Calibrated probabilities for real-time decisions
Reality Signal™ turns raw model scores into true probability + uncertainty, so you can safely automate, reroute to stronger models, reprompt, or escalate — and keep reliability stable as your system evolves.
AI engineers building decision systems: agents with side effects, routing layers, compliance and fraud triage, or any workflow where “confidence” must map to real risk.
For each model prediction, send a score (or richer features). Reality Signal returns prob_true, uncertainty, and a recommended boolean decision computed against your threshold.
Raw scores aren’t probabilities. Miscalibration creates hidden cost: you automate when you shouldn’t, or you over-escalate when the system could safely handle the case.
How it works
Call /decide with a score or feature payload. You receive a calibrated probability and uncertainty plus a decision_id.
Route based on prob_true + uncertainty: automate, reprompt, reroute to a stronger model, or escalate to a human.
When the outcome is known, send /feedback with the decision_id. Reality Signal learns from real-world outcomes and monitors drift.
A minimal integration
In the simplest setup, you send a single score. (If you send a features dictionary instead, Reality Signal uses a feature-based calibration mode.)
curl -X POST https://sowl-api.jollysand-1b9ed42e.swedencentral.azurecontainerapps.io/decide -H "x-api-key: YOUR_API_KEY" -H "Content-Type: application/json" -d '{
"score": 0.82
}'{
"decision_id": 123,
"prob_true": 0.62,
"uncertainty": 0.08,
"decision": false,
"threshold": 0.70
}decision is false, route the case (reprompt / reroute / escalate). If true and uncertainty is within your guardrail, automate.curl -X POST https://sowl-api.jollysand-1b9ed42e.swedencentral.azurecontainerapps.io/feedback -H "x-api-key: YOUR_API_KEY" -H "Content-Type: application/json" -d '{
"decision_id": 123,
"feedback": 1
}'0 or 1 (ground truth). Reality Signal links it back to the original decision via decision_id.Key concepts
Your environment locks its calibration mode based on the first successful calls. Score-only requests select score-based calibration; feature payloads select feature-based calibration. If you need to switch schemas, use the reset workflow.
If you configure decision economics (cost vs loss), Reality Signal can derive an optimal probability threshold automatically — so your automation policy matches business impact.