The current main branch and the latest published 0.2.x source tag are supported. Until the first 0.2.x tag is published, the immutable v0.1.0 source release remains supported. Pre-release branches may change without notice.

Do not post secrets, private source, customer data, credentials, or exploit details in a public issue. Prefer GitHub's private vulnerability reporting form. If that route is unavailable, send a concise report to contact@nymrel.com. Include the affected version or commit, Python and operating-system versions, a minimal synthetic reproducer, and the expected and observed behavior. No response-time SLA is promised.

This package is a truthfulness helper for a caller-supplied read result. It is not access control, authentication, authorization, input validation, a sandbox, source-integrity verification, or a guarantee that an enforcement point obeyed the result.

The package cannot detect when a caller:

• labels a failed read as successful;
• omits a required source;
• supplies forged or stale data;
• chooses an inappropriate computed verdict; or
• uses allow_empty=True without a valid policy basis.

The source-release verifier rejects symbolic links, reparse points, hardlinks, and other non-regular public-tree entries. It snapshots a supplied ZIP through one identity-checked open and requires the complete bytes to match the canonical deterministic build. These controls prevent allowlisted paths from following excluded files and prevent unreferenced bytes from hiding outside validated ZIP entries.

The archive builder stages bytes in a fresh exclusive sibling, verifies them, and atomically replaces the destination. It rejects linked output paths, redirection in output-directory ancestors, non-regular destinations, and attempts to write inside the source tree outside dist/, including alternate path spellings detected by filesystem identity. A failed staged write or pre-replacement verification therefore does not truncate an existing reviewed archive or a hardlinked victim. The destination is checked again after atomic replacement; a failure at that final check is reported but does not restore the prior output, which is another reason to use an exclusively controlled output directory and retain reviewed artifacts separately.

Packaging validation consumes the exact SHA-256-locked build closure in requirements/build.lock, disables index access during each build, compares two wheel builds, snapshots each wheel through one identity-checked open, rejects duplicate archive and RECORD names, extra metadata headers, and unreferenced ZIP-envelope bytes, verifies every recorded digest against that snapshot, and installs the result into an isolated runtime environment. The runtime package itself has no dependencies.

Passing the verifier does not prove that Git history is clean, that every printable secret pattern is absent, or that hosted repository protections are enabled. The release process separately reviews the complete public tree, immutable digest, public commit history, provider checks, and tag. Run verification and archive creation from immutable or exclusively controlled source and output directories. Path, link, identity, staging, and snapshot checks reduce static-link and partial-write risk, but pathname APIs cannot eliminate every race against a concurrently malicious directory owner on every supported operating system. Hash locking also does not prove upstream code is safe, preserve artifact availability, or authenticate the interpreter, bundled pip bootstrap, package index, and hosted runner image.

The package never logs or renders a stored value. Dataclass representations also hide labels and reasons. Generic problem summaries are safe by default. Labels and reasons remain caller-controlled and may contain paths, usernames, query data, exception messages, or credentials. Calling detailed_problems() is an explicit decision to reveal them; sanitize first. The method renders control, format, surrogate, private-use, unassigned, line-separator, and paragraph-separator characters as visible escapes to prevent multiline or invisible-control log injection, but it does not redact sensitive printable text.

assess_check() is the recommended integration API because its required boolean outcome selects exit 0 or 1. The lower-level assess_sources() keeps arbitrary verdict text independent from its exit code and defaults that code to 0; it never interprets failure-looking text. In either API, callers must use the effective exit code or explicit is_* properties. Ordinary object truthiness and an empty source-problem tuple do not mean pass.

Unbounded or infinite iterables can consume memory because assessment materializes the input once. Exceptions raised during iteration propagate. Convert a genuine source-operation failure into SourceRead.unreadable(...); do not catch broad process-control exceptions merely to produce an assessment.