
     <div class="figure-head">
-      <p class="eyebrow">How it’s measured</p>
       <h2>The benchmark loop</h2>
       <p class="sub">One ideal answer per layout; many producers supply HTML; the engine converts; a deterministic
         scorer grades the result against the ideal tree; every run is logged with provenance so progress is auditable.</p>

benchmarks/visuals/fonts/Geist.woff2 benchmarks/presentation/fonts/Geist.woff2benchmarks/visuals/fonts/Geist.woff2 renamed to benchmarks/presentation/fonts/Geist.woff2

@@ -60,20 +60,12 @@ body {
 }
 
 .wrap {
-  max-width: 980px;
+  max-width: 1280px;
   margin: 0 auto;
-  padding: 56px 28px 80px;
+  padding: 56px 44px 80px;
 }
 
 /* ---- header ---- */
-.eyebrow {
-  font-family: var(--font-mono);
-  font-size: 11.5px;
-  letter-spacing: 0.14em;
-  text-transform: uppercase;
-  color: var(--faint);
-  margin: 0 0 12px;
-}
 .page-head h1 {
   font-size: 34px;
   line-height: 1.1;
@@ -106,8 +98,8 @@ body {
   border: 1px solid var(--line);
   border-radius: 16px;
   box-shadow: var(--shadow);
-  padding: 26px 28px 22px;
-  margin: 28px 0;
+  padding: 40px 44px 34px;
+  margin: 40px 0;
 }
 .figure-head { margin-bottom: 8px; }
 .figure-head h2 {