Public records, traced to the row.

Sources

Row counts are from the latest analytics snapshot (2026-09-04); “—” means the source has not been loaded or is not counted. Live status: /api/health.

How money is counted

USAspending publishes one row per award transaction. Each row's federal_action_obligation is the amount obligated (or deobligated) by that action alone. An award's value is therefore the sum of those amounts over all its transactions that have not been deleted by a later correction. Glassbox VA stores that sum as net obligations in app.award_current, one row per award.

The source files also carry cumulative columns (current_total_value_of_award and similar). They reflect the award at file-build time, not at each modification, so they are kept for reference and never used for money. Each record page shows every transaction's own amount next to the running total, so any figure can be re-added by hand.

Fiscal years are U.S. federal fiscal years (October 1 – September 30), labelled by the year in which they end. Award-level views (office, program and recipient pages) attribute each award to the fiscal year of its first action; the fiscal-year trends page instead counts each transaction in the fiscal year of its own action date, so multi-year awards are spread across the years they were obligated, and it marks years where the source is incomplete. Recipients are identified by their SAM.gov Unique Entity ID (UEI); awards without one are grouped by name and flagged.

Glassbox VA's own choices

These are choices made by this project, not definitions from any agency:

• Relative thresholds. Most detectors flag the top decile (90th percentile) of a peer group — same administration, office or product code — with an absolute dollar or ratio floor as a second guard.
• Rolling windows. Silence and concentration are measured over the 36 months before the latest award action in the data.
• Set-aside checks from 2023. Certification is checked only for awards from January 1, 2023, when SBA's VetCert program took over verification from VA's Center for Verification and Evaluation.
• Evidence strength. A 0–1 score of how far a match sits past its threshold, discounted for data-quality problems. Formula: src/lib/patterns/strength.ts.
• Registered-agent deny-list. Addresses known to host thousands of unrelated companies (for example 30 N Gould St, Sheridan, WY) are excluded from address-cluster detectors.

Detectors and the tables they read

• SDVOSB Set-Aside to an Uncertified Firm
  app.award_current · sba.cert
• SDVOSB Flag Without Certification
  app.award_current · sba.cert
• Award to an Excluded Party
  app.award_current · sam.exclusion
• Certified-Firm Address & Contact Clusters
  sba.cert · sba.address_denylist · app.award_current
• Registration Lapse After Award
  app.award_current · sam.entity
• Sole-Source & Modification Creep
  app.award_current · usa.contract_txn
• Vendor Concentration
  app.award_current
• Zombie Recipients
  app.award_current
• Ghost Capacity
  app.award_current · sam.entity
• Modification Purpose Drift
  app.award_current · usa.contract_txn
• Duplicative Funding
  app.award_current
• Related Parties
  sam.entity · sba.cert · sba.address_denylist · app.award_current
• Geographic Reconciliation Variance
  gdx.expenditure · gdx.usa_state_year · gdx.band · gdx.reconciliation · usa.assist_txn · usa.contract_txn

Independence

Glassbox VA is an independent project of Pythagorithm. It is not affiliated with, endorsed by, or sponsored by the U.S. Department of Veterans Affairs or any federal agency. Data are drawn from public government sources and may contain errors.

Pattern matches are observations from public records. They are not findings of fraud, waste or misconduct, and they are not determinations of eligibility. Known benign explanations are listed with every match.