The open ERP contract dataset

Every row is a public record with a public URL. Twelve of the buyers are public bodies; four are private companies whose contracts are public because they were filed as exhibits with the US Securities and Exchange Commission, which is a different thing and worth keeping straight. The documents span January 2001 to October 2025, across twelve distinct delivering parties and roughly seven platform families. No client or prospect document appears in this dataset. The firm has no clients, and nothing here is anyone’s confidential paper.

Each document is read against the same fixed list of eleven checks, in the same order. The dataset records, for each check and each document, the state, the severity the instance actually fired at, whether it counts toward the standing score, the clause it rests on, where that clause sits if our record captured a reference, and a note explaining the call. The standing score counts checks C1 to C7 only, so it runs 0 to 7; checks C8 to C11 are recorded where the document addresses them but are never counted.

We say “contracts” rather than “statements of work” for this set as a whole, because one of the seventeen is an executed agreement body whose statement-of-work exhibit was never in the file we retrieved.

This is a sample, not an industry benchmark. Five of the sixteen supplier-written documents come from a single publisher, Tyler Technologies — the largest single skew in the set. Nothing here supports a statement about the ERP services market as a whole, and the dataset does not rank vendors.

The eleven checks, by their column prefix: C1 data migration listed as a buyer responsibility with no supplier conversion hours; C2 integrations counted but not designed; C3 testing not budgeted by the implementer; C4 change management with no deliverable; C5 hypercare ending before the buyer’s first month-end close; C6 assumptions that shift risk silently to the buyer; C7 re-baselining or re-pricing rights for the supplier; C8 a post-design re-estimation right; C9 a no-fit-gap method with test-stage change orders; C10 a quoted fee that is contractually only an estimate billed on time and materials; C11 a single go-live with deferred functionality contractually dropped.

The analysis built from these rows, with the verbatim clauses and the score distribution, is the benchmark page.

The five states a check can read

Every _state cell holds one of exactly five values. Two of them look similar and are not, and we keep them apart on purpose.

Neither of the last two is ever folded into a score as though it were a pass. That is the whole reason the two are separate values rather than one.

The document_type column

Three values, and the distinction between them changes how a row should be read.

Licence. This dataset is published under the Creative Commons Attribution 4.0 International licence (CC BY 4.0). You may copy, redistribute, adapt and build on it, including commercially, with attribution.

Citation.

The two files agree with each other row for row and field for field, and both agree with the tables below. The tables show the columns that fit on a page; the files carry all seventy-nine, including the quoted clause, the section reference and the note for every check.

If a finding here does not match the underlying document, that is a correction we want to hear about, and we will publish it: donzou@psi888.com.

Check states by document

Where a check fired, the tier it fired at is shown beside the state. Only C3 has a tier that changes whether the instance is counted in the standing score: it counts at high and never at medium.

1. What the review of this dataset found. Before publication, every count on these pages was re-derived from the seventeen row-level records rather than carried over from any summary. That pass found five errors, all of them in aggregate sentences and none in the rows themselves: one document had been listed twice in the score distribution, at two different scores; three check denominators had been computed on inconsistent rules, so two fire rates were stated against the wrong base and one counted documents the check was never run against; and one counter-example count was understated by one. Every figure published here is the re-derived one. The rows passed unchanged. We record this because it is the second kind of finding in this dataset: a count taken from a summary of a document set is not the same thing as a count taken from the set.

2. Every row is re-derived, never copied from a summary. Each cell is taken from that document’s own stored scoring record. Where a count printed in a summary sentence disagreed with the row-level records, the rows won.

3. Two scores were corrected on 2 October 2026. The Village of Downers Grove contract (Tyler Technologies) and the City of Sutter Creek contract (Tyler Technologies) had both been counted at C3 HIGH on the same ambiguous Tyler data-conversion-verification sentence that had already forced the San Rafael contract to not assessable on 30 September. Downers Grove now reads C3 fired at medium, not counted in the standing score; Sutter Creek now reads C3 not-assessable. Both scores move from 5 of 7 to 4 of 7. Both were re-examined from the stored record without a new fetch, because both statement-of-work exhibits remain unreadable. The direction of the underlying testing finding is unchanged; the severity split is softer than earlier published versions. It is set out in full on the benchmark page.

4. A mapping error on check C2, found and corrected 1 October 2026. While this export was being built, the C2 check was found to point at the supporting evidence patterns belonging to two other checks rather than its own. It was corrected the same day. No other check’s mapping was re-audited in that pass, so this was a single incidental fix rather than a sweep, and we say that rather than implying a full audit.

5. Two cells once counted a check as fired with no quoted clause recorded against them. The Chandler and Orlando rows each carried a fired cell with no supporting quote. As of 2 October 2026 all four of those cells — Chandler’s C3 and C9, Orlando’s C6 and C9 — carry quoted clauses with section references. There is now no cell in the dataset that reads fired with an empty quote.

6. Some quotes are close paraphrases or summaries, not the sentence itself. Each quote reproduces the text recorded at scoring time. Where that record captured a short description of the clause rather than the sentence, the cell holds the description. Nothing was reconstructed or reworded to read more like a quotation.

7. Section references are often missing, and none has been invented. A _section cell reads “not recorded” wherever the original scoring record did not capture one. We would rather publish a gap than a plausible-looking section number.

8. One document was retrieved through a third-party rendering. The Home Interiors & Gifts contract with PricewaterhouseCoopers was read from a third party’s hosted rendering of the underlying SEC exhibit; the specific filing URL could not be located, so completeness rests on the document’s own internal consistency — a dated signature block and a sequential deliverable and fee schedule — rather than on a second-source fetch. Its sibling private-sector documents were read directly from the filings.

9. Redaction is not the same as silence. In the NuVasive contract with IBM Global Business Services, total hours, total fee and all per-role hourly rates are redacted under an SEC confidential-treatment request, and the go-live and support section is redacted in full. Those are the filer’s own excisions, not a retrieval failure, and they are recorded as not assessable rather than as effort that does not exist.

10. Scores of 2 in this set are floors. The Citizens Property Insurance row scores 2 with its governing exhibits absent; the Warnaco row scores 2 with four of the seven standing checks not assessable after four exhibits of a very large filing could not be retrieved. Read both as the least the document can be shown to do, not as a clean contract.