Crash testing service for builders & businesses
AI tools tested under pressure. Decisions backed by evidence.
Independent, human-led testing of AI products, agents and workflows under real-world conditions. Understand performance, failure points, limitations and the human oversight required before your next decision.
- Human verdicts
- Tested methodology
- Evidence backed
- No sponsorships
Every test, published in full.
Dia Browser
ceiling
Couldn't reach the repo or see a render, so build work needed a second tool.
Notion AI
6,000 words
One 6,000-word doc cut to its opening ~1,400 words.
Perplexity AI
citation check
One response had 3 of 8 citations fabricated.
Four phases, same method.
The method gathers the evidence. A human makes the call.
Understand the claim. Establish the baseline.
Understand the product, its claims, intended use and initial experience. The tool meets a real task cold, with no prompt engineering and no setup coaching.
I stop right where a rushed user would stop. If the tool needs a manual to look competent, that's already the finding.
Push beyond the happy path.
Apply realistic pressure and look for failure, friction, inconsistency, edge cases and the limits of what it can do.
Nothing fails on the happy path. I deliberately pick the messy, real project — the one I'd never volunteer for.
Capture what actually happened.
Preserve outputs, evidence, operator observations, unexpected behaviour, interventions and failures.
The evidence is the asset. Every number in the verdict traces back to something I logged.
Turn evidence into a clear conclusion.
Establish what holds, what fails, who it is useful for, its boundaries and what the evidence supports.
The number on the page is the number I'm willing to defend. If I can't defend it, it doesn't publish.
Notes from the operator.
The thinking behind the tests — what a score alone can't tell you.
Currently in the rig.
Verdict pending