Static Cloudflare Worker serving the model-benchmark dashboards at bench.unsigned.gg. Two pages, one growing pile of numbers:
/(public/index.html) — the local bank: decode speed + MTP speculative-decoding gains, measured first-hand on an RTX 5090./frontier— the frontier leaderboard: models we can't run at home, each score tagged with its provenance (independent / corroborated / self-reported / disputed).
Migrated out of the agentic monorepo (packages/local-models/bench) into its
own repo so it ships on its own CI/CD cadence.
Numbers live as JSON so adding a benchmark is a data edit, not HTML surgery.
They sit under public/ because only the assets root is served; the pages
fetch() them at load.
| File | What it holds | Schema |
|---|---|---|
public/data/frontier.json |
vendor frontier chart | { snapshot, source, provenance, model_sources, higher_is_better, models[], benchmarks[] } — each benchmark { name, scores[] }, scores aligned to models order |
public/data/swe-pro.json |
independent SWE-bench Pro board | { benchmark, source, url, verification, rows[] } — each row { model, score, flag, note } |
public/data/candidates.json |
models awaiting a run | { note, candidates[] } |
Winners, runners-up, and provenance badges are computed in the page — never hand-marked. Add a row, and the tables, win tally, and stat tiles recompute.
Humor is allowed; lying is not. Snark about a score all you want — the score itself stays honest, and unverified provenance gets labeled (see the Sakana / Fugu disclaimer), never laundered into fact.
npm install
npm run dev # local preview at http://localhost:8787
npm run check # wrangler dry-run build (no deploy, no auth)Every push to main runs .github/workflows/deploy.yml: it validates the JSON
data, dry-run-builds the Worker, scans for secrets (pinned gitleaks), then
deploys to bench.unsigned.gg via
cloudflare/wrangler-action.
One-time setup — add the deploy token (the deploy step self-skips with a warning until this exists, so CI stays green in the meantime):
- Cloudflare dashboard → My Profile → API Tokens → Create Token, template
"Edit Cloudflare Workers" (scopes:
Account · Workers Scripts · Edit, plus the zone forunsigned.gg). - Add it to the repo: Settings → Secrets and variables → Actions → New
repository secret, name
CLOUDFLARE_API_TOKEN.
The account ID is committed in wrangler.jsonc (not a secret). Wrangler and the
custom-domain route are pinned; bump deliberately.
Manual deploy (from a machine with wrangler login):
npm run deploybench.cerebral.work has its own source tree (cerebral-src/), generated artifact (cerebral/) and Worker config (wrangler.cerebral.jsonc). It presents Scorcher measurements, Cortex reference coverage and Saga simulation/fixture evidence in the Cerebral V2 visual system.
npm run test:cerebral
npm run build:cerebral
npm run check:cerebral
npm run dev:cerebralSee the implementation and evidence guide. The Cerebral PR workflow checks and packages the site; production deployment is a separate operator-approved step.
The Cerebral site includes /visualizations/: 174 entries from the pinned Observable D3 gallery, adapted to Three.js with Rust/WebAssembly data kernels. Local demonstration values are labeled separately from benchmark measurements; every entry links to its original author. The featured-libraries collection is linked alongside the attribution guide.
The Cerebral build now needs Node 22+ and a Rust toolchain with wasm32-unknown-unknown installed:
rustup target add wasm32-unknown-unknown
npm ci --ignore-scripts
npm run build:cerebral
npm run test:gallery
npx playwright install chromium
npm run test:gallery:browserCHROME_PATH=/usr/bin/google-chrome can select an existing browser for the last command. Native kernel tests: cargo test --manifest-path visualization-wasm/Cargo.toml -- --test-threads=1.
See the gallery implementation guide and validation evidence. No remote Observable iframe or notebook runtime is required.