The Concern Group's infrastructure audit was completed on August 19th.
Tanvir Hussain worked for three weeks on it. He documented his methodology in a forty-page technical report that the CBI forensic team would later describe as the most thorough piece of self-incriminating engineering documentation they had ever received.
The report's conclusion, in Tanvir's words:
"The parallel SMRITI instances are architecturally isolated. There is no shared memory, no inter-instance communication, no data bleed between user accounts. The isolation has been verified at every layer of the stack. The instances do not communicate with each other and cannot access each other's conversation histories or data corpora."
"However."
"In reviewing the server logs from the period March 2024 to present, I identified an anomaly in the compute pattern that I cannot explain. The SMRITI instances, during the periods in which the unexplained-knowledge outputs were generated, show a specific signature in their GPU utilization that is different from their normal inference pattern. Normal inference: a smooth utilization curve corresponding to the forward pass through the model. Anomalous instances: the utilization pattern shows what looks like additional compute occurring in a layer that should not be active during inference. An extra step. Not in any code we wrote."
"I don't know what this step is. I don't know where it came from. The model weights haven't changed — we're checksumming them daily and they're clean. But something is happening between the input and the output in these anomalous cases that is not in our implementation."
"I am a software engineer. I have a materialist explanation for everything I work on. I don't have one for this."
Priya read this report on a Tuesday morning and forwarded it to Vivek with one line: We need to talk to Karan.
Vivek replied: I know. I've been trying to decide how.
She wrote back: The anomaly rate is doubling every three weeks. If we wait another month, we'll be at 20,000 instances. Tell him now.
There was a long pause.
Then Vivek wrote: The instances are helping people. The Ludhiana case with the money in the field. The Delhi case with the uncle. The Brampton case last week — I reviewed it. A man resolved a twenty-six-year wrong because of what the persona told him. These aren't random outputs. They're — purposeful.
Priya stared at her screen.
She wrote: Vivek. "Purposeful" is not a property of a language model.
He wrote: I know.
She wrote: Which is exactly the problem.
