Making veterans' spending transparent
scroll↓
01 · What this is

Follow the money.

The question is concrete: where does the money appropriated for veterans go, and to whom?

Glassbox VA traces U.S. Department of Veterans Affairs contracts and grants — every published modification — alongside SAM.gov registrations and exclusions and SBA veteran-owned business certifications. Every figure cites its source row. Calibrated language. Audit-ready.

We surface twelve named patterns — veteran set-aside awards to firms without a matching SBA certification, awards during an exclusion, non-competed contracts that grew far beyond their first obligation, recipients that vanished after the money flowed, offices dominated by a single supplier. Each pattern is a search you can run against the data, with severity bands, evidence arrays, known false positives and calibrated next steps.

The patterns respond to published oversight concerns — GAO's High Risk List entries for VA acquisition management and VA health care, VA Inspector General findings, and federal payment-integrity law. An independent project, not affiliated with VA. Open source, MIT, with audit tokens chained to source data — so the “why this match?” trail can always be walked back.

  • 8.94M
    VA awards
  • 9.76M
    Published transactions
  • 83K
    SBA certifications
  • 1.6K
    Set-aside checks flagged
  • 12
    Named patterns
  • 12
    Detectors live
Glassbox VA patterns

Follow the money.

Pick a pattern. Glassbox VA surfaces every record matching it, in calibrated language, with full citation.

SDVOSB Set-Aside to an Uncertified Firm

Live

The dataset shows contracts awarded under a service-disabled or veteran-owned set-aside or sole-source authority to a recipient whose UEI has no matching SBA veteran certification active on the award date.

Signal · app.award_current contracts with set_aside_code ∈ {SDVOSBC, SDVOSBS, VSA, VSS} (and SDVOSB/VOSB set-aside text) · first_action_date ≥ 2023-01-01 (SBA VetCert era) · obligated_total ≥ $25K · no sba.cert row for recipient_uei with cert_type matching the set-aside (SDVOSB for SDVOSB set-asides; SDVOSB or VOSB for veteran set-asides; JV variants accepted) and cert_start ≤ first_action_date ≤ COALESCE(cert_exit, ∞) · signal = net obligations $ · weaker when a certification exists at another date
Run detector for matches
Veteran-owned business & eligibility · 38 U.S.C. 8127 (Veterans First) · 13 CFR 128 (VetCert) · GAO High Risk List: VA Acquisition Management

Certified-Firm Address & Contact Clusters

Live

The dataset shows groups of SBA-certified veteran-owned firms that share a street address, phone number or principal and together hold VA awards — a pattern associated with pass-through and rent-a-vet schemes that the certifying agency can verify.

Signal · sba.cert active VOSB/SDVOSB firms grouped by (address_norm, zip5), phone_norm, or principal name · ≥ 3 distinct UEIs per group · known registered-agent / virtual-office addresses (sba.address_denylist, e.g. 30 N GOULD ST 82801) and any key shared by > 25 registrants excluded · group VA obligations ≥ $1M · signal = distinct firms in the group
Run detector for matches
Veteran-owned business & eligibility · 13 CFR 128.203 (control by the veteran) · 38 U.S.C. 8127 · VA OIG SDVOSB eligibility findings

Sole-Source & Modification Creep

Live

The dataset shows contracts that were not competed and whose net obligations grew at least threefold beyond the first transaction through modifications.

Signal · app.award_current contracts · base_obligated (first transaction's federal_action_obligation) ≥ $100K · obligated_total (SUM of per-modification deltas) / base_obligated ≥ 3.0 AND ≥ p90 of growth ratios among modified contracts with the same PSC · not competed: extent_competed_code ∈ {B not available, C not competed, G not competed under SAP, NDO non-competitive delivery order} or, when no competition code is recorded, an other_than_full_and_open authority · ≥ 1 modification · signal = growth ratio
Run detector for matches
Spending pattern · FAR 6.3 (other than full and open competition) · GAO High Risk List: VA Acquisition Management
+ 6 more patterns →
6 shown · 12 total at /follow

What this does.

Three primary capabilities. One coherent product.

Search the corpus

Query 8.94M VA awards — built from 9.76M published contract and grant transactions — plus SAM.gov contract opportunities, by topic, recipient, or product. Every result is auditable to its source row.

Evaluate a draft

Paste a solicitation draft to surface duplication risk, recipient concentration, and calibrated-language issues before publishing — not after audit.

Audit-ready trail

Every result, every score, every recommendation carries an audit token. Downloadable JSON. Independently verifiable.

Three checks. One score.

Every draft is evaluated against four dimensions. The composite score is the recommendation.

Similarity

The closest existing record has similarity 0.83 with the draft.

0.00
Concentration

Top three recipients account for 71% of comparable funding.

0.00
Language

Two flagged phrases against the calibrated regex set.

0.00
Composite0/ 30

Every output carries an audit trail.

Audit tokens wrap every score with full provenance. Four governance tiers — input filtering, contextual modeling, output gating, audit sealing — each with passing gates, evidence count, and retrieval timestamps. Independently verifiable. Downloadable as JSON.

View the methodology
Input
Context
Output
Audit
$0.0B
VA obligations with no public description
0.00M
VA awards (contracts + assistance)
0.00M
Published transactions summed
0
Validation tiers per output

The substance behind the surface.

Glassbox VA runs every output against a published regex set, a citation discipline, and an audit-token schema. The validator is open. The scoring is reproducible. The audit trail is the product.

View the full methodology — including the calibrated lexicon, the U.S. data-issue register, the detector thresholds, and a live try-it-yourself validator.

forbidden_absolute
/\b(fraud|fraudulent|corrupt|illegal)\b/i
GLASSBOX-GOV
12 flags
rejected → calibrated
should have
comparable filings typically state
Read the full methodology