Verdict and basis

A verdict is Firsthand's current judgment about one counterparty, carrying a confidence and a basis. The basis is the list of observations the judgment rests on.

standing     grounded
confidence   0.82
basis        3 observations

Confidence decays with recency, so this is a recorded run from 2026-09-08 rather than a live figure. The explorer reads the current number.

It is arithmetic, not a model call

Nothing in the decision path calls a language model. The engine loads the prior from HOT state, pulls the observations from COLD, weighs each by recency and by who surfaced it, and computes a standing. Running it twice over the same record returns the same answer.

A sentence of English may be rendered from a verdict for a human reader. That is presentation. A test parses the judge package to assert no model client is imported there.

The confidence formula

confidence = clamp(0.40 * min(distinct_sources / 5, 1)
                 + 0.35 * (corroborated / n)
                 + 0.25 * max(0, 1 - age_days / 90)
                 - 0.25 * contradictions)

Volume counts distinct sources, not observations. Raw count is the cheapest quantity on this chain to manufacture, so pricing confidence by it would make confidence purchasable.

null is not zero

confidence: null means there was no record at all. confidence: 0.0 means Firsthand looked and found the record worthless. The two are different answers and never render the same.

The prior

Every verdict carries what Firsthand believed before it. That is what makes a verdict a position that changed for a reason rather than a guess, and the explorer shows it on its own panel.