Crash Tests // The data layer
AI Crash Tests
Every tool put through the same four-phase Test Rig — realistic conditions, real pressure, the evidence logged. Honest verdicts, no sponsors — ever.
2Verdicts
published
1In the rig
verdict pending
4Phases
same method
100%Independent
no sponsors
Verdict
Type
All crash tests
Crash Test 001Mar 2026
Perplexity AI
Research / Answer engine
Fast research, shaky sources
37.5%Citation
fabrication
Primary Failure PointIt fails its own citations under load.
Full report →
Crash Test 002◇ Latest verdict
Notion AI
Productivity / Workspace AI
Convenient but shallow
23%Long-doc
coverage
Primary Failure PointSilently truncates long docs and invents names.
Full report →
Crash Test 003◆ In the rig
Claude Sonnet 4.5
LLM / Reasoning
The score, re-run
Back on the rig under the full four-phase protocol. Verdict publishes when it clears every station.
0 of 4 stations cleared · Verdict pending
Follow the run →
No tests match that filter yet — more runs are on the sheet.