LLM Errata: please try to break this experimental AI-memory correction protocol

If an AI system imports memory from another system, should it also accept responsibility for future corrections?

LLM Errata is Thomas Willner's experimental, vendor-neutral conformance proposal for corrections, supersessions, and erasures that arrive after AI memory has already been copied or used to derive summaries, embeddings, profiles, graphs, or caches.

The proposed lifecycle is:

observe -> quarantine -> rebuild -> test -> attest

An importer authenticates an erratum, blocks the affected root and known descendants before recall, rebuilds from still-valid inputs, runs negative, positive, and preservation checks, and returns a signed receipt. If a relevant store or descendant cannot be inspected, the result must say unknown rather than silently claiming success.

This is a human-owned, substantially AI-assisted project. Thomas Rainer Willner directed the work, is the publishing author, and accepts responsibility for its claims. The repository records prior-art collisions, schemas, semantic probes, test vectors, a runnable reference implementation, threat boundaries, and explicit conditions for narrowing, redesigning, or retiring the proposal.

Please challenge it

Useful responses include:

- an earlier public implementation or normative profile covering the complete conjunction;
- a counterexample that breaks quarantine ordering, lineage coverage, receipt integrity, privacy, sequencing, or preservation;
- evidence that the design is impractical for real memory systems;
- an independently authored adapter or separately authored validator; or
- a smaller design that achieves the same user outcome.

The active review target is commit ac4468faf73c2cc7949dd29b2a2a151f5bd23116, canonical surface digest 7e0d6c88c1ca3a87743ac70ba2a3dfea0b350d112d2d3c59a3c6cbb537568f12:

- Repository: https://github.com/thomaswillner/llm-errata
- Review contract: https://github.com/thomaswillner/llm-errata/blob/ac4468faf73c2cc7949dd29b2a2a151f5bd23116/REVIEW_REQUEST.md
- Independent implementation contract: https://github.com/thomaswillner/llm-errata/blob/ac4468faf73c2cc7949dd29b2a2a151f5bd23116/INDEPENDENT_IMPLEMENTATION.md

Recent external feedback is still being assessed. It includes a reported preservation-test blind spot where duplicated retained facts may remain recallable and therefore escape a presence-only check. A separate independent reproduction reports 28/28 published wire/controller vector cases passing, while execution tracing found that only 1/28 reached an adapter; its author explicitly states this does not validate adapter behavior or move G2/G4. The external harness also found and mutation-tested three vacuous-pass defects in its own checks before publication. Treat the repository as active work, not a settled standard.

Status and use boundary

Current verdict: NOT_PROD_READY. Internal code and tests do not establish independent conformance, audited constant-time production cryptography, two independent implementations plus a third-party validator, operation across three real memory systems, or complete operational evidence. Publication and invitations do not close those gates. No endorsement, partnership, certification, audit, or commercial commitment is implied.

Independent commercial and non-commercial implementations of the specification are permitted under the repository licence. Any product or service implementing a material part must provide reasonably accessible credit to LLM Errata, Thomas Willner, and the repository. Reference code under prototype/, scripts/, and tests/ remains personal-use unless separately licensed.

Public technical findings should return to the repository's issue process; security-sensitive findings should follow its SECURITY.md instructions.