What an LLM-built knowledge graph keeps, and what a repair pass against its sources changes

A data note on one deployment of an agent memory system whose knowledge graph was extracted from its corpus by a language model, each relation citing the chunk of text it came from. Two studies, each under its own public pre-registration, ran on the same frozen uniform sample of 300 relations. The first asks how often the cited chunk supports the relation as written. Two judges from different model families, a check that each judge's quote occurs in the chunk, and a blind human grade of 173 rows give 0.913 (274 of 300, Wilson 95% interval 0.876 to 0.940). Both judges passed a negative citation control and a mismatched-evidence control. On 60 rows drawn at random from those the two judges agreed on, agreement between the human and the judges was low (kappa 0.155), which the note reports as a limit. The second tests a repair pass of the kind recently described for agent memory as "dreaming": a model rereads each relation against its chunk and keeps, rewrites or deletes it. Run once with claude-opus-5-5, the pass raised precision among the relations it left from 0.913 to 0.990 (+7.6 points, bootstrap 95% interval +4.7 to +10.7) and kept none of 30 relations shown with another document's chunk. The 180 relations it kept unchanged were already 0.989 supported, so most of the measured gain came from which relations it chose to leave alone, not from the content of its rewrites. It also rewrote 87 of the 274 relations that were already supported, 30 of them into a relation between a different pair of entities, so it kept 0.650 of the supported relations unchanged. Three human grades were regraded after a spreadsheet emptied their evidence cells; with the first grades, precision after would be 0.979. A repair pass should be measured on retention of the original facts as well as on precision. The frozen sample, the repairs and the human grading sheets are private and not released; their hashes are in the public commitments. The archive carries the note, its provenance file, both pre-registrations, the commitments, both readings, the census and the scripts.