Our validation — what we've actually measured

The methodology page on this site ends with a promise: no accuracy number for BeltUp LLC's own systems without the methodology alongside it. This page is that disclosure — the first one, kept honestly narrow to what has actually been measured so far.

What is being measured

ViralDecay's engine makes two separate kinds of call about a piece of tracked content, and — as the methodology page argues — these are not interchangeable, so they are reported separately rather than folded into one headline number:

• State and direction — is this content stable, still gaining momentum, declining, or already effectively dead? This is closer to the discrimination task described on the methodology page than a timing prediction: the question is whether the model correctly separates "still going" from "already over," not exactly when the ending arrives.
• Time-to-death (the countdown) — a bounded estimate of how many hours or days remain, analogous to the Remaining Useful Life problem surveyed on the predictive maintenance page, including that field's convention that a countdown's accuracy is expected to tighten only as the end actually approaches.

The number: state/direction accuracy

Against an early real-world labeled validation corpus — content tracked and manually confirmed against its actual outcome — the engine's state/direction classification is currently correct 83% of the time. The classifier is built as a multi-detector consensus rather than a single rule, the same architectural family as the controlled-false-alarm surveillance methods described on the epidemic dynamics page: several independent checks have to agree before a state changes, which is deliberately biased toward fewer false alarms over a marginally higher headline number.

This corpus is small and growing — an early disclosure, not a final one. The number will be re-measured and republished here as the labeled corpus grows, consistent with the actuarial convention surveyed on the mortality page: a validation figure is only as good as its most recent re-check, and a single early measurement is reported as exactly that.

What is not yet validated

Why a modest, disclosed number is still a differentiator

As documented on the methodology page, a review of major commercial social-analytics platforms found that none of the enterprise incumbents publish a validated, methodology-disclosed accuracy number for virality or decay prediction at all. An 83% state/direction figure, reported with its methodology and an explicit acknowledgment of what is not yet measured, is a different kind of claim than either the silence from the market leaders or the unsourced percentages from smaller tools — not because the number is large, but because it is checkable.