home

Business applications and what they let users do

The data behind the draft “Trust, but Audit: Measuring an LLM Agent Pipeline that Extracts Software Capabilities from Documentation” (RELIABLE-SE 2027) and the pub-10 dataset release. Every number in the paper can be traced to rows here.

Internal review build, 28.10.2026. Not anonymised, not licensed, not for distribution. Source texts are not shown (vendor copyright); every statement carries its verbatim excerpt, URL and the sha256 of the archived page.

Read the paper draft (PDF, 4 pages) · Download the release files (CSV/JSON + datasheet) · Prompts and briefs

25applications (12 open source, 13 commercial)
217archived documentation pages
3,819statements from two blind runs
2,021final capability statements
79.2%overlap between the two runs
97.3%acceptable in the blind audit (n = 150)

How a statement gets into the dataset

Archiverchooses and fetches pages
217 of 241 fetched
→
Extractors A and Bblind to each other
1,879 + 1,940
→
Script checkquote verbatim? form?
0 dropped
→
Verifierconfirm / correct / reject / add
1,787 duplicates removed
→
Adjudicatorrules on every non-confirmation
136 verdicts reviewed
→
Final set
2,021 statements

All steps except the script check are Claude models under fixed briefs (Sonnet 5.5 for archiving, extraction and verification; Opus 5.5 for adjudication; medium effort). Every agent call is listed in provenance.

Start here

25 random statements →

The quickest way to judge the data. Reload for another sample.

All final statements →

Filter by application, verification outcome, form, or which run found it.

Blind audit →

An independent Opus pass over 150 random rows, without pipeline verdicts.

Rows the audit rejected →

Each with the auditor's note: which check failed and why.

One app, every step →

Each candidate from both runs, with the verifier's and adjudicator's reasons.

Corrections →

Original vs final wording for the 77 reworded statements.

Rejections →

The 60 statements rejected for a reason other than duplication.

Missed capabilities →

56 proposed by the verifier, 49 accepted.

Deliberately skipped →

Marketing, testimonials, pricing, platform claims: what the extractors set aside, and why.

Run A vs run B →

Per application: found by both, by one run only, or added.

Sample screening →

576 candidate applications, each with the screener's verdict and reason.

All metrics →

Every pipeline count, recomputed from the stage files.

Pilot →

The 317 pilot statements behind Tables 1–2 of the paper.

Three examples

F2-055-S073 · MaintainX · confirmed
The system lets a user create a purchase order specifying the type and quantity of parts to order from a vendor.
“A purchase order in MaintainX specifies the type and quantity of specific parts you want to order from a vendor.” — https://help.getmaintainx.com/about-vendors
F1-017-S029 · ZenTao (now branded Sanplex) · confirmed
The system automatically calculates the total man hours of a project when the project team is set.
“After the team is set, ZenTao will automatically calculate the total man hours for the project.” — https://www.sanplex.com/book/zentaomanual/free-open-source-project-management-software-createprojects-35
F2-019-S100 · Xero · added
The system lets a user benchmark business performance against the industry.
“Benchmark against the industry” — https://www.xero.com/us/accounting-software/

The 25 applications

AppNamePCF categoryPagesStatementsBoth runsA onlyB onlyAdded
F1-001Operately1.0 Develop Vision and Strategy14102824133
F1-002It's a Plan2.0 Develop and Manage Products and Services8865712143
F1-005DeskcommCRM3.0 Market and Sell Products and Services568476114
F1-011Asp.Net Core Inventory Order Management System (Indotalent WMS)4.0 Manage Supply Chain for Physical Products33736010
F1-017ZenTao (now branded Sanplex)5.0 Deliver Services1380581651
F1-023Helper6.0 Manage Customer Service41917020
F1-029Wukong HRM7.0 Develop and Manage Human Capital41615100
F1-034WeCMDB8.0 Manage Information Technology (IT)43628422
F1-037FOSSBilling9.0 Manage Financial Resources114633760
F1-043QloApps10.0 Acquire, Construct, and Manage Assets54744300
F1-046Probo11.0 Manage Enterprise Risk, Compliance, Remediation, and Resiliency1286522734
F1-047Mayan EDMS13.0 Develop and Manage Business Capabilities6124833641
F2-001Asana1.0 Develop Vision and Strategy77363622
F2-007Userback2.0 Develop and Manage Products and Services14101721856
F2-015Pipedrive3.0 Market and Sell Products and Services579524194
F2-019Xero4.0 Manage Supply Chain for Physical Products7100755173
F2-025Salesforce Field Service5.0 Deliver Services875574113
F2-031Salesforce Service Cloud6.0 Manage Customer Service106757361
F2-037RUN Powered by ADP7.0 Develop and Manage Human Capital276100
F2-043ServiceNow IT Service Management8.0 Manage Information Technology (IT)2415311717181
F2-049Sage Intacct9.0 Manage Financial Resources12108911340
F2-055MaintainX10.0 Acquire, Construct, and Manage Assets101351081647
F2-061Vanta11.0 Manage Enterprise Risk, Compliance, Remediation, and Resiliency8101781931
F2-067Givebutter12.0 Manage External Relationships1221117410243
F2-073Success.co13.0 Develop and Manage Business Capabilities96459410

Raw access

Every table: browse the database, run your own SQL, or append .csv / .json to any table or query URL. Built by tools/business-apps/site-db.ts from data/business-apps/run/.

Powered by Datasette · Data license: Not yet licensed — internal review only · Data source: Ahmed ElBatanony, Innopolis University