Jev Decision Evidence¶

Inspect the evaluation that produced a binding without changing the binding's native type or making another provider call. For a field-by-field API reference, see Decision Evidence.

Decision Evidence in Kedi¶

Jev-backed output bindings retain their decision evidence without changing the normal output type:

```
from kedi import decision_info
```
> adapter: pydantic
> model: typesafe/jev-latest

[ticket] = I was charged twice for the same invoice.
>> It is [duplicate_charge: bool] that <ticket> describes a duplicate charge.

`print(decision_info("duplicate_charge").probability)`
`print(decision_info("duplicate_charge").threshold)`

[ticket] := I was charged only once.
`print(decision_info("duplicate_charge").matches_inputs(ticket=ticket))`

The helper resolves the nearest lexical binding. For an interactive session, session.decision_info("duplicate_charge") inspects a top-level binding without executing another fragment. Unknown names raise NameError; values without Jev evidence return None.

Evidence includes the provider/model, source location, Noul probability, the threshold used for boolean conversion, and Choice/Score confidence and distributions when the provider supplied them. Noul does not gain an invented confidence value. Multiple fields produced by one template share one call record while retaining field-specific evidence. For a nested or multilabel output, answers contains each matching child path separately; Kedi never merges several Noul answers into one probability.

state_fingerprint, criteria_fingerprint, config_fingerprint, and request_fingerprint identify the evaluated provider state, resolved criteria, decision configuration, and their combined request. These are one-way identities, not recoverable prompt copies.

matches_inputs(**values) performs a local fingerprint comparison over the explicit <name> inputs. It makes no model or tool call and does not claim that history or external state is unchanged. Reassigning an input leaves the historical record intact, so this check can report False. Reassigning the output removes the record from that current binding. Kedi does not automatically re-evaluate a decision or apply a new threshold.

Decision records participate in strict interactive-session dump/load. They contain versioned fingerprints, not recoverable copies of the original input text. A fingerprint is an identity check, not anonymization; do not publish snapshots that may contain sensitive provider evidence.

Capture Decisions from Python Calls¶

kedi.capture_decisions() collects the completed template decisions produced by @kedi.query and @kedi.bind calls without changing their return types:

import kedi

@kedi.query(adapter="pydantic", model="typesafe/jev-latest")
def review(ticket: str) -> bool:
    """kedi
    >> It is [duplicate: bool] that <ticket> describes a duplicate charge.
    = `duplicate`
    """

with kedi.capture_decisions() as run:
    duplicate = review("I was charged twice for the same invoice.")

for decision in run.decisions:
    print(decision.field_path, decision.probability, decision.threshold)

decisions is a tuple of immutable DecisionInfo views, one per output binding, in call completion order. Multiple fields share their call_id. These historical snapshots remain accessible after the function returns. Nested capture scopes also contribute to their enclosing scope; separate concurrent scopes are isolated. Use regular with around await calls and await desired work before exiting. Exit does not wait for pending work; the capture stops accepting records.

Capturing makes no additional model calls. Response-cache hits, failed calls, and outputs without provider evidence do not create decision records. Successfully completed decisions remain available if a later call raises. This captures typed template outputs, not implicit control-flow claims or arbitrary SDK calls.