Skip to content

refactor(site): condense homepage to hero, scoreboard, evaluation - #4

Open
hiyouga wants to merge 1 commit into
mainfrom
claude/condense-results-page
Open

refactor(site): condense homepage to hero, scoreboard, evaluation#4
hiyouga wants to merge 1 commit into
mainfrom
claude/condense-results-page

Conversation

@hiyouga

@hiyouga hiyouga commented Aug 10, 2026

Copy link
Copy Markdown
Contributor

Page was doing three jobs at once (homepage / experiment report / usage docs), which pushed the result below the fold. Cut to three sections.

Hero — one-line thesis, two actions, three numbers. Dropped the 4 hard-gate layers tile (it restated the Evaluation section).

Results — now the body of the page. The five-row ranking card and the eight-row table said the same thing twice, so they are merged into one ranked table: rank column + inline accuracy bar, sorted by accuracy. The "best current default" callout becomes a single summary line above the table.

Evaluation — three-step contract, then the repository link.

Removed from the homepage: four capability cards, 15-case roster, quick-start terminal block. Staging/scoring commands live in README.md and EVALUATION.md, which stay authoritative.

Disclosure

The evaluator note keeps the BankerToolBench regrade and now also carries two comparability limits that only README.md had:

  • Penguin settings ran with Gemini visual judging enabled; Claude Code and Codex did not — and two cases (BankerToolBench, DV-World) need a visual gate for official PASS.
  • Each harness ran the model it is normally paired with, so Cost is not a same-model comparison.

Every configuration string now names its model, so the $0.1995 and $34.27 rows can be read against each other without leaving the table.

Also

Removed CSS for the deleted blocks (~65 fewer lines), gave the horizontally scrollable table a visible focus ring, added tabular-nums to numeric columns.

Numbers are unchanged from main — Claude Code stays at the regraded 10/15.

Rendered light and dark headlessly; both check out.

🤖 Generated with Claude Code

https://claude.ai/code/session_018tqUGuivaUjXD6c7mhYnqY

The page was carrying three jobs at once — project homepage, experiment
report, and usage docs — which pushed the benchmark result, the thing
readers come for, below the fold.

Cut to three sections:

- Hero: one-line thesis, two actions, three numbers (the "4 hard-gate
  layers" tile went away; it restated the evaluation section).
- Results: now the body of the page. The separate five-row ranking card
  and the eight-row table said the same thing twice, so they are merged
  into one ranked table — rank column plus an inline accuracy bar — and
  the "best current default" callout becomes a single summary line above
  it. Rows are sorted by accuracy.
- Evaluation: the three-step contract, then the repository link.

Dropped from the homepage: the four capability cards, the 15-case roster,
and the quick-start terminal block. Staging and scoring commands are
already in README.md and EVALUATION.md, which stay authoritative.

The evaluator note keeps the BankerToolBench regrade and now also records
the two comparability limits that only README.md carried: Penguin settings
ran with Gemini visual judging enabled while Claude Code and Codex did
not, and each harness ran the model it is normally paired with, so the
cost column is not a same-model comparison.

Every configuration string now names its model, so a $0.1995 row and a
$34.27 row can be read against each other without leaving the table.

Also removes the CSS for the deleted blocks, gives the horizontally
scrollable table a visible focus ring, and adds tabular figures to the
numeric columns.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018tqUGuivaUjXD6c7mhYnqY
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant