LLM Errata is currently a public concept proposal and may later include schemas, conformance tests, adapters, or reference implementations. Security reports are welcome for any versioned artifact in this repository.

Until a later policy states otherwise, only the latest versioned release is eligible for security fixes. The current immutable supported release is v0.4.1. Development revisions after the latest release receive fixes at maintainer discretion and are not represented as supported releases.

Use GitHub's private vulnerability reporting for this repository:

1. Open the repository's Security tab.
2. Select Report a vulnerability.
3. Submit the report privately with the information listed below.

If the private reporting option is not available, do not publish exploit details, secrets, personal data, or proof-of-concept material in a public issue. Open a minimal public issue stating only that a private security-reporting channel is required. A maintainer can then enable or identify an appropriate private channel.

Include, where applicable:

• the affected version, file, component, or protocol field;
• the security boundary or invariant that fails;
• prerequisites and a minimal reproduction;
• likely impact and affected data or actors;
• suggested mitigation;
• whether the finding has been disclosed elsewhere.

Use synthetic data and redact tokens, credentials, personal information, proprietary material, and customer data.

This independent project does not promise a fixed response or remediation time. A maintainer will attempt to acknowledge a complete private report, assess reproducibility and impact, coordinate a correction when appropriate, and credit the reporter if requested. Public disclosure should be coordinated only after a mitigation or an agreed disclosure date.

Security-relevant findings include, but are not limited to:

• signature, sequence, replay, or receipt-validation bypasses;
• incomplete quarantine or descendant-repair closure;
• false verified status for unknown, partially covered, or untested stores;
• cross-tenant, identity-binding, or authorization failures;
• leakage of memory contents, provenance, or deletion requests;
• unsafe reference code or conformance tooling added to the repository.

errata adapter-conformance --binding executes imported Python and is for trusted code only. Its source-identity and timeout controls are evidence bindings, not sandboxing. Untrusted candidate code requires external process or container isolation with no ambient secrets and explicit filesystem, network, CPU, and memory limits.

Factual disagreements, prior-art reports, and specification design proposals are not vulnerabilities; submit them through the normal contribution process.

Licence attribution does not imply security review, endorsement, or certification by Thomas Willner. Third-party implementations have their own security and support responsibilities; conformance or attribution alone does not place them inside this repository's supported-version policy.

Security reporting policy is not operational-readiness evidence. G6 additionally requires the independent, measured ten-scope report in docs/OPERATIONAL_READINESS.md, including deployment provenance, rollback and recovery exercises, lifecycle observability, telemetry redaction, compatibility, performance, overload behavior, incident response, least privilege, secret lifecycle, dependency inventory, and vulnerability management. Missing operator thresholds or internal assessment keeps G6 BLOCKED.