Golden Bench

About

What this is

Golden Bench measures how well software agents play a 1997 console first-person shooter, running natively from a community decompilation: a harness turns each moment of play into a text and structured description, the agent picks an action, the harness carries it out, and the run is scored on objectives, completion time and survival. Language models are the first contestants; the harness is meant for decision models too.

Results are numbers: scores, times, decisions, costs, and the positions and health recorded along each run. Paths and timelines on this site are our own drawings of those numbers on blank grids.

What this is not

The boundary

The game belongs to its rights holders: its code, levels, art, sound, music, on-screen text, characters and marks. None of that is on this site. Specifically:

If the decompilation or the harness disappeared tomorrow, every page here would still make sense: it is results and a description of the rules.

How it's built

Runs happen offline on the operator's Mac. A sweep plays each agent on each season level and seed, verifies each run by replaying its input log, and scores it. An export step gathers the sweeps' result bundles into a handful of JSON files (an index with the leaderboard, a level table, a run list and one file per run), keeping only numbers and our own words. This site is a static build over those files with Astro: no server, no database, no tracking. Charts are inline SVG drawn at build time; a small script per page animates the replays and is not needed to read anything. The only thing loaded from elsewhere is one typeface.

Data from 5 sweeps, harness commits 32a2193, 674047e, ac6c751, b4ba945.