Corpus updates: compare evidence and answers against a versioned manifest

Record source identities and versions, additions, modifications, deletions, parser configuration, chunking and tokenizer artifacts, embedding model, index configuration, hierarchy derivations, and access-policy versions. Include query rewriting, reranking, packing, prompt, generator, and decoding configuration for answer-level replay. Distinguish source publication time, ingestion time, and effective time. A document can be newly ingested but already superseded. Keep build and validation outcomes so a missing source can be traced to exclusion, parsing failure, permission, or index publication.

Build a coherent candidate generation and validate its source mappings before making it the intended serving snapshot. The actual storage and search implementation determines what atomicity or consistency is available; a single manifest file does not make several databases update together. Requests need a way to identify the generation they used or detect mixed-version reads. Parent stores, derived summaries, caches, and citation resolvers must follow compatible versions. Rollback can restore a previous compatible generation only if its retained artifacts and access rules still permit serving it.

Create tests for changed facts, removed evidence, new relevant documents, permission-only edits, boundary changes, and unaffected control questions. A correct answer may legitimately change when the authoritative applicable source changes. Update its expected evidence and adjudication with a recorded reason rather than preserving an obsolete string. For stable questions, compare retrieval coverage, selected source versions, final context, citation support, and answer quality. Reusing old exact-output assertions can reject an improved paraphrase while missing an unsupported but familiar sentence.

Aggregate accuracy can hide individual regressions. Use paired case identities to count gains, losses, unchanged successes, and unchanged failures under a comparable criterion. Slice results by document type, language, tenant scope, multi-hop requirement, and freshness sensitivity. If labels legitimately changed, distinguish that group from stable-label comparisons. Include abstention and access checks as independent outcomes; a new model that answers more questions can still leak restricted evidence or cite deleted spans. Review the concrete failure traces before accepting a net score improvement.

Probe source fetch, retrieval, expansion, answer generation, and citation resolution using the candidate generation. Verify deletions and revocations across warm caches as well as cold paths. Confirm newly relevant documents can displace old ranked results, including cached negative answers. Record generation identifiers in evaluation traces and examine any response combining incompatible versions. Evaluate latency and cost with the same correctness and access gates, because an answer regression can be hidden by silently omitting evidence or falling back to another snapshot.

The Markov lab offers a conceptual two-state propagation model only. Reset, set p=1 and q=0, predict that one Step puts all probability in state one, then Step once. Separately Reset, set p=0.50 and q=0, predict state-one probability 0.75 after two Steps, then Step twice. State zero can be called old and state one new for this analogy. The instrument has no deployment, corpus version, or consistency checker. A probabilistic migration curve cannot prove every real response uses the required coherent snapshot.