[acgxv]

Derived memories can omit a subject, effective date, or qualification. Agents also copy long quotes and generate a second citation list that can disagree with native claims. Add a bounded evidence pack that keeps selected notes linked to deduplicated immutable source windows, and resolve native source references under current read permissions, hashes, and byte bounds. Expose the same contract through HTTP and MCP; keep semantic support explicitly unevaluated.

Apply source-context and provenance patterns reviewed in Hindsight, Graphiti, RAGFlow, LightRAG, and LlamaIndex without adding a database, model call, or graph-truth authority. Diagnose graph fragmentation and missing temporal/semantic edges separately. Derive answer-level citation lists from native claims without changing facts or references.

Validation:

• 460 Rust unit tests, 105 integration tests, two focused source-reference integration tests, 181 Python contracts, Clippy, strict style, format, and documentation checks passed.
• Retain 36 paid answers with unchanged questions, initial candidates, sources, and full 4096-dimensional Qwen embeddings. Native-consumer replay passes jointly on 16/16 original test, 14/16 contrast, and 2/4 development cases. All 36 native bindings pass; this is not a semantic truth certificate.
• Preserve 24 raw redundant-envelope failures and raw strict joint results (4/16, 7/16, 1/4). Replay derives only the redundant list. The final shorter prompt has no new paid generation. This is one new read-path condition, not a rerun of the previous 132-case matrix or a competitor comparison.
• Reproduce prior 132-case evidence, unchanged verdicts on 131 accepted legacy payloads, and all eleven database fingerprints. New and prior evidence verifiers pass offline.
• Add $0.19228509; cumulative paid $8.09924852, conservative exposure $9.18304892 against the authorized $10 ceiling.

Remaining gaps: missing initial candidates and two historical-time interpretations. Shared-source graph connectivity is navigation, not proof of causality. Owned test containers, volume, and API process are removed; raw evidence and verified backup remain. No production deployment.

See docs/evidence/benchmarking/2026-10-12-evidence-context.md for the full evidence and limitations.