Probabilistic Foundations
The SDK uses calibrated probabilistic models under the hood. Every belief carries a full probability distribution, not just a confidence number, which is why the system can distinguish between "confident with no evidence" and "confident with strong evidence."
The SDK exposes confidence as a 0–1 number. This page documents the underlying probability model for contributors and advanced users.
What the Math Enables
The probabilistic foundations are what make several key capabilities possible:
Two-channel uncertainty. A flat confidence score cannot distinguish between "we have not investigated" and "we have investigated extensively and the answer is genuinely uncertain." The underlying math can, because it tracks both the direction of evidence and the amount of evidence separately.
Evidence-weighted updates. When new evidence arrives, it shifts beliefs proportionally to the quality and independence of the source. A verified measurement moves beliefs more than an inference. Independent corroboration matters more than redundant confirmation.
Temporal decay. Beliefs lose certainty over time because the underlying distributions gradually relax toward their uninformative starting points. The rate is configurable, and decay is applied when beliefs are loaded, so the runtime always works with time-adjusted values.
Multi-agent fusion. When multiple agents contribute beliefs, the system merges their evidence proportionally to each agent's trust weight. Contradictions are detected through the distributions, not through string matching.
Coherence detection. The system identifies when two confident claims contradict each other by analyzing the relationship between their distributions. This surfaces conflicts that would be invisible in a flat-confidence system.
For Researchers and Contributors
If you are building on the beliefs infrastructure or want to understand the mathematical foundations in depth, contact us at devs@thinkn.ai. We are happy to discuss the underlying models with researchers and contributors.