Built on Alberta TRACE.

TRACE — Targeted Review of Alberta's Contracts and Expenditures — is a program of the Alberta Ministry of Technology and Innovation. It uses agentic AI on open government datasets to identify patterns of interest in public spending: charity funding loops, ghost-capacity entities, sole-source amendment growth, and several others. TRACE's data outputs (entity matching, loop detection, anomaly flags) are published in the cra.* and general.* schemas of the read-only Render replica that Glassbox queries.

Source citation: Nate Glubish, “We're taking Alberta's AI to Ottawa”, Substack, April 13 2026 (as recorded in Glassbox's Session 8 TRACE integration report).

A note on naming. The term “TRACE” does not appear anywhere in the upstream repository Glassbox builds on (the Alberta “AI For Accountability Hackathon” data platform — its FED/, CRA/, AB/ and general/ modules). The mapping between the Ministry's published description of TRACE and the upstream analysis scripts listed in the provenance table below is Glassbox's interpretation, not a label the upstream authors applied. Where a Glassbox pattern has a direct upstream script, the path is cited; where it does not, the pattern is marked Glassbox-native.

• Federal corpus expansion. TRACE is built on Alberta provincial data; Glassbox extends every TRACE-derived pattern to the 1.33M-row federal grants & contributions corpus.
• Calibrated-language discipline. Every Glassbox output passes the calibration sweep — no “fraud”, no “clearly shows”, no “should have”, no causal claims. Patterns surface correlation and observation, not guilt.
• Audit-token provenance. Every output carries an audit token chained to the source data tokens, so the “why this match?” trail can always be walked back.
• Public-facing UX. TRACE is an internal review program. Glassbox is its public interface — searchable, browsable, citable.

Several thresholds and windows in Glassbox are Glassbox choices. They are listed here so no reader mistakes them for upstream or Ministry definitions.

• Loop-score attention threshold of 12 is a Glassbox choice. The upstream scoring script (CRA/scripts/advanced/02-score-universe.js) reports the score distribution at ≥ 15 and ≥ 10; it does not define an “attention” cut. Glassbox's 12 / 15 / 18 bands are calibration choices for the observation / attention / flag pills.
• Zombie recipients use a rolling 36-month window. The upstream script (FED/scripts/advanced/05-zombie-and-ghost.js) uses a fixed cutoff — last agreement start before 2022-01-01. Glassbox instead measures silence relative to the refresh date (last agreement start < today − 36 months), so the set moves forward with each refresh rather than growing indefinitely against a stale date.
• Evidence strength is a Glassbox score. Every stored match carries an evidence-strength value in [0, 1] — a read on how much of the published record supports the match (row volume, recency, corroborating rows across sources), separate from severity. Labelled weak (< 0.40), moderate (< 0.70) and strong (≥ 0.70). Formula reference: src/lib/patterns/strength.ts.
• Thresholds are relative, with rolling windows. Where v1 used fixed dollar and ratio cutoffs, v2 detectors express thresholds as percentiles of the comparable population (same department, program or fiscal-year window) and recompute them on every refresh. A match is therefore “top decile of its peers this window”, not “above a number chosen in April”.
• Amendment values are cumulative. In the federal corpus, agreement_value on an amendment row is the running total, not the increment (landmine F-3). Every growth ratio Glassbox reports is latest ÷ original for the chain keyed by (ref_number, business number or legal name) per landmine F-1 — never a sum across amendment rows.

The operator-facing description of the v2 pipeline is in docs/METHODOLOGY-V2.md.

Which Glassbox surface uses which upstream data product, and whether the detection logic comes from an upstream script, from Glassbox, or both. Upstream paths are relative to the upstream repository root.

Glassbox is designed to align with Alberta's public-sector AI policy: sovereign compute (no third-party AI for pattern detection — calibration enforcement is local regex, semantic retrieval is opt-in), open-source models, audit trails on every output, and accountability via citation rigor. See the compliance mapping for the line-by-line policy correspondence.

Alberta Ministry of Technology and Innovation has not endorsed Glassbox. We credit the TRACE methodology lineage; we do not claim partnership. Glassbox observations from public records are not findings of misconduct.