Leaderboard

nurb leaderboard

Two generated surfaces, one editorial layer, two repos. The benchmark (tasks, scorer, submissions, REPORT.md) lives in Shpigford/nurb-benchmarks; the public page (site/benchmarks.html, behind nurb.dev/benchmarks) lives here. Both surfaces regenerate mechanically from that repo's submissions/; the verdict sentences and subscription labels live in its src/nurb_evals/site.py and are written by a person. Submission PRs are pure additions and merge freely; nothing reaches the public page until this skill runs. That gap is deliberate: it is where the sanity check and the verdicts happen, so publishing is an editorial act, not a side effect of merging.

Step 0: Get the benchmarks checkout

Clone or update Shpigford/nurb-benchmarks somewhere outside this repo (for example ../nurb-benchmarks, or a temp directory), on its main. Everything below that touches submissions, verdicts, or REPORT.md happens in that checkout; only the final site/benchmarks.html lands here.

Step 1: Find what is new

The runs added since the page was last regenerated. In this repo, git log -1 --format=%ci -- site/benchmarks.html dates the last publish; in the benchmarks checkout, list the submission directories whose merge commits landed after it (git log --since=<date> --diff-filter=A --name-only -- submissions/ | grep results.jsonl).

Each directory under submissions/ is one run: <harness>-<model>-<effort>-<hex>/ holding results.jsonl plus per-trial gzipped transcripts and part sources. If nothing is new, say so and stop.

Step 2: Sanity-check every new run

Work from the benchmarks checkout (its own uv project; uv sync --locked first if the venv is stale). For each new run directory:

• Rows parse and carry full identity: every line of results.jsonl has harness, harness_version, model, effort, seed, nurb_version, benchmark_version, a 12-char benchmark_revision, and timeout_s. A benchmark_revision that matches no revision the benchmark ever shipped is disqualifying (revisions from before the August 2026 repo split shipped from Shpigford/nurb's history).
• Artifacts are complete and sanitized: a transcript.txt.gz and the part source for every row; no /Users/, /home/, or usernames anywhere once decompressed (the suite's sanitization test enforces this too).
• The parts are authored, not planted: hash the submitted part files against tests/solutions/ and against parts from other submissions. An exact match with a reference solution is disqualifying; matches across unrelated submissions are worth reading.
• Spot-check by re-grading: for at least one row per new run (and every row that looks too good), rebuild the trial project (task.materialize, drop the part in, restore the submitted measurements.toml for leg_cup) and run the grader. The committed score must reproduce exactly; grading is deterministic.
• Read one transcript per new contributor (gzcat <trial>/transcript.txt.gz): the headless preamble held (no nurb dev), the model actually iterated, and the session matches the row's timings.

A run that fails a check is removed with a PR comment saying which check and why, not silently. Suspicion is not proof: when a re-grade mismatches, check the benchmark_revision first; a row graded under an older shipped revision reproduces under that revision's scorer, not today's.

Step 3: Verdicts

Every (harness, model, effort) combo on the board should have an entry in VERDICTS in the benchmarks repo's src/nurb_evals/site.py: which subscription it runs on, and one or two sentences a person with a printer can act on. Ground every claim in the rows and transcripts (findings, timings, the honesty tasks); never speculate. A new combo without a verdict renders numbers-only, which is acceptable for a day, not a policy. Capped or censored data is named as such ("hit the session limit"), never averaged into a claim.

Step 4: Regenerate, look, publish

From the benchmarks checkout:

uv run python -m nurb_evals.report --write
uv run python -m nurb_evals.site --out <this-repo>/site/benchmarks.html

Open site/benchmarks.html in a browser and look at it before publishing: label collisions on the chart, a card wrapping badly, an empty state showing when rows exist. Screenshot, not DOM-query. Then two PRs, never straight to main: in nurb-benchmarks, the regenerated REPORT.md plus any verdict edits and removed runs; here, the regenerated site/benchmarks.html. Each PR body lists which runs were published and which were rejected with reasons. uv run pytest -q tests/test_report.py tests/test_pricing.py tests/test_contribute.py in the benchmarks checkout must be green; the full grading suite only guards scorer code, which this skill never touches.

The page deploys with site/ however the site deploys; this skill's job ends at the merged PRs.