Intelligence ModelsOverview

Intelligence Models

Repute AI processes public Telegram data through five specialized intelligence models. Each model targets a distinct structural dimension. Together, they compose into the REP score.

The Five Models

ModelFull NameWhat It MeasuresOutput RangeFeeds Into REP
AVIAttention Velocity IndexAcceleration of attention over timeUnbounded (normalized)Yes
SISStructural Integrity ScoreAuthenticity of Telegram community0–100Yes
NSMNarrative Strength ModelStrength and compounding power of narrativeUnbounded (normalized)Yes
BPMBreakout Probability ModelProbability of structural breakout0–1Yes
TRITelegram Risk IndexBot risk + decay + fragmentation (inverted)0–100Yes (as penalty)

How They Combine Into REP

REP is a weighted composite of the five model outputs, scaled to a 0–1000 range. TRI acts as an inverse input — higher TRI reduces REP.

The general composition formula:

REP = scale(
  w_AVI × AVI_normalized +
  w_SIS × SIS_normalized +
  w_NSM × NSM_normalized +
  w_BPM × BPM_normalized −
  w_TRI × TRI_normalized
)

Where scale() maps the weighted sum to the [0, 1000] range using a monotonic transformation, and weights reflect the relative importance of each structural dimension.

Model weights are not disclosed in full to prevent gaming. The relative ordering of importance is: SIS > AVI ≈ NSM > BPM > TRI (inverted).


Model Independence

Each model is designed to measure a structurally independent dimension. A community can have:

  • High AVI (rapid acceleration) but low SIS (bot-driven acceleration)
  • High SIS (authentic community) but low NSM (narrative not spreading)
  • High BPM (breakout-ready structure) but high TRI (bot risk undermining it)

This independence means composite REP cannot be gamed by optimizing a single dimension. All five must cohere.


Read Each Model in Detail