How GLODA computes what it shows

GLODA stores the notice text a source publishes, its structured fields, the retrieval timestamp, and a content hash. Attachments and documents are linked to the source, not warehoused; a per-source terms review precedes any change to that. Raw payloads are kept immutable for a bounded retention window so a record can be re-parsed, then removed.

Publisher · source system and method · source record link and id · publication date, with whether the source stated it · last seen at the source · indexed and last updated · record hash and revision count · jurisdiction of the publisher · licence where recorded. “Last verified” in GLODA means the last time the record was seen at its source; it is not a claim of human or official verification.

A state describes GLODA’s confidence in its own field, at record level and per field. It is never a judgement about the notice, the buyer or any supplier.

✓Data quality: Verified

Verified: The record’s provenance traces to the publisher’s own channel, or its URL falls under a domain in GLODA’s official-domain allow-list (Phase 6). Everything else shows the neutral “Not yet verified” — visible either way, never hidden.

✓Data quality: Source-backed

Source-backed: The field is exactly what the publisher’s own channel (API, feed, or portal page) stated, or the issuer entered it themselves.

≈Data quality: Extracted

Extracted: GLODA derived the field from the notice text with a deterministic rule (a conservative regex) or a labelled AI extraction; the original text is one click away.

~Data quality: Inferred

Inferred: GLODA inferred the field from weak signals (keywords, an ingestion-time date, a buyer name matched from free text). Treat it as a hint.

◌Data quality: Incomplete

Incomplete: Several expected fields are missing from the source record itself.

–Data quality: Not available

Not available: The source record does not carry this field. Missing data is never shown as a negative.

!Data quality: Needs review

Needs review: GLODA’s own quality checks flagged the record (placeholder title, redirect page, weak summary). Check the original before relying on it.

Record-level rule, in order: review-flagged → needs review; details inferred from free text → inferred; any key field regex- or AI-extracted → extracted; three or more expected fields missing → incomplete; otherwise source-backed for official channels and issuer listings.

• Official API or feed
• Official document or portal page
• Entered manually
• AI-assisted extraction — used only when a record is thin and always labelled AI; the model may not add facts beyond the notice.
• Third-party dataset — official bulk datasets (for example historical contract awards) loaded with their vintage.
• Inferred relationship — keyword sectors, regex-matched amounts, ingestion-time dates, buyer names matched from text.

Freshness compares the record’s last update with now: fresh, recent, aging, stale. Records whose deadline has passed leave default search and are marked closed by a nightly job; records that stop appearing at a healthy source are archived. Nothing is deleted from the audit history.

• Contract value — a structured amount from the source is source-backed; an amount matched from the text by a conservative regex is extracted (confidence 60) and marked; USD conversions use the European Central Bank rate on the publication date, or the nearest available rate, flagged as approximate.
• Sectors — published CPV / UNSPSC / NAICS codes are source-backed; GLODA sectors mapped deterministically from those codes are marked as crosswalk; keyword-only sectors are inferred and drawn de-emphasised with a tilde.
• Buyer and reference — as published, or inferred from free text and marked; an inferred buyer never merges records or scopes a procurement thread.
• Publication date — as stated by the source; when the source states none, GLODA records when it first saw the notice and says so.
• Documents and contacts — as published; document hints inferred from text are excluded from counts.
• Bid-readiness analysis — on request, for Analyst plans: a language model lists what the notice states (requirements, submission steps, dated steps, cautions) and each row is kept only when its quote is found word for word in the notice; the questions it leaves open are labelled as gaps, not facts. One analysis per notice version, cached. It never scores a tender, predicts an outcome or advises whether to bid.
• Debrief and challenge rules — GLODA’s own summaries of the funder’s or jurisdiction’s published procurement rules, each linked to the official text with the section cited and a review date; sets marked partial state only what is on file. Not model-generated and not legal advice.
• Requirements checklist — conditions the notice states, extracted by a language model and kept only when the quoted passage is found word for word in the notice text; unverifiable rows are discarded. A reading of the notice, never an assessment of a bidder.
• Summary — the source’s own text; a model-generated summary replaces a missing or placeholder one and is labelled AI.

The same opportunity published on several channels is matched conservatively (normalised title, buyer, deadline day, countries). Matches are linked, never merged: one record is elected canonical for search, every sibling stays reachable with its own provenance, and the brief lists the other sources.

A procurement is modelled as a process with an append-only ledger of stage events, each naming its evidence (a notice, a notice revision, or an award record), the source, the method and a confidence. The current stage is the highest-ordered stage any evidenced event asserts; a submission deadline that an amendment extended re-opens a process an earlier notice had closed. Stages after award are never derived without a document that states them.

Rules Watch tracks published changes to the procurement rules of the markets GLODA covers — thresholds, directives, laws, policies and notice-format guidance. Each entry is GLODA’s own plain-language summary of the publisher’s official text, linked to that text, tagged to its market, and shown on the Rules Watch timeline. It is not scraped from an official feed: entries are curated and published through a source labelled as GLODA manual entry, so the provenance says exactly that. Summaries are GLODA’s words, never model-written, and are not legal advice — the publisher’s official text governs.

Every entry carries a confidence, and “reviewed <date>” — the day GLODA last checked the summary against the official text. Confidence is set only after that check.

reviewed

A person at GLODA read the change against the publisher’s official text and confirmed the summary; the review date is when that check was last done.

summary

GLODA’s plain-language summary of the official text, written from the source and linked to it — accurate to what the publisher states, without a separate line-by-line confirmation.

partial

Only the facts on file are stated; the source carries more than GLODA has summarised. Read the official text for the rest.

A change moves through published states as its instrument advances; a later instrument marks an earlier one superseded, and both stay on the timeline.

Sanctions and debarment list updates are covered by integrity screening, not Rules Watch, and only official public exclusion lists are ever reported as findings — with their source and date, and a name match is never identity.

• Several development-bank notice portals block datacenter access and are ingested through a browser lane only when a runner is available; the coverage manifest marks them dark when it is not.
• Lot-level data is stored only where a source publishes lot identifiers; GLODA never invents lots.
• Contract signature, effectiveness, variation and completion data are not yet modelled; no source in the index publishes them in a structured form.
• Financing-project links exist for World Bank project numbers; other funders’ project identifiers are wired as their sources expose them.
• Historical award datasets can be partially anonymised at the publisher (individual consultants); the record says so rather than filling the gap.

A GLODA record is an index entry, not the original document. Cite the original publisher as the source of a notice, award or rule, and cite GLODA as where you accessed it, with the record link and the date you read it. For example: “Contract award notice, [publisher]. Accessed via GLODA, gloda.org/records/<id>, on <date>.”

Every public record page carries its publisher, source system, source record link and id, publication date and record hash in the provenance block — everything a citation needs. Where a field was extracted or inferred rather than published, the record says so; cite the original text, not GLODA’s derivation, for anything load-bearing. The publisher’s official text always governs.

If a record has a wrong deadline, amount or a dead link, tell us: every public record page and in-app brief has a “Report an issue with this record” control. The report goes to GLODA’s review queue and is checked against the source; where the source has changed, the record is updated, and where the source itself is wrong, the record says so rather than quietly diverging from it.

Corrections never erase history. A record keeps its revision trail and content hashes, so a change to a deadline or value is visible as a revision, not a silent rewrite. A record removed at a source is marked closed, not deleted from the audit history.