Running pattern detector against the corpus
The dataset shows the gap between named policy priorities (emissions, housing, reconciliation, healthcare) and the actual flow of funds — concrete spend versus stated plan.
Glassbox surfaces correlations and observations, not findings of misconduct. Severity is a calibrated read on how strongly the dataset matches the pattern definition — a starting point for review, not a verdict.
The match meets the pattern threshold but the signal is the lower band. Normal review queue.
The match exceeds the routine threshold. Worth a closer look — pull the source rows, walk the citation.
The match is in the highest band of the calibrated scale. Recommend: prioritise for active review.
Severity is computed from the detection signal · ILIKE keyword match on prog_purpose_en / prog_name_en, summed over the rolling 5 years to the corpus as-of date, compared to a calibrated stated annual commitment per priority · signal = actual ÷ stated
The dataset shows federal grant & contribution spending tagged to climate & emissions reduction ran at ~$3.5B/year over the 5 years to 2026-09-01 (35,256 agreements across 11 departments), versus a stated annual baseline of ~$9.5B (36% of stated). Glassbox shows the gap; the policy decision-maker confirms whether the baseline measure is right.
Why this may be benign · Program-purpose keyword matching is a proxy: programs can serve a priority without naming it, and stated commitments are budget-cycle estimates, so the gap measure is indicative only.
Aggregated from the matches below. Severity bands are calibrated per pattern; counts and dollar totals reflect this snapshot only.