Skip to content

Evidence and Confidence

TrustSight labels evidence along two independent axes, and neither is a probability.

  • The evidence tier (A-D) describes how a signal is sourced and how much external state it needs before it means anything.
  • The evidence layer (L1-L8, plus the optional L9 artifact lane) describes what kind of check produced a finding.

"Confidence" in TrustSight is not a statistical probability attached to a package. It is the combination of three visible facts: which evidence tier the signal comes from, whether the check had complete coverage, and how much history the database holds. All three are reported; none is hidden behind a single score.

Evidence tiers

Tier Source Cold-start behaviour
A Structural Rule matches against the parsed recipe and diff Always available
B Priors Static source-bucket classification Always available, deterministic
C History URL and maintainer novelty from the local database Gated by maturity; meaningless on run one
D Verification Declared checksums, keys and pinning Reported at weight 0, never credited

The full taxonomy, including maturity gating and the cold-start rules, is in Evidence Tiers.

Evidence layers

The layer is a report and coverage view, not a sequence of barriers. A finding at L4 does not mean an attacker passed L1-L3; it means the check that caught it belongs to the evasion-shape category. A layer that stayed silent is not proof of safety.

Assurance Layers owns the layer model, the coverage states and a worked report example. Evidence Categories owns the classification table and the rule-family mapping.

Confidence in practice

  • A weight-0 finding is context, not lower confidence. It is shown because its absence would be misleading.
  • A finding on an incomplete analysis is still evidence; it does not license conclusions about the parts that were not read.
  • A cold database withholds novelty rather than inventing it, and may produce Inconclusive when novelty is the only thing holding the score up.
  • A declared practice (checksum, key, pin) is a statement by the recipe, not a verification performed by TrustSight.