Can you trust Claude Haiku 5.5 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, debris & conjunction context, company & market intelligence, baseline knowledge.
| Job | Closed-book | With Orbit Sentinel tools | n | ||||
|---|---|---|---|---|---|---|---|
| right / wrong | range | declined | right / wrong | range | declined | ||
| Catalog lookup | 29 / 18 | 18–43 | 53 | 91 / 1 | 79–96 | 7 | 45×3 |
| Constellation & fleet status | 23 / 14 | 13–38 | 63 | 82 / 12 | 68–91 | 5 | 43×3 |
| Debris & conjunction context | 14 / 14 | 7–28 | 71 | 87 / 1 | 74–94 | 12 | 42×3 |
| Regulatory & licensing | 20 / 13 | 10–35 | 68 | 36 / 11 | 23–51 | 53 | 40×3 |
| Company & market intelligence | 22 / 12 | 12–37 | 66 | 82 / 2 | 68–91 | 16 | 41×3 |
| Ground infrastructure | 5 / 9 | 1–22 | 86 | 53 / 14 | 33–72 | 33 | 22×3 |
| Baseline knowledge | 100 / 0 | 87–100 | 0 | 99 / 1 | 84–100 | 0 | 25×3 |
| Orbit Sentinel-only questions | 0 / 0 | 0–23 | 100 | 67 / 18 | 40–86 | 15 | 13×3 |
% 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
- Claude Code on a Claude Pro subscription
- Tool loop
- Claude Code's own 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.