Skip to content

Corpus Behavioral

The package's position in the corpus, or its deviation from a corpus baseline. These are Class D and Class E: they run once per metadata cycle in trustsight full-aur, after the per-package loop, and each returns one finding per cluster with the members in params.members.

They are silent without a prior snapshot, which is enforced rather than assumed: the calibration gate is fire_rate(no_baseline) == 0. R112 and R107/R111 are prioritisation rather than accusation; they say what a compromise would reach, not that anything is wrong.

See the rule system reference for the field table, the severity weights and the reserved identifier ranges.

R073: Accelerated Release Cadence

  • Target: programmatic (git commit graph)
  • Severity: metadata field, never a scored finding
  • Category: temporal-metadata
  • Condition: The HEAD commit has 3+ ancestors within the last 24 hours (rapid-fire pushes).

Why it does not score: Bursts of commits are overwhelmingly benign activity - a maintainer fixing a bad checksum, then a typo, then a rebuild bump. This is precisely the "measuring activity, not risk" failure the accuracy work's CI gate (|pearson(score, diff_lines)| < 0.3) exists to catch. At any non-zero weight it becomes a census on active maintainers.

R073 is therefore metadata only: recorded as a boolean on the PackageFact (recent_commit_burst: bool). It is not appended to triggered_rules and contributes nothing to the score. If future corpus analysis shows that burst cadence pairs with other signals (burst + maintainer takeover + verification removal), the burst multiplier can be applied to those signals alone - never as a standalone finding.

Origin: pnpm's minimumReleaseAge (24-hour cooldown on new versions) and uv's exclude-newer (reject packages published within a configurable window). Both tools impose a registry-side cooldown: do not install a version until it has existed long enough for the community to vet it. R073 takes the opposite perspective - instead of blocking recent versions, it notes that multiple commits landed in a short window, recording the cadence as context for other signals to use.

R093: Orphan/Adoption Dependency

  • Severity: MEDIUM (weight 15)
  • Category: dependency
  • Condition: A package depends directly on a package orphaned or adopted in this cycle.

R107: Transitive Exposure

  • Severity: INFO (weight 0)
  • Category: dependency
  • Condition: A package's transitive closure, at least [thresholds] r107.min_hops (default 2) hops away, reaches a package adopted out of the orphan state this cycle. Context only, never additive.

R111: Transitive Orphan Exposure

  • Severity: INFO (weight 0)
  • Category: dependency
  • Condition: A package's transitive closure includes a currently orphaned package. Context only, never additive.

R112: Dependency Centrality

  • Severity: INFO (weight 0)
  • Category: dependency
  • Condition: A package is depended on by at least [thresholds] r112.min_dependents (default 50) AUR packages. Prioritisation only: it says what a compromise would reach, not that anything is wrong.

R125: Introduction Rate Deviation

  • Severity: MEDIUM (weight 15)
  • Category: adoption
  • Condition: The corpus-wide introduction rate for a cycle deviates from the baseline mean by at least [thresholds] r125.z_score (default 3.0), once at least r125.min_history_cycles (default 3) cycles of history exist. An immature history is quiet.

R106: Known Indicator of Compromise

  • Severity: tiered by the indicator's confidence: confirmed is FATAL, high is CRITICAL, medium is HIGH, an untiered entry is MEDIUM
  • Category: ioc
  • Condition: A declared surface exactly matches an entry in iocs.toml.

Four surfaces are read: the package's own name; names added to depends/makedepends/optdepends/checkdepends/provides/replaces; the host of any URL and any bare host token; and any hex digest of digest length. R106 also reads the current PKGBUILD, not only the diff, so a dependency on a package later published as malware is reported on every review rather than only on the one that introduced it.

Matching is exact. Normalisation is limited to what is not part of the identity: case, a trailing root dot, IDNA spelling, surrounding quotes. A host is never stripped of a subdomain and a name is never stemmed, because that would turn equality into resemblance. A malformed entry is dropped with a warning rather than coerced.

The shipped list is empty. TrustSight does not invent indicators, so a fresh install cannot fire R106 at all, and a miss is uninformative. trustsight corpus pivot <indicator> inverts the rule: given one indicator, it lists every corpus package referencing it, reading only stored material and never the network.

A confirmed indicator cannot be suppressed through overrides.json.