Can you trust gpt-oss 20B for space work?
By Anthony Vincent Caracappa, Founder and Self-Directed Space Researcher, Viventine Space Systems · updated
On its own, trust it for baseline knowledge. Connected to Orbit Sentinel, trust it for catalog lookup.
| Job | AloneClosed-book | With dataWith Orbit Sentinel tools | n | ||||
|---|---|---|---|---|---|---|---|
| right / wrong | range | declined | right / wrong | range | declined | ||
| Catalog lookup | 7 / 67 | 2–18 | 27 | 92 / 1 | 80–97 | 7 | 45×3 |
| Constellation & fleet status | 12 / 62 | 5–25 | 26 | 36 / 28 | 23–51 | 36 | 43×3 |
| Debris & conjunction context | 5 / 69 | 1–16 | 26 | 58 / 13 | 43–72 | 29 | 42×3 |
| Regulatory & licensing | 10 / 56 | 4–24 | 33 | 48 / 18 | 33–64 | 33 | 38×3 |
| Company & market intelligence | 2 / 66 | 0–13 | 32 | 70 / 11 | 55–82 | 19 | 41×3 |
| Ground infrastructure | 5 / 45 | 1–17 | 50 | 46 / 20 | 31–61 | 34 | 40×3 |
| Baseline knowledge | 96 / 4 | 80–99 | 0 | 85 / 5 | 66–94 | 10 | 25×3 |
| Orbit Sentinel-only questions | 8 / 20 | 2–25 | 72 | 67 / 23 | 47–82 | 11 | 25×3 |
% right / % wrong per cell; color is the verdict (green trust it, amber verify, red don't use). Range is the 95% interval for % right, counting each question once; n is questions × runs for the tools track.
- How it was run
- Open weights, run locally (Ollama on one RTX 4090)
- Tool loop
- SpaceBench tool loop
- Runs
- 1 on its own, 3 with tools; rates pool all runs, ranges count each question once
- Verdict rule
- Trust it: wrong ≤ 5% and right ≥ 80%. Don't use it: wrong > 20%. Otherwise verify its answers.
Limits, harness differences and corrections: methodology.