What Process-Aware Detection Can—and Cannot—Promise
The central question is not whether messages still matter. It is whether a conversational scam can be judged safely when the most important evidence is spread across messages, responses, and changing objectives. The evidence examined in this series supports a conditional answer: slow-burn scams should be represented and detected as processes because chronology changes meaning, but process-aware detection remains decision support, not proof. Its value lies in creating time and clarity for verification while keeping its uncertainty visible.
The unit of risk is a trajectory
A single message can contain a decisive artifact. A fraudulent link, a mismatched sender address, or an impersonated brand may justify immediate inspection. But many conversational scams do not begin with a message that announces its own purpose. The opening creates contact. Engagement makes the persona and the relationship feel credible. Later turns introduce a promise, a problem, a request for secrecy, or a financial step. The significance of that request depends on the state built by the earlier exchange.
That is why the process framing is more than a change in terminology. It changes what the detector must preserve. A message-only system asks whether the current text resembles a known malicious pattern. A process-aware system must also ask what identity the sender claimed, which expectations were created, what the recipient has already accepted, how the conversation moved between channels, and whether the current request changes the direction of the exchange. The same words can have different risk when they appear at different points in that trajectory.
The Scam Kill Chain makes this progression visible without claiming that real conversations follow a perfect script. Initial Contact describes how an interaction becomes possible. Engagement describes the sequence of actions and responses through which attention, trust, pressure, or promised gain may be shaped. Termination describes the point at which the accumulated relationship approaches an extraction attempt or another decisive outcome. Psychological-technique labels make an action’s persuasive function easier to analyze, but the label remains an interpretation grounded in a reported move, not a diagnosis of a person.
The representation also needs to preserve what it does not know. A victim report is not necessarily a transcript. It can compress turns, omit uncertainty, or describe an outcome without recording the exact path to it. The benchmark construction funnel makes progression measurable by selecting reports with enough recoverable structure, but that selection means the resulting cases are not a prevalence sample and not a complete view of conversational fraud. A structured instance is evidence organized for a task; it is not the original interaction restored without loss.
This distinction carries into model evaluation. PreScam separates the question of whether risk is approaching Termination from the question of what scammer actions or techniques will follow a partial history. That separation matters because early warning and exact forecasting have different failure costs. A detector can raise useful attention without naming the next action precisely. A language model can produce a fluent continuation without recovering the step that actually advances the scam. The reported results show that local cues and broad manipulative intent are easier for current systems than temporally coherent escalation.
The architecture paper reaches a related conclusion from the system side. A recent-message window can miss a distant identity claim or promise. Full-history prompting preserves more raw text but grows in cost and can dilute salient details. A running summary can preserve selected relationships at bounded context cost, but it can omit subtle evidence. None of these choices removes the underlying problem. They are different ways of deciding which parts of the trajectory cross into the next decision.
The resulting mental model has three layers. First, there is observed evidence: turns, claims, responses, artifacts, and reported outcomes. Second, there is a structured representation: phases, actions, psychological techniques, verification records, and an evolving summary. Third, there is inference: a risk estimate, an alert, or a forecast of future action. Each layer can be useful, and each can be wrong. A confident prediction cannot repair an omitted turn. A neat phase label cannot transform a retrospective summary into a full transcript.