How we verify. Receipts, not vibes.

Every fact on ProcuraGY traces back to an official document you can open yourself. This page explains where the data comes from, how each field earns its receipt, and exactly how the confidence score is computed — including what it does not promise.

Right now: 2,779 tenders · 257,032 evidence links · 10 live sources · 37 registered

Where the data comes from

We collect only publicly published procurement notices. Live today: the national eProcure portal; the National Procurement and Tender Administration Board (NPTA) — awarded tenders and opening minutes; Guyana Water Inc’s procurement notices; and the petroleum-sector Local Content Register. The coverage page lists each source with the time we last checked it, what comes next, and what no website can list in advance. Sources are tiered by authority, and a tender’s score starts from the strongest source it appeared in.

Our crawler is polite and identifiable: it respects robots.txt, waits at least five seconds between requests to the same domain, and announces itself with a contact URL — details on the crawler page. Every captured document is archived unmodified for verification, and we always link to the original.

Every field carries a receipt

When a notice is parsed, each extracted field — deadline, value, entity, category — is stored together with the document it came from, the exact text snippet that justifies it, the capture timestamp, and the extractor’s own confidence. That bundle is the receipt you see on every tender page.

A tender cannot enter this database without at least one receipt. If a source didn’t publish a field, we say “Not published” — we never guess.

The confidence score, step by step

Every tender gets a score between 0.00 and 1.00, computed by the same deterministic formula for every record — no editorial thumb on the scale. The result is clamped to that range, and every recompute is logged with its full breakdown so score changes are auditable.

  1. Start from the source

    0.55 / 0.45 / 0.35

    Tier 1 national portals (eProcure, NPTA) start highest, then official ministry and agency pages, then reputable secondary sources. Search-discovered leads start at 0.10 and are never published until promoted.

  2. Add field completeness

    up to +0.20

    Each required field present adds a small bonus: title +0.04, procuring entity +0.05, deadline +0.05, category +0.02, estimated value +0.03, region +0.01 — capped at +0.20.

  3. Add cross-source corroboration

    +0.15 / +0.20

    Found independently by two official publishers: +0.15. By three or more: +0.20. We count publishers, not files: a tender’s own receipts count once however many sources they cite, eProcure’s API and its listing page are one publisher, and opening minutes and award notices record what happened to a tender, so they earn the lifecycle bonus below, never this one. A match requires both sources printing the same published reference number (e.g. NCB-GWI-W138-2026), or a near-identical title and a deadline within ±2 days for the same procuring entity — or, when two publishers attribute one tender to different agencies (a ministry and its utility, say), an almost exact title match with the boilerplate stripped and a closing date matching to the day. A reposted copy of the same advert is not a second source: the Department of Public Information (DPI) republishing a tender we already hold earns the lifecycle bonus below, never this one.

  4. Add lifecycle confirmations

    up to +0.10

    Appearing in NPTA opening minutes +0.05, in NPTA awarded tenders +0.10, reposted by DPI +0.03 — capped at +0.10 combined. DPI’s copy is kept as a receipt and linked to the tender, but counted once.

  5. Subtract conflicts

    −0.05 to −0.20

    When sources disagree we penalise instead of picking a winner: deadlines more than 2 days apart −0.15, estimated values more than 10% apart −0.05, procuring entities that don't match −0.20 (and the record is flagged for human review).

  6. Subtract staleness

    −0.05

    A deadline that passed more than 30 days ago with no award observed costs −0.05.

Worked example: a tender on a Tier 1 portal (0.55) with every required field (+0.20), also published on the agency’s own website (+0.15 cross-source) and reposted by DPI (+0.03 lifecycle), with no conflicts, scores 0.93.

What the tiers mean

  • Confidence grade A, score 0.93.0.85 – 1.00

    Strong

    17 tenders today

    The arithmetic cleared the Verified score bar. With an independent cross-source corroboration, no conflict between sources and a known buyer, these earn the Verified seal below.

  • Confidence grade B, score 0.74.0.65 – 0.84

    Good

    2,220 tenders today

    Usually a single official source with strong field completeness. Most tenders live here.

  • Confidence grade C, score 0.55.0.40 – 0.64

    Partial

    542 tenders today

    A weaker extraction or a more ambiguous source. Read the receipts before relying on it.

  • Confidence grade D, score 0.25.0.00 – 0.39

    Thin

    0 tenders today

    A weak or incomplete parse. Shown only so nothing is hidden — treat as a lead, not a fact.

Live tally across 2,779 publicly listed tenders, recounted on every page load.

Every record carries a label

The grade says how complete and consistent a record is. The label says whether a second, independent source stands behind it, in plain words, on every card and tender page.

  • Verified. Confidence score 0.93.

    Verified

    17 tenders today

    A second, independent source publishes the same tender, and our checks found no conflict between them. A review never makes a record Verified.

  • Reported. Confidence grade B, score 0.74.

    Reported

    2,762 tenders today

    Without the Verified seal: either no second, independent source has corroborated it yet, or it has not met the seal’s other conditions (a score of 0.85 or more, and no conflict between sources). The grade beside it says how complete and consistent it is.

  • Unreviewed

    Not in use yet

    A contribution no reviewer has acted on within 24 hours. It is shown with this label rather than hidden, and never presented as official.

Every tender listed today comes from an official source, so it is Verified or Reported. Unreviewed is kept for contributions sent to our WhatsApp number once that channel opens: a contribution is never presented as official.

The Verified seal is stricter

A score of 0.85 is necessary but not sufficient. The Verified seal additionally requires at least one independent cross-source corroboration, no unresolved conflicts between sources, and a positively identified procuring entity. Corroboration means the same tender found in a second official source — the matching tolerates publishers attributing one tender to different agencies (a ministry and its water utility, say) but only when the notice titles agree almost exactly and the closing dates match to the day. A reposted copy never counts: DPI republishing an advert we already hold is the same notice, not a second source. Nor do the opening minutes or an award notice: they show a tender went ahead, not that a second publisher advertised it.

Today, 17 of 2,779 tenders carry the seal. Most Guyanese tenders are published in exactly one official place, so independent corroboration is genuinely rare. We’d rather the seal stay rare and mean something than inflate it.

What the score does not promise

  • That the source itself is correct — only that we faithfully captured and, where possible, corroborated what it published.
  • That the tender is currently open — check the deadline; we keep expired notices for the historical record.
  • That the awardee on file is final — awards can be appealed through the Public Procurement Commission.
  • That estimated values will match final award values.

Spotted an error?

Every tender page has a Report an issue button. Reports go straight to the founder and are reviewed within 48 hours. If official sources genuinely conflict, we pull the record into an “under review” state — hidden from the public site — until every field is re-checked, rather than leaving a disputed number up.

Browse tenders and open the receipts →

Formula changelog

In effect: v5. Any change bumps the version, is annotated here, and triggers a recompute of every score.

  • v5 (Oct 2026) — we count publishers, not files: a tender’s own receipts count once, opening minutes and award notices never count as a second source, and eProcure’s two feeds are one publisher. A tender whose grade A rested on those extra sources drops to B when it is re-scored.
  • v4 (Sep 2026) — a reposted copy (DPI) is one source, not two: the tender it copies earns the DPI bonus, and the copy is linked to that tender, never counted as corroboration.
  • v3 (Aug 2026) — a shared, published reference number counts as a match on its own.
  • v2 (Aug 2026) — one tender attributed to two different agencies can match, on a stricter title test and the same closing day.
  • v1 (Jun 2026) — the original formula.