How Concessio Extracts Contract Clauses From Primary Documents

Last updated: August 5, 2026

Published August 5, 2026. Figures as of August 2026 — full methodology at /docs/methodology.

A clause is only as useful to a diligence team as the paper trail sitting behind it. This page walks through how a clause moves from a primary concession document — a gazette notice, a regulator filing, a procurement portal upload — into Concessio's clause library, why "extracted" doesn't mean "trusted by default," and what keeps a clause out. That library holds 3,750 clauses today. All of them carry a source URL, none of them cite a banned source, and 1,663 of them — 44% — carry a citation pinned to a specific page of the source document.

How does a clause get into the library in the first place?

Extraction starts with the document, not the database. Concessio's clause pipeline is built to run only against primary sources: concession contracts and annexes published by government PPP units, procurement portals, sector regulators, gazette publications, stock-exchange disclosure systems (SEC EDGAR, HKEX and peers), and multilateral development bank project databases. A fund's portfolio page or a Wikipedia summary is not an acceptable source under that standard, and every clause in the visible library today is screened against it — which is why the zero-banned-source figure above is zero, not "close to zero."

What happens when the source document is a scan, not searchable text?

Not every concession contract shows up as searchable text. Photocopied gazette pages, scanned regulatory annexes, and PDF exports of paper originals are a routine part of the source mix. Before any clause typing can run, those documents go through an OCR pass that converts the scanned image into machine-readable text. This step is where a lot of extraction pipelines quietly lose fidelity — a skewed page or a low-resolution photocopy can turn a defined term into gibberish — which is part of why the audit step described below exists downstream of it, not instead of it.

How is a clause typed and pulled out of the document?

Once the text is machine-readable, each contract section is classified into one of Concessio's clause types. The table below shows the twelve largest types by volume — not the full taxonomy; smaller clause types make up the remainder.

These 3,750 clauses span 832 distinct concessions. Coverage is concentrated, not uniform: most of the 22,000+ concessions Concessio tracks have no extracted clause yet. The clause library is a growing subset of that universe, not a claim of comprehensive coverage.

What does "verbatim extraction" actually mean?

The clause record stores the contract language itself — the actual sentence or paragraph as written in the source document — not a paraphrase of what the clause does. The extraction rule for structured fields (trigger conditions, cure periods, notice periods, remedies) is that they are pulled only from that quoted text, not from general knowledge of how a clause of that type "usually" works. If the quote doesn't support a structured field, the field stays empty rather than being filled in from assumption.

What does the page-level citation capture?

Alongside the source URL, each clause record stores the specific page (or, where pagination doesn't apply, the section reference) of the source document the quote came from. That's the difference between "this deal has a clause about this somewhere in a lengthy concession agreement" and "here is clause 7, page 17." Right now 1,663 of the 3,750 visible clauses (44%) carry that page-level citation; the rest cite the source document without a page pinpoint yet. That's an honest gap, not a rounding error.

Why is extraction audited rather than assumed correct?

Because OCR misreads, a classifier can mistype a clause, and a quote can be transcribed with a dropped clause number or a misread date — and none of those failure modes are visible just by looking at the final record. So extraction is followed, as a matter of policy, by an independent re-derivation check: a separate pass goes back to the same cited page of the same primary source and re-derives the clause independently, confirming that the quoted language and structured fields actually appear where the citation says they do. That check is the bar a clause has to clear before it's treated as usable — it is the standard the library is built on, not a claim that every record has already been re-audited or that extraction is trusted at face value.

What does an extracted clause actually look like?

Take Highland Schools PPP1, a Social Infrastructure concession in the United Kingdom. The library's record here is a duration clause — the provision setting out when the agreement takes effect and when it expires, not a termination clause — cited to Clause 7, page 17, sourced to highland.gov.uk. The extracted quote reads:

"7.1 This Agreement shall take effect on the Contract Award Date. 7.2 Subject to the provisions of clause 39 (Survival) this Agreement shall terminate on the Contract Expiry Date unless it is terminated prior to such date in accordance with..."

That's the pattern for every clause in the library: a named contract section, a page number, a source domain, and the actual contract text — not a summary of what the clause is understood to say.

Does the pipeline work on contracts that aren't written in English?

Yes. The Kumamoto Airport concession in Japan has a change-in-law clause in the library, cited to page 38, section 第4.-2.-(3), of a filing hosted on mlit.go.jp — Japan's transport ministry. The clause is extracted directly from the Japanese-language primary concession contract, not from a translated summary or an English-language secondary description of the deal. Clause typing and citation capture run against the primary document in its original language; the OCR and extraction steps aren't limited to Latin-script sources. Of the 1,663 page-cited clauses, 499 are in Latin-script languages.

What's the honest summary?

3,750 clauses, all source-cited, zero on a banned source, 832 concessions covered, and 44% with a page-level citation today. Extraction, as designed, runs through a document-acquisition stage, an OCR stage, a typing stage, a verbatim-quote stage, a citation-capture stage, and an independent re-derivation check before a clause is treated as confirmed — and the page above describes what's live now, gaps included, not a finished state.

Figures as of August 2026. Full methodology at /docs/methodology.

Request Demo