Aggregating recommendations across live detectors
What the dataset suggests a contracting officer, grants officer or auditor might review — calibrated actions distilled from 551 pattern matches across VA contracts, grants, SBA certifications and SAM.gov records, prioritised by severity and dollars at stake.
Glassbox VA observes; it does not direct. Each recommendation cites its underlying pattern matches and source rows. The responsible officials decide the response.
Sequenced execution map · dependency-cascaded · not policy advice
Highest severity or largest dollars at stake. Surface to the audit committee or program lead.
The dataset shows 50 groups of VA recipients sharing a registered address, phone number or principal ($11.2B combined). Affiliation matters for size status and competition; most shared attributes are benign.
The dataset shows 50 groups of three or more SBA-certified veteran-owned firms sharing a street address, phone number, email or principal, together holding $5.1B in VA obligations. Shared attributes are common for benign reasons; they are also the pattern VA OIG has described in pass-through and control cases.
The dataset shows 50 recipients that received substantial VA obligations and have had no VA award action for at least 36 months — $4.2B in total. Silence is not evidence of non-performance; it is the prompt to confirm closeout, final invoices and any unliquidated obligations.
The dataset shows 50 contracting-office × PSC cells with an HHI of at least 1,500 and above their administration's 90th percentile over the last 36 months, covering $2.9B. Concentration can be structural (AbilityOne, utilities, a single local market); persistent concentration elsewhere suggests testing the market.
The dataset shows 50 VA contracts awarded since January 1, 2023 under a service-disabled or veteran-owned set-aside or sole-source authority where the recipient's UEI has no matching SBA certification active on the award date — $2.2B in net obligations. Some reflect joint ventures certified under another UEI or CVE-era firms whose transferred dates differ; others are eligibility questions under 38 U.S.C. 8127.
The dataset shows 50 VA contract awards whose federal record flags the recipient as service-disabled-veteran-owned without an SBA SDVOSB certification on or before the award date ($1.2B). Because these awards were not set aside, the likely effect is on small-business goaling credit rather than eligibility; FY2024 NDAA §864 requires certification for awards counted toward SDVOSB goals.
The dataset shows 50 VA contracts recorded as not competed whose net obligations reached at least three times the first obligation — $1.1B in total. Growth within priced options is expected; growth beyond the options ceiling or the original justification's scope is the question for review.
The dataset shows 50 recipients holding both VA contracts and VA assistance awards with overlapping stated purposes; the smaller side totals $664M. Grant and contract officers can confirm the two instruments fund distinct work.
The dataset shows 50 recipients whose SAM.gov registration expired or disappeared within a year of their latest VA award action ($497M in net obligations). Lapses after contract completion are routine; lapses during performance deserve a check.
The dataset shows 1 VA awards first actioned inside a SAM.gov exclusion period for the same party (1 matched by UEI, 0 by name and state only), totalling $60K. Awards to excluded parties require a written compelling-reason determination; name-only matches may be different firms.
The dataset shows 50 recipients of substantial VA obligations that cannot be tied to a SAM.gov registration — no valid UEI on the award, or a UEI absent from the SAM.gov extracts — totalling $7.0B. Many are pre-2022 DUNS-era records or exempt award types; the rest are identity questions.
Observation-band patterns. Background monitoring; act if the trend persists.
The dataset shows 50 VA awards where the latest substantive modification description shares under 30% of its keywords with the original. Often this is FPDS data-entry practice; sometimes it is a change of scope that should have been competed (FAR 43.1).
Glassbox VA runs every live pattern detector against the source corpus, aggregates the matches, and emits structured recommendations through src/lib/recommendations/build.ts. Priority is computed from severity (flag / attention / observation) × dollars-at-stake. Each recommendation cites its underlying matches and sources — the trail walks back to app.award_current and the per-modification rows in usa.contract_txn row by row.
Recommendations frame as observations and options — never as directives or causal claims. “The dataset suggests reviewing X” rather than “X is fraud.” Severity bands describe match strength, not proven misconduct. Read the methodology for the full calibration rules.