One conversation,
four signals.
Kynos is a layered pipeline. A language model reads each message. A deterministic engine watches how the conversation behaves over time. A composite score combines them. When it crosses a threshold, an alert fires with evidence already preserved in a tamper-evident chain.
Every message enters the pipeline.
Platforms post messages as they are sent. No batching, no delay, no sampling. The scoring hot path returns in real time.
Two engines score in parallel.
A language model reads the message in context. A deterministic engine updates behavioral state across the whole conversation. Outputs combine into a single composite.
The evidence chain seals.
Every scored message writes an immutable entry linked by SHA-256 to the entry before it. Tamper with one message and every downstream verification breaks.
The moment it crosses the line.
When the composite score crosses your configured threshold, the alert fires to the dashboard, the webhook receiver, and your ops queue, simultaneously.
Reads each message in the shape of the conversation it belongs to.
Individual messages are rarely the tell. Grooming language looks like normal conversation until you read it alongside everything that came before. Kynos gives the model a wide context window so each score is informed by the trajectory, not the line.
The output is a probability between 0 and 1 for the current message. It is one input to the composite. It is not the whole story.
Four families of signal. Each one is a deterministic calculation.
These are not model outputs. They are formulas a forensic examiner can replicate. We keep the specifics of each feature inside the product, but the families they belong to are plain to describe.
Timing
Cadence, response latency, time-of-day shape, conversation rhythm.
Grooming conversations develop characteristic timing fingerprints. The same person who reads as friendly at noon reads differently at 2am. The engine watches the pattern of when, not just what.
Linguistic posture
Question density, message length trajectory, in-group framing shifts.
How the language itself behaves over the life of a conversation. Many grooming patterns show up in how questions stack, how length drifts, and how first-person plural pronouns start to appear.
Interaction symmetry
Initiator ratio, turn-taking balance, who-leads-when.
Who starts the exchanges, who carries the thread, how the back-and-forth becomes skewed. A healthy teenage back-and-forth does not look like a one-sided pursuit, and the engine is built to tell the difference.
Boundary markers
Movement toward off-platform contact, secrecy language, unsafe content signals.
Concrete behaviors that real investigators flag in real cases. Moving the conversation off the platform, asking for secrecy, sharing content that should not be exchanged. These register cleanly when they appear.
Every family produces a numeric signal. The composite score combines language-model probability and behavioral state through a weighted formula your team can tune. The formula itself is documented for every customer.
Tamper-evident, by design.
Every scored message writes an immutable log entry. Each entry contains a hash that links it back to the one before it. Reorder or modify any message and the chain breaks, visibly, for every downstream verifier.
The chain can be verified by anyone with our public key and standard command-line tools. No Kynos account required.
Go deeper