Current language models are masters of statistical mimicry, producing fluent text without any concept of whether that text is true. They learn what is plausible, not what is real. The Belief-Aware Language Model introduces a fundamental architectural change: a second output head — the Belief Head — that runs in parallel with the standard language prediction head.
Where a conventional LLM outputs a token, BALM outputs a Sentient Token: the text paired with a continuous degree of belief on a scale from −1 (refuted) to +1 (established). This is not post-hoc confidence calibration. The Belief Head is trained as an integral component through a composite loss function, forcing the model to simultaneously learn fluency and factual awareness.
Underpinning this is a Bayesian weight framework where model parameters are probability distributions, not fixed values. When new evidence arrives, the posterior updates without catastrophic forgetting — yesterday's knowledge becomes today's prior. The model genuinely learns over time, rather than remaining frozen at its training cutoff.
Read full post ↗
The same architecture produces different epistemic value across domains. In search, BALM annotates every source with a degree of belief: a government database (.gov) registers differently than a forum post. The synthesized answer reflects genuine epistemic assessment rather than SEO ranking.
In medicine, the model understands evidence hierarchies. A peer-reviewed meta-analysis from the New England Journal of Medicine carries a different degree of belief than an influencer's health blog. When claims conflict — as they often do in clinical research — BALM flags the disagreement explicitly rather than confidently presenting one side.
In finance, degrees of belief map naturally to signal conviction. A macro forecast grounded in multiple central bank communications and consistent economic data carries high belief. A contrarian thesis built on a single data point registers lower — not because it's wrong, but because the evidential basis is thinner. The investor sees the degree of belief and decides how to weight it.
Read full post ↗
The BALM API returns a streaming response that separates content from analysis. As the model generates its answer, text content streams in real time with inline degrees of belief annotating key assertions. When generation completes, a structured summary follows containing the overall degree of belief and a breakdown of assertion-level confidence.
Each assertion in the response carries a degree of belief reflecting the strength of the underlying evidence. Claims grounded in government and academic sources (.gov, .edu, peer-reviewed journals) typically register high belief. Established news organizations land moderately high. Claims sourced from unverified or commercially motivated outlets register lower — transparently, with the reasoning visible.
Integration is straightforward: send a query, receive a stream, parse the separator, render the scored results. The API handles web search, source credibility assessment, content generation, and belief scoring in a single call. We're currently onboarding partners for early access.
Request API access ↗