Data-Quality Scoring

Every sample and feature carries a quality score summarizing noise, validation, imputation, and drift, so downstream models weight it correctly.

Quality as a first-class number

Not all data is equally trustworthy, and the fabric refuses to hide that. Every sample and every engineered feature carries a data-quality score that summarizes how much to trust it: was it in-band, was the sensor live, was it imputed, how noisy is the channel, how much drift correction was applied. Models and the twin weight inputs by this score.

What the score combines

• Validation verdict: used, filtered, imputed, or rejected.
• Signal-to-noise against the channel's known noise floor and ADC resolution.
• Imputation flag and the confidence of the graph model that filled it.
• Drift-correction magnitude since the last calibration.
• Cross-channel agreement with redundant diagnostics.

Propagation through features

A feature's quality is bounded by its worst essential input: an instantaneous Q built on a drifting neutron calibration inherits that uncertainty; a pressure map with imputed Thomson channels carries a reduced score in the affected region. The fabric propagates quality through the feature graph so the number that reaches the twin is honest about itself.

Acting on quality

Quality scores let the sensor fusion down-weight bad channels, let the anomaly ensembles avoid firing on artifacts, and let MLOps detect drift as a population-level fall in quality. The score is stored with the feature and recorded in lineage. It is computed identically for the breeder and the burner and is a design specification for machines not yet built.