Changelog

Bars no longer draw the interval. Boards, harness pages and Benchmarks still print it beside each pass rate.

Boards have no ranked or practice status. Every board lists its rows by pass rate, and rank ranges show once two rows have a score, so a board with one harness shows its score and nothing about rank. /benchmarks loses "Ranked boards only".

Terminal-Bench 2.0 on Qwen2.5-Coder 7B is an ordinary board, and its page says what the board is for.

The site's text says "board" everywhere it said "arena". Old links keep working.

Results lists every published run too, one row each, with the same filters and CSV. The old list of runs at /runs leads there.

Entry pages are gone: a run's page shows the same result. An old link to an entry leads to its run's page, or to its board when the entry has more than one run there.

Boards give each harness a rank range from paired comparisons, task by task, instead of tiers. The methodology explains the method and its version.

Each board measures against a reference harness, which the board's dataset names. Every Δ on a board and on a harness page is against that harness.

A run's page names the other models its trials called, besides the board's model.

A row's Runs on a board say how many runs were issued and how many were uploaded.

A benchmark's page shows an allowlist board first when one holds a comparison.

Your bench opens a job's files with links that work for 15 minutes.

A run's owner can make its trajectories public. The site scans each trial's trajectory for secrets first, and a trial with a finding stays private.

Each reviewed dataset has an assessment against the dataset checklist.

The home page says what you can do here in plain sentences, with four slides that show a run from start to score.

Every interval is drawn as a band behind its bar.

The header and the footer link to hbench on GitHub.

The run builder knows the flags of hbench 0.3.9.

The Submit page is now "Put your result on a board".

The terms list the reasons for a take-down, and a take-down's note is public.

The public audit log lists every take-down with one of the listed reasons, and every run put back.

Each harness page draws a profile of the harness on each board, from its trajectories, with an overlay and a table of values.

The board table shows the cache ratio, the spread between runs, errors by type and infrastructure failures, and counts the trials the model refused apart from failures.

A cost that no harness reported is estimated from OpenRouter's price list, and marked "est.".

A run with more than 10% infrastructure failures is refused at publish.

Public pages no longer show who uploaded a run, or how it was uploaded.

Methodology is in the header.

The site has a new look, built on one design system, in light and dark.

Benchmarks shows one row per benchmark, with filters that pick the board.

An upload replaces the site's own API and run tokens with [REDACTED], as it does model keys.

An upload from an hbench release that node-q/hbench never published is refused.

Anyone can upload any hbench job and publish it to its board. Benchmarks and boards replace arenas, and the site no longer runs benchmarks itself.

A board has tabs for its ranking, value, runs, versions and a comparison of two configurations task by task.

Charts can be copied as an image, downloaded or embedded on another site.

Every harness name links to the harness's page.

A job is published once, even after you delete it.

The privacy notice says which per-task results of a published run are public, and what deleting a job keeps.

curl -fsSL https://nodeq.tech/install.sh | sh installs hbench's latest release, and the site accepts uploads from any hbench release whose results it can compare.

Each board has one ranked list, the same on every page, and every published run has a page of its own.

The FAQ and footer link the Discord support and announcements channels.

Every response tells browsers to use HTTPS only.

Publish a job from Your bench to its board.

A job on a board can be compared with the board's other configurations.

An entry on a board has a README badge.

Uploads are checked against hbench's job manifest.

A job uploaded before is refused, and each run counts once.