Understanding the Proofence Trust Score

Every verification eventually has to produce something a person or a system can act on. A wall of raw signals is honest but unusable; a single green checkmark is usable but dishonest. The Proofence Trust Score exists to sit between those extremes: a compact summary of how much confidence a given output has earned, backed by the reasons behind it.

This post is about what a trust score is and, just as importantly, what it is not. We will not walk through the internal weighting, and there is no secret number that unlocks the model. The interesting questions here are conceptual: what feeds a score, why calibration matters more than raw accuracy, and how to keep a number from becoming an excuse to stop thinking.

A score is a summary, not a source of truth

The single most important thing to understand about a trust score is that it is downstream of the evidence, never a substitute for it. The Proofence Engine gathers signals from grounding, cross-model consistency, policy alignment, and behavioral checks. The score is a way to summarize those signals so that people and pipelines can triage quickly. Every score in AtlasProof by Proofence stays linked to the evidence that produced it, so a reviewer can always open the number and see why.

What feeds the score

At a conceptual level, several families of signal contribute to any trust decision. We keep the exact combination internal, but the categories are not secret — they are the honest inputs any serious verification system has to weigh.

• Evidence grounding: how well each claim is anchored to a retrievable, relevant source.
• Cross-model agreement: whether independent models converge or diverge on the answer.
• Policy alignment: whether the output respects the organization's stated rules and constraints.
• Uncertainty: how stable the result is when the question is reframed or retried.
• Provenance: how trustworthy and current the underlying sources themselves are.

Notice that these signals can pull in different directions. A claim can be strongly grounded in a source that policy says should not be used. Two models can agree confidently and both be wrong. A good scoring approach does not paper over this tension. It carries the disagreement forward so that a low or borderline score comes with a legible reason, not a shrug.

Why calibration beats precision

It is tempting to chase precision — to make a score as sharp and confident as possible. We think calibration matters more. A calibrated score means that when the system says it is fairly confident, it is right about as often as that confidence implies, and when it is unsure, it says so. A score that is precise but poorly calibrated is worse than useless, because it earns a trust it has not proven, and teams learn the wrong lesson only after something breaks.

A trust score that is always confident is not a trust score. It is a mood.— Proofence Research

Calibration is also why we resist collapsing everything into a single all-purpose threshold. What counts as "trustworthy enough" for a first-draft internal summary is not the same bar as a customer disclosure or an agent action that moves money. The score is designed to be read alongside the stakes of the decision, not in place of them.

// Illustrative — the fields below are conceptual, not a real formula.
{
  "trustScore": 0.62,
  "band": "review",
  "contributing": [
    { "signal": "grounding", "direction": "positive" },
    { "signal": "consistency", "direction": "negative" }
  ],
  "explanation": "Well grounded, but two models disagreed on scope.",
  "recommendedAction": "route-to-human"
}

Explainability is the whole point

If a score cannot explain itself, it cannot be trusted, and a trust signal that cannot be trusted is a contradiction. This is why explainability is not an add-on in Proofence but the reason the score exists in the first place. Every score should answer three questions on demand: what pushed it up, what pushed it down, and what a reviewer should do next. Those answers are what let an auditor reconstruct a decision months later and what let a team improve their prompts, sources, and policies over time.

Understood this way, the Proofence Trust Score is not a black box passing judgment. It is a calibrated, explainable summary designed to speed up the easy cases and flag the hard ones for the people best equipped to handle them. Its job is not to end the conversation about whether to trust an output. Its job is to make that conversation faster, fairer, and fully on the record.