VeriTender
METHODOLOGY · AUDIT PIPELINE

How we audit a tender.

Every finding VeriTender surfaces is reproducible, citation-anchored, and tamper-evident. The methodology is engineered to meet the evidence standard regulators apply to public-procurement oversight — not the credibility standard of a typical SaaS dashboard.

PIPELINE

Five stages, evidence at every step.

I · Ingest

Live capture from 22 jurisdictions across 7 procurement data standards.

II · Normalise

Schema mapping and multilingual translation into a canonical tender model.

III · Score

ML pattern detection across risk, collusion, fairness, and competition axes.

IV · Audit

Human analyst review with confirm / escalate / dismiss rationale on every flag.

V · Report

Hash-chained audit log and signed evidence pack export for formal disputes.

I

Live capture across 22 jurisdictions, 7 data standards.

22 jurisdictions live. Sources include TED EU (eForms 2024), ProZorro (Ukraine), eOJZ/OCDS (Georgia), Bund.de (Germany), SAM.gov (United States), Find a Tender Service (United Kingdom), e-tender.gov.az plus 5 SOE feeds (Azerbaijan), the World Bank Procurement Notices API, EBRD ECEPP, ADB, ISDB, AIIB, EIB, and UNGM.

Refresh frequency: hourly for live OCDS feeds, daily for batch SOE portals, weekly for IFI pipelines. Backfill goes deep enough to compute cohort baselines (12–36 months per source).

Standards supported: OCDS, eForms 2024, ProZorro JSON, SAM.gov CDB, eUKPROC, BMI XML (Bund.de), GE OCDS extension.

II

Schema mapping + multilingual translation.

Each source schema is mapped to a canonical tender model: buyer, lot, value, currency, deadline, awarded bidder, document set. Non-English titles are auto-translated via Claude Sonnet (UK→EN, AZ→EN, KA→EN, RU→EN) with the original language always preserved alongside the translation.

All money values are normalised to EUR at tender publication date so cross-country benchmarks are computed against a stable reference, not against today's spot rate.

Translation provenance is preserved — original language always available alongside the EN translation, with model version and timestamp recorded for audit.

III

ML-based pattern detection across 4 risk axes.

Every tender and bidder cohort is scored on four axes:

Risk — spec irregularity, vendor-specific requirements disguised as generic. Collusion — rotational bidding, price anomaly z-score below −2, UBO overlap, shared infrastructure, document-metadata leaks. Fairness — short bid window, missing notification, narrow eligibility. Competition — cohort size, win-rate concentration, repeat-winner patterns.

Seven canonical collusion patterns: shell-company bidding, rotational bidding, price anomaly, document-metadata leak, UBO overlap, complementary pricing, shared infrastructure.

IV

Human-in-loop review with expert escalation.

Automated scores produce flags, not verdicts. Each flag enters an analyst review queue with the underlying evidence — source documents, computed metrics, comparator cohort. Analysts confirm, escalate, or dismiss with a written rationale.

Escalation paths are defined per customer institutional context: audit chamber, donor agency, enterprise compliance. The platform never auto-routes a finding to a binding decision-maker without analyst sign-off.

No fully-automated decision-making — Article 14 of the EU AI Act respected by design.

V

Tamper-evident audit log + exportable evidence pack.

Every flag, review action, and rationale is hash-chained into an append-only audit log, with the chain enforced at database-trigger level rather than at the application layer. Customers can export the full evidence trail per case as a signed DOCX/PDF pack — admissible in formal procurement disputes.

The explainability layer attaches to every ML score: SHAP-style top features plus human-readable rationale that an analyst, regulator, or bidder can read without ML literacy.

Article 12 (logging) + Article 13 (explainability) of the EU AI Act both addressed in the data model, not as an afterthought.

MODELS

ML + LLM hybrid, no black boxes.

Pattern detection uses classical ML: gradient-boosted trees for risk scoring, isolation forests for outlier detection on bid-value distributions, graph algorithms for bidder-network and UBO-overlap analysis. The LLM layer (Claude Sonnet 4.5 and 4.7) is reserved for tasks where language understanding is the core requirement — translation, document-metadata extraction, and human-readable rationale on each flag.

Models are versioned with the tender release they scored. Retrospective re-scoring is possible — when a new collusion pattern is added, historical tenders are re-evaluated against it. All model versions and weights are referenced in the audit log, so a regulator can reproduce any historical score from the raw data and a published model card.

Model card per release available on request. Quarterly review with the customer's data-science team included in the Cloud Annual tier.

EVIDENCE

What counts as a finding.

Three criteria, applied to every flag before it leaves the analyst queue:

Traceable — every claim points to a specific source document with cryptographic hash and retrieval timestamp.

Reproducible — any third-party reviewer with the same data and model version arrives at the same score.

Falsifiable — every flag includes an explicit "this is wrong if X" condition, allowing the bidder or buyer to challenge.

"We do not surface speculation. If we cannot point to a document, we do not raise a flag."
NEXT STEP

Walk through the methodology with our principal.

A 45-minute briefing covers the audit pipeline against your priority jurisdictions, the models we apply, and the evidence standard behind every finding.

Request a briefing →