Determinism - Kviria

Same input, same answer

Kviria’s engine does not sample, rank by model opinion, or hedge. The same anchor and the same words produce the same decision every time. When your user asks the same question twice, your agent gives the same answer twice - no drifting, no alternating between candidates.

The evidence trail

Every response carries an evidence array: the tests that produced the decision, each with a result and a machine-readable reasonCode.

"evidence": [
  { "test": "resolve", "result": "pass", "reasonCode": "single_survivor" }
]

The trail is an audit surface, not user-facing copy:

Debugging: when an answer surprises you, the reason codes say why the gates decided the way they did.
Trust: the coordinate your agent presents is backed by a trail it can show a skeptical integrator.
Never render raw evidence to end users - your agent phrases the outcome.

Honest failure

Determinism cuts both ways: when nothing matches, Kviria says notFound - with the absent category when that is the useful fact - instead of inventing the closest plausible-sounding place. An agent that never hallucinates a location is worth more than one that always answers.

Coverage

Kviria is in beta with coverage in San Francisco (approximately lat 37.70-37.84, lng -122.54 to -122.36). Outside that area every query returns an honest notFound.