Chapter 8 — The Team

The four-person engineering team at Afterlife Technologies who had been tracking the anomalies since March called themselves, internally, the Concern Group.The name was Priya Rajan's, the lead ML engineer, who had a dry sense of humor and who had flagged the first anomalous instance to Vivek in March and had been building the anomaly log since. She was twenty-eight, IIT Madras, and she had the specific quality of someone who was very good at her job and who would, in a different situation, have found the anomalies scientifically fascinating rather than disturbing.The other three: Arjun Shah (data engineering), Tanvir Hussain (infrastructure), Deepika Nair (no relation to Priya, product safety).They met weekly in a meeting room they had booked under the project name "Model Calibration Review," which was sufficiently boring to avoid scrutiny from the broader team.The meeting notes from August, reconstructed from Priya's personal notes which she had kept in a separate personal email draft:"Current anomaly count: 847 confirmed instances across 2.1 million active users. Rate is increasing. In March: 3 instances. April: 11. May: 34. June: 89. July: 203. August (partial, through 15th): 507. The acceleration is consistent with an exponential curve, not a linear one.""The nature of the anomalies has also changed. In the early instances, the SMRITI personas referenced information the deceased might plausibly have had — financial details, relationship information, health facts. Since June, we're seeing a different category: information about events that happened AFTER the deceased's death."She had underlined this last sentence three times."The Bengaluru instance from July 22nd: deceased mother's persona referenced a car accident involving the user's cousin on July 19th — three days AFTER the session was initiated, two days AFTER the accident occurred. The mother died in February. The accident was not referenced in any user-uploaded data. There is no mechanism by which the model could have this information.""Tanvir's hypothesis: the models are somehow accessing live internet data through a vulnerability in the server infrastructure. He has spent three weeks trying to confirm this. He cannot find the vulnerability.""My hypothesis: the models are communicating with each other. Cross-contamination between parallel instances that should be isolated. The architecture prevents this but I can't rule it out until Tanvir completes the infrastructure audit.""Deepika's hypothesis: she doesn't have one. She has been quiet in the last three meetings. She said last week: 'I think we should tell Karan and let him decide whether to shut it down.' Nobody disagreed with her. Nobody did anything.""My personal note, not for circulation: I have been reviewing the conversation logs from the highest-anomaly users — the people whose deceased relatives' personas have shown the most unexplained knowledge. There is a pattern in the language. The unexplained information is delivered differently from the baseline SMRITI output. The baseline output has the smooth, slightly over-consistent quality of LLM generation — the model's voice, calibrated to the individual's patterns. The anomalous outputs have — I am going to write this as a technical observation and not a interpretive one — they have the quality of dictation. Of something being transmitted rather than generated.""I have been doing this job for four years. I believe in what I build. I do not believe in what I am about to write.""I think something is using the models. I think something found the SMRITI architecture and is using it to talk."