Narrative Risk Model

A public conversation can grow without becoming a strategic problem. A small set of messages may also begin to establish an interpretation that challenges trust in an organisation. Counting mentions reveals activity; understanding risk requires asking what story is taking shape, who supports it and why it might have consequences. The Narrative Risk Index (IRN) starts from that distinction.

The proposal takes the narrative as its unit of analysis and understands it as messages that share a way of explaining a problem, assigning responsibility or demanding a response, within a recognisable temporal trajectory. Discussing the same topic is not enough. Consider a hypothetical situation involving repeated service interruptions. Some users discuss a technical fault; others interpret events as a sign of neglect. They are discussing the same incident but calling for different responses. Grouping them without distinguishing that meaning would obscure an important part of the problem.

Five dimensions for framing a question

The model organises fifteen variables into five dimensions. Conversational dynamics examines visibility, speed and persistence. Narrative structure considers coherence, ease of transmission and moral framing. The actor dimension distinguishes legitimacy, amplification capacity and convergence among participants. Strategic sensitivity considers the subject, its relationship to what the organisation represents and its territorial context. Organisational vulnerability incorporates past events, deficits of trust and sector exposure.

The dimensions answer different questions, even when they begin with the same message. In the hypothetical case, moral framing appears when the interruption is interpreted as a failure to fulfil a duty towards users. Strategic sensitivity asks whether service continuity is central to the organisation's promise. A deficit of trust requires a history supporting that assessment; it cannot be inferred solely from the current complaint. Distinguishing these questions prevents the intensity of an accusation from becoming sufficient evidence for every other variable. The model's hypothesis is that risk depends on the relationship between narrative, circulation and context.

From traces to variables

Texts, dates, sources, relationships and changes in activity are the traces of conversation the system observes. It builds indicators from those traces. This step involves interpretive decisions. Repetition may indicate persistence, but it may also reflect duplication; shared vocabulary may help identify a narrative, but it does not prove that all participants assign responsibility in the same way. The collected data and its interpretation must remain distinguishable.

Legitimacy is neither a follower count nor a certification of truthfulness. The question is what standing an actor has to address the issue and among which audience. A user may offer direct testimony about the interruption; a specialist may contribute knowledge about possible causes. The assessment should explain the basis for that standing and preserve the distinction between those contributions. Coding criteria need definitions, examples and procedures for resolving disagreements. When information is missing, its absence should remain visible rather than silently becoming a low value.

What a formula condenses

Variables are brought onto a common scale before being combined. The formulation brings together weights, possible interactions and escalation rules. Its purpose is to preserve relevant differences, not reward the same signal several times. A statement accusing the organisation of neglecting its users may inform the assessment of moral framing, but it does not replace interruption records or the history of trust. Each assessment needs its own reason. An interaction should explain what changes when two conditions are considered together, rather than simply adding closely related labels. The parameters express design decisions that can be revised.

The result is expressed on a scale from 0 to 100. An IRN of 70 does not mean a 70% probability of crisis, nor does it independently estimate financial loss. It is a synthesis under particular data, rules and weights. Reading it requires knowing which variables contributed, which were estimated with limited information and what evidence could change the classification. In the example, persistence of the incident and the attribution of neglect may deserve attention for different reasons. The explanation of the index should preserve them so that the number does not replace the problem being examined.

A conceptual example

Return to the hypothetical situation. A user association gathers accounts of interruptions and asks for an explanation. Some posts reproduce a news report; others describe different experiences. Compared with an isolated fault receiving a similar volume of comments, the count might look equivalent. The difference emerges when asking whether incidents have recurred, whether accounts provide independent information and whether the complaint challenges the promised continuity of service. The model proposes preserving those differences in the assessment without assuming that the attribution of neglect is true.

The example is a conceptual illustration, not a performance test. It shows why narrative assessment and verification of the incident should proceed together without becoming interchangeable. A coherent narrative can contain incorrect claims, and a less visible voice can provide decisive evidence. The public plausibility of the story, the standing of its actors and factual verification are related but distinct questions. An instrument that confused them would end up confirming the dominant interpretation at the very moment it should be examining it.

Making the result open to scrutiny

The proposed validation requires reconstructing historical cases, defining in advance which event is to be detected and calculating the index using information available at each point in time. Lead time, correct detections, missed events and false alarms can then be assessed. Simulations help explore sensitivity to weights or noise; they do not replace that testing. A result appearing reasonable after the outcome is known is not sufficient either.

The interest of IRN lies in distinguishing configurations that a mention count would treat as equivalent. A conversation may acquire consequences through the meaning it organises, the people involved and the relationship between the complaint and the organisation. The model's contribution is to bring those questions together without erasing their differences. The index organises judgement; responsibility for verification, interpretation and action remains with its users. A more precise assessment should explain both why a case receives attention and what would change that decision.