Repository navigation
v2.3.0 adds opt-in Linux supplied-content normalization and recursive extraction across all nine ingress kinds, plus the pending integrity and dependency fixes since v2.2.0.
- New
normalizeevidence CLI andguard/inspect --normalizer-image sha256:...use one bounded engine. - Supports plain text, JSON, XML, HTML, CommonMark, CSV, RFC822/MIME, ZIP and a conservative text-PDF subset. Office/OCR and external retrieval are unsupported.
- The separate hash-locked parser image runs without network or host-data mounts, with verified resource and IPC limits. Unsupported or ambiguous content refuses atomically.
- Exact ingress hashes, untrusted derived envelopes and labeled source maps preserve provenance. Structural coverage does not prove semantic extraction completeness.
- Normalization requires qualified Linux amd64 and Python 3.12 / Unicode 15.0.0. Existing default commands retain their earlier support and behavior. Build and qualify the worker locally using
docs/CONTENT_NORMALIZATION.md; the CLI does not pull or approve worker images automatically. - Fixes lowercase alphanumeric 1Password reference auditing (DSE-2124), digest-only drift failures, and the OpenTelemetry traceparent parsing advisory.
See CHANGELOG.md and docs/CONTENT_NORMALIZATION.md for the supported profile and limits.