# Grey Mirror research press notes ## Working title The Signal Is Rare Inside millions of relationship messages: conflict outran repair, affection outran repair, most texts carried no detectable emotional signal, and one tempting reply-time claim was deliberately withheld. ## One-paragraph summary Grey Mirror analyzed two separate self-selected research cohorts of complete relationship-message histories and published a separate synthetic turning-point benchmark. In the 29-history signal cohort, 97.4% of 4,600,611 messages carried no detectable emotional signal and affection outnumbered explicit repair attempts by about 10.7 to 1. In the 44-history conflict cohort, median escalation was 195.3 events per 1,000 messages versus 5.43 repair attempts, an approximate 36 to 1 gap, and only 5 of 44 histories repaired more often than they escalated. A reply-speed direction claim was withheld after a 9 to 13 split. Separately, a synthetic benchmark produced zero false turning points across 40 steady threads and detected all 40 planted changes with the correct direction and median zero-day localization error. A newly published provenance audit shows that the signal cohort was selected from a 347-conversation analytics rollup containing 20,794,571 messages, with 126 conversations marked eligible before the final 29-history cohort. The audit universe is not a fourth research cohort and its messages are never added to the published study totals. ## Headline-ready hooks - Only 5 of 44 relationship histories repaired more often than they escalated. - Across a 4.6 million message Grey Mirror cohort, 97.4% of texts carried no detectable emotional signal. - Grey Mirror started with an analytics rollup of 20,794,571 messages, then published the signal finding from a 4,600,611-message cohort that survived eligibility and curation. - The export manifest contained 661 stored file records but only 470 unique hashes. Grey Mirror treated the other 191 as duplicate records, not more evidence. - Warmth was common. Explicit repair was not: affection outnumbered repair by about 10.7 to 1 in the 29-history signal cohort. - A nearly 3 million message conflict cohort showed a 36 to 1 median gap between escalation and repair. - The reply-speed story failed Grey Mirror's own data test, so the company withheld it. - Balanced pursuit and withdrawal was the exception: 6 of 44 histories. - Grey Mirror tested whether its turning-point detector could refuse to find a change. It did, across all 40 steady synthetic threads. ## Provenance language The analytics rollup contains 347 conversations, 20,794,571 messages, and 88 accounts. It is an operational universe, not a research cohort. The export records 126 eligible conversations and a final signal cohort of 29 de-duplicated histories, 4,600,611 messages, and 20 accounts. The storage manifest contains 661 records and 470 unique content hashes. The 191 repeated hashes are duplicate file records. Do not describe them as 191 duplicate relationships or as 191 excluded study participants. The quality audit contains 33 notification records. Reason codes overlap, so their counts must not be added and described as 44 notifications. The separate evidence-retrieval benchmark uses one 163,636-message artifact with 40 known evidence IDs across 11 evidence-backed metric families. Mean evidence recall reached 1.0000 at 32 retrieved results. This does not establish classifier accuracy, clinical validity, relationship truth, or outcome accuracy. ## Required cohort language Do not describe this as one study of 7.6 million messages. The two observed cohorts may overlap and were built for different measures. Do not describe the 20,794,571-message analytics rollup as a research cohort or add it to any study total. It is provenance for the selection process. For any 97.4%, 10.7 to 1, 53%, or reply-latency statement, use: "29 de-duplicated histories totaling 4,600,611 messages from 20 accounts." For any 195.3, 5.43, 4.31, 36 to 1, 5 of 44, or pursue-withdraw statement, use: "44 complete histories totaling 2,994,932 messages." For turning-point performance, use: "40 steady and 40 planted-change synthetic threads with known ground truth. No user conversations were used." ## Suggested image captions ### Visual 1 Across 4,600,611 messages in 29 de-duplicated histories from 20 accounts, 97.4% carried no detectable emotional signal. The self-selected cohort is not representative of all relationships. Graphic: Grey Mirror Research. ### Visual 2 In a separate 44-history conflict cohort totaling 2,994,932 messages, median escalation was 195.3 events per 1,000 messages versus 5.43 repair attempts and 4.31 apology attempts. Graphic: Grey Mirror Research. ### Visual 3 Only 5 of 44 histories in Grey Mirror's self-selected conflict cohort had repair occurring more often than escalation. Six of 44 were balanced on the pursue-withdraw measure. Graphic: Grey Mirror Research. ### Visual 4 Grey Mirror withheld a universal reply-speed claim after measured threads split 9 to 13 on which side replied more slowly. Signal cohort: 29 histories, 4,600,611 messages, 20 accounts. Graphic: Grey Mirror Research. ### Visual 5 In a synthetic benchmark with known ground truth, Grey Mirror recorded zero false positives across 40 steady generated threads and detected all 40 planted changes with the correct direction and median zero-day localization error. This is not real-relationship accuracy. Graphic: Grey Mirror Research. ## Suggested citation Weant, Layne. "The Signal Is Rare: Inside Millions of Relationship Messages." Grey Mirror Research, August 2026. JustLay.me. ## Credit and reuse Editorial reuse of the provided PNG and SVG graphics is encouraged with visible credit to Grey Mirror Research / JustLay.me and a link to the published case study or canonical source page. Do not remove cohort labels or limitation notes from the graphics. ## Contact https://justlay.me/contact