Concept
Decision Attribution
Decision Attribution is identifying why an autonomous decision missed, including the honest question of whether the agent was at fault or conditions were adverse.
In one sentence
Decision Attribution is identifying why an autonomous decision missed, including the honest question of whether the agent was at fault or conditions were adverse.
Why it matters
Knowing a decision missed is only half the story. The half that changes what you do next is why. An insurance claims agent denies a claim; on appeal it is reopened and paid. Was the agent wrong, or did the appeal surface a document that did not exist when it decided? The two answers lead to opposite actions. If the agent reasoned badly on the evidence it had, you fix the agent. If the world moved after the decision, tightening the agent would only make it more confidently wrong on the next genuinely hard case. Attribution is the discipline that keeps you from fixing the wrong thing.
Business context
Decision Attribution shows up in the meeting after a bad outcome, when someone asks what to change. In an insurance, lending, or operations function, misses come from two very different places: the agent reasoned poorly, or the world moved after it decided. A leader cares because acting on the wrong one is expensive. Retraining an agent that was actually right only makes it more confidently wrong on the next genuinely hard case.
On the hard question
"Was the agent wrong, or did the world move?" is the question Decision Attribution exists to ask of every miss. This library names the question and why it matters. How a given system answers it is method, and method is out of scope here.
Where it fits
Attribution acts on the results that Outcome Intelligence measures, and it reads the record that Runtime Evidence keeps. It turns a miss from a number into a reason you can act on.
→
Decision Attribution
→
Right fix
When it applies
Any time a run misses its outcome and someone has to decide what to change. It matters most where misses are expensive and their causes are mixed: some the agent's fault, some the world's. Without attribution, every miss looks like an agent defect, and teams over-correct on noise.
Common misunderstandings
It is not the same as knowing a decision missed. The miss is the number; attribution is the reason behind it.
It is not an assumption that every miss is the agent's fault. Some misses belong to the world, and treating those as defects trains on noise.
It is not a method this library prescribes. Attribution stays a question every miss should face; how a given system answers it is out of scope here.
Related concepts
In the product
Where this shows up in Provy. When a run misses, Provy points at the likely
reason instead of leaving you a bare failure. How it separates an agent's fault from adverse
conditions is method this library does not describe.
See the product →
Related insights
Further reading
- Provy Research. Why observability isn't enough.
- Provy Research. Decision Assurance.