AI screening

Is a month unusual for this facility?

A robust z-score on the seasonally adjusted residual of flaring, venting, fuel use and methane performance is the gate: it applies an absolute criterion, so a series with no unusual months raises nothing. An isolation forest corroborates a point the gate has already flagged and never flags on its own. Median and median absolute deviation throughout, so one genuine outlier cannot hide the next.

Output: A question in the review queue, with the ranked signals behind it and the model version.

z threshold

3.5

min records

36

gate

robust z on the seasonal residual; the isolation forest corroborates only

Does a satellite retrieval over a facility disagree materially with the inventory?

Matches each top-down retrieval to a facility and period and scores the divergence against the bottom-up figure. A retrieval cannot attribute to a process, sees only large emitters under good conditions and samples an instant, so it cross-checks the inventory and never replaces it.

Output: A question in the review queue when the divergence is material; the inventory figure is unchanged either way.

material divergence pct

50

Of the returns that passed every rule, which should a reviewer open first?

A gradient-boosted tree over engineered, record-level features, with an attribution for each. It orders a reviewer's attention; it never blocks a return (the deterministic rules do that) and never produces a figure. Until enough reviewed outcomes exist it is not trained, and the queue falls back to a stated ordering by severity and age.

Output: A priority on each queue item, labelled with whether it came from the model or the fallback order.

min labelled reviews

200

requires both outcomes

yes