# SpaceBench

> A two-track space-domain LLM benchmark: the same questions asked closed-book and grounded via MCP tools against the Orbit Sentinel satellite catalog & regulatory database. Scoring is calibration-aware (correct +1, incorrect -2, abstention 0) with negative controls. Grounded answers re-freeze from re-runnable SQL as reality changes, so the benchmark resists training-data contamination.

Current corpus: 83-81e122a887e2 (83 questions), published 2026-07-22, 20 runs. Results data license: CC BY 4.0. Created by Viventine Space Systems LLC (https://viventine.com).

## Pages

- [Leaderboard](https://spacebench.net/): current results, both tracks, gap and cost charts
- [Methodology](https://spacebench.net/methodology/): grading rules, scoring, contamination resistance, disclosures
- [Questions](https://spacebench.net/questions/): full grounded corpus with SQL ground truth; reasoning-set examples
- [Archive](https://spacebench.net/v/83-81e122a887e2/): immutable snapshot of this corpus version

## Machine-readable data (JSON)

- [Current pointer](https://spacebench.net/data/current.json): corpus fingerprint + publish date
- [Leaderboard](https://spacebench.net/data/v/83-81e122a887e2/leaderboard.json): per-run summaries (score, outcomes, accuracy numerators/denominators, cost) + closed-vs-MCP gap pairs
- [Corpus](https://spacebench.net/data/v/83-81e122a887e2/corpus.json): grounded questions in full; reasoning questions as id/category/tier
- Per-run detail: https://spacebench.net/data/v/83-81e122a887e2/runs/{track}--{model}.json (per-question outcomes, tokens, tool calls)

## Full text

- [llms-full.txt](https://spacebench.net/llms-full.txt): complete current leaderboard and methodology as plain text