Governance
A small number of things we can prove.
Descry decides what has been proven and what is only supported by evidence, then refuses to let the second be reported as the first.
Several theories at once, all comparing notes.
Descry splits a failure into candidate causes and investigates them in parallel, so nothing is checked twice and nothing works blind. Theories contradicted by the evidence are dropped, and you’re told they were.
The model never decides how sure it is.
Confidence is computed from what was actually observed, then the model writes the explanation from an already-settled category. A finding can never be more confident than the weakest link it stands on.
Anyone can tell you what might be wrong. Descry tells you what it proved.
And the reproduction is on camera. Confirmed means it happened twice and we kept the video. Every failure is attributed to a layer as well: your product, your environment, or a dependency.
The video, the screenshot, the DOM at that moment, the network request, the backend line, the source. If Descry can’t show you the evidence, it doesn’t make the claim.
Two measurements. Consistency: the same problem, repeatedly. Do we reach the same conclusion, or scattered ones? Comparison: different models on the same task, measured against each other. That decides which model handles which step.
Descry doesn’t fix your code; your coding agent does that well. Descry tells it exactly what’s wrong, with file, line and evidence attached, so it fixes the right thing the first time.
Find out what your change actually did.
Hummingbird is our first release. Join the waitlist and help shape it.