{
  "failed_cases": [
    "post/memory-0-0",
    "post/combined-001"
  ],
  "shared_missing_fact": "STREAM-0-286",
  "stage": "native Memory retrieval",
  "observations": [
    "The correct original is stored as raw message 15. Derived episodic message 16 also contains the correct decision with valid_at 2026-08-15.",
    "Neither failed reader context contains STREAM-0-286. Both include partial August 22 instructions and unrelated Cobalt entries.",
    "The shared reader abstains on the absent delivery mode while correctly returning the newer retry policy.",
    "Native Memory queryMessage explicitly requests valid_at descending order, and the Elasticsearch connector passes that sort into the query body.",
    "Raw message validity timestamps are import timestamps; some derived records use event dates, while other extracted records keep import time. Some extracted text loses the subject or assigns the import date to the event."
  ],
  "inference": "Sorting mixed timestamp meanings favors later-imported or current-dated records. Combined with fragmented or generic notes, this can displace an earlier but still-valid project decision from top 6. The stored decision is not lost, and the reader is not shown it in the two failed cases.",
  "not_claimed": [
    "A sort-policy intervention was not run.",
    "This is not an optimal-configuration quality ceiling.",
    "The two failed questions are not two independent root defects."
  ],
  "source_urls": [
    "https://github.com/infiniflow/ragflow/blob/v1.0.0-rc1/internal/service/memory.go",
    "https://github.com/infiniflow/ragflow/blob/v1.0.0-rc1/internal/engine/elasticsearch/chunk.go"
  ]
}