One opened artifact: the consumer view plus the owner's close. Whoever holds this owns the connection's lifetime; hand consumers cveDbLookup only.

Open an osv.db artifact and build the owning handle over it, or reject the artifact (CveDbRejected) with its connection already closed. Nothing an artifact carries can make this throw: acceptance admits only a conformant STRICT schema whose stored values the integrity walk verified, so the provenance decode below it is total. Faults below the artifact contract (an unopenable file) still throw, and then too the connection is already closed: an exception never leaks it.

Advisory questions about one ecosystem's artifact -- the read-only view a consumer (a rule evaluation) is handed. It deliberately cannot release the underlying connection; that is the owning CveDb's capability.

Names and versions are the artifact's own vocabulary: the OSV wire package name (scope inline, e.g. @scope/name) and verbatim version text. Callers render their domain values to that form at the boundary.

One advisory segment recorded against a package: the advisory's identifier, its CVSS base score (0 to 10, Nothing when unscored), and the affected interval's bounds as the artifact stores them (verbatim version text). The lower bound arIntroduced is inclusive (Nothing == from the beginning); the upper bound is arFixed (exclusive) or arLastAffected (inclusive) or neither (open-ended). An exactly-enumerated affected version is a point segment (introduced == last_affected).

Why a downloaded artifact was refused before a handle was built over it.

A rejection is a value, not an exception: the caller (the sync task, once it exists) has a real decision to make, keep the last known-good database and alarm, rather than a fault to unwind from.

A query the accepted advisory database could not answer: the SQLite edge threw mid-query (an I/O error on the database file, a connection released out from under a straggling reader). Carries which handle field was asked and the rendered SQLError for the log line; the artifact's content can never produce this (openCveDb acceptance made the row decodes total), so it marks an infrastructural fault, not a data fault.

A confined typed exception, the same shape as CredentialError at the breaker leaf: it is thrown at the SQLite edge inside the handle and absorbed by the one boundary every advisory query runs under -- the rules engine's resilience harness (runEffectfulRule), which resolves it to an Unavailable evaluation and advances the rule's circuit breaker. It never crosses that boundary, so no caller above the rules engine sees it. The alternative (an Either on every CveLookup field) would reshape every rule's evaluation type for a fault only the harness ever handles.

An artifact version marker: S3's ETag, opaque text compared for equality only. Two objects with equal ETags carry equal bytes, so an unchanged ETag is "nothing to do", a rejected artifact's remembered ETag is "still the same bad artifact", and it names the exact advisory database a rule decision was recorded against.

Is this version inside the advisory segment's affected interval, under the ecosystem's version ordering? The interval is introduced <= v bounded above by either v < fixed (exclusive) or v <= last_affected (inclusive), whichever the segment carries, or unbounded when it carries neither. A point segment (introduced == last_affected) is affected only at that exact version.

Fail-closed. Both the remediation fast lane (a fix must not fast-track while it sits inside another advisory's affected range) and the deny gate (an unvettable version must not be admitted) want the same polarity, so every unprovable comparison, an unparseable bound or version, counts as inside: trust is only ever granted on evidence.

Does this advisory segment's severity meet or exceed the threshold (a CVSS base score, 0 to 10)? arSeverity is the artifact's normalised numeric score (Pilot reduces a CVSS vector or a qualitative label to a number at ingest), or absent.

Fail-closed for the deny direction. The threshold gates a deny, so a severity that cannot be shown to fall below it counts as meeting it: an unscored advisory (Nothing -- most of the npm malware feed) returns True. Only a score strictly below the threshold returns False.