The methodology.

Output describes what the data shows. It does not assign verdicts. It does not aggregate intent. It does not direct the reader.

Forbidden patterns

Hover any line for a description of what the pattern catches.

Rejected → calibrated

• ✗ The contractor failed to deliver.
  ✓ The dataset does not show a final modification or closeout for this award.
• ✗ Evidence of fraud in the set-aside award.
  ✓ The award's UEI has no SBA SDVOSB certification active on the award date.
• ✗ VA should have competed this contract.
  ✓ The award is recorded as not competed under FAR 6.302-1.
• ✗ The firm got the contract in exchange for favors.
  ✓ The firm's SAM.gov registration lists the same address as three other VA recipients.
• ✗ This raises serious questions about oversight.
  ✓ The dataset shows 14 modifications raising net obligations from $120,000 to $2.4M.

Try it yourself

Type any sentence below. PYTH-GOV runs the calibration sweep against it in real time, the same way every output is gated before it ships.

Accepted phrasings

• · The dataset shows net obligations of $134M on VA awards between FY2017 and FY2024.
• · Pattern consistent with sole-source growth across the modification chain.
• · Records indicate the award's last modification was dated July 3, 2023.
• · Comparable solicitations typically state the period of performance and the place of performance.
• · Public records do not contain a certification for this UEI on the award date.

Every prose claim ships with at least one source pointer. No quote may exceed 15 words. No source may be quoted directly more than once.

• Tier 1 — GAO and VA Office of Inspector General reports. Highest authority.
• Tier 2 — VA budget submissions and performance reports, Congressional hearing records, CRS reports.
• Tier 3 — Major-outlet news with editorial review.
• Tier 4 — Other news. Requires multiple corroborating sources before any claim ships.

Every output of an accountability AI must itself be accountable. The token records what evaluated the input, what model produced the analysis, what gates passed, and what audit was issued.

Glassbox VA is independent and has not been reviewed by any agency. Its patterns respond to concerns that public oversight bodies have already documented. They are listed here as context, not as endorsement.

• GAO High Risk List — Improving VA Acquisition Management; Managing Risks and Improving VA Health Care
  GAO's standing list of federal programs at high risk of fraud, waste, abuse or mismanagement.
  Frames the spending patterns (competition, concentration, contract growth, modification scope).
• Payment Integrity Information Act of 2019 · OMB Circular A-123, Appendix C
  The federal framework for identifying, reporting and reducing improper payments.
  Frames zombie, ghost-capacity and excluded-party checks as payment-integrity prompts.
• 38 U.S.C. 8127 (Veterans First) · 13 CFR Part 128 (SBA VetCert)
  VA's set-aside and sole-source preference for veteran-owned small businesses, and SBA's certification rules.
  Defines what the set-aside and certification detectors check.
• FAR (48 CFR Ch. 1) · VAAR (48 CFR Ch. 8)
  Federal and VA acquisition regulations: competition (Part 6), exclusions (9.405), SAM registration (4.11), modifications (Part 43).
  Cited in each detector's definition and recommended review actions.
• VA Office of Inspector General
  Independent audits and reviews of VA programs, including veteran-owned small business eligibility and contract oversight.
  Context for which patterns reviewers have found consequential; not a source of matches.

The public data carries documented traps. Every query guards against them explicitly; the full register with live counts is on /transparency/data-quality.

• US-1
  USAspending's cumulative award totals are not frozen per modification — reading them per row misstates value
  Guard: Money is the SUM of per-transaction federal_action_obligation over non-deleted rows (app.award_current); the record page shows each delta and the running total
• US-2
  Corrections delete earlier transactions (correction_delete_ind = D)
  Guard: Deleted rows kept for audit with is_deleted = true and excluded from every total
• US-3
  Deobligations produce negative transactions
  Guard: Summed as published; detectors discount evidence strength when present
• US-4
  Awards to individuals appear as MULTIPLE RECIPIENTS / REDACTED DUE TO PII
  Guard: Counted in money totals, never as recipients or in entity detectors
• US-5
  Awards without a valid UEI (DUNS era, exempt types)
  Guard: Grouped by normalised name with a name-only penalty and caveat

Pattern detection runs offline into precomputed match tables. Several thresholds are explicit Glassbox VA choices rather than regulatory definitions. They are listed here so no reader mistakes them for agency rules; the data lineage is on /methodology/sources and the operator description is in docs/METHODOLOGY-V2.md.

• Evidence strength (0–1) is separate from severity.
  Every stored match carries an evidence-strength score: how far past its threshold it sits, discounted for data-quality flags (missing UEI, name-only matching, deobligations, joint ventures). Severity says how large the match is; strength says how well-documented it is. Labels: weak < 0.40, moderate < 0.70, strong ≥ 0.70. Formula reference: src/lib/patterns/strength.ts.
• Thresholds are relative percentiles over rolling windows.
  Detectors compute cutoffs as percentiles of the comparable population — same administration, office or PSC — with an absolute floor as a secondary guard, and recompute them at every refresh. A match means 'top decile of its peers', not 'above a number that stopped being reviewed'.
• Silence and concentration use a rolling 36-month window.
  Zombie recipients and vendor concentration are measured relative to the latest award action in the data, so the window advances with each refresh instead of growing against a fixed date.
• Set-aside checks start on January 1, 2023.
  SBA took over veteran-owned small business verification from VA's Center for Verification and Evaluation on that date. CVE records are not public in bulk, so earlier awards are not checked against certifications.
• Growth is net obligations over the first obligation.
  US contract money is published per modification. Sole-source creep and award growth divide the sum of every transaction's obligation by the first transaction's obligation — never a cumulative column read from one row.
• Data quality is scored, not just listed.
  The /transparency/data-quality register records, per documented issue, how many rows and dollars it touches at refresh time and how it is guarded. These are properties of the published data, not findings about any entity.

Each agent operates within bounded perception. The strata are not decorative — they are how the architecture is scoped.

• PYTH-LEAD · Task agent (descends along provenance strands)
  Orchestrator, schedule, scope cuts
• PYTH-DATA · S4 source-linked
  Award-layer SQL, data-issue guards, keyword retrieval, ingest
• PYTH-SYN · S3 semantic
  Similarity scoring, awardee patterns, calibrated recommendation text
• PYTH-FE · S2 procedural
  Every UI component, every page, every motion. Reads DESIGN-SYSTEM.md as authoritative.
• PYTH-GOV · S3 semantic
  Calibrated-language gate, Proof token completeness, route-existence + visual smoke gates. Veto authority.