27 October 2026

Private sale

Private company data, licensed by the companies that own it

Archives, code and recordings, each from one company working with Rightmark. Owners are named at introduction.

4 lots

Company records

Operating records: mail, chat, documents, tickets, CRM and outreach

Lot SE-A

Exclusive: a B2B sales-software company's own records, 8 tools, more than 10 years, up to 15,900 call hours

One company's operating record across eight tools: about 350 mailboxes, team chat from 2015 to October 2026, wikis, Jira, support conversations, the CRM and up to about 30,000 recorded sales and customer-success calls. It shows how an established B2B software company sells, supports its customers and plans its work, in private records that span more than ten years and have never been licensed.

years of records; team chat runs from 2015 to October 2026
more than10
tools from one company: mail and documents, chat, two wikis, work tracking, support, CRM, a call recorder
8
mailboxes over the years
about350
hours of recorded sales and customer-success calls, by the owner's count; may be lower
up to15,900
people over the years; 50 to 120 at a time, about 75 today
about300

What is inside

  • Snapshot: B2B sales-software company · about 12 years old · 50 to 120 staff over the years, about 75 today · incorporated in the US · most staff work outside the US.
  • Email and documents: about 350 mailboxes over the years; the owner puts email at about 99% English, mostly with customers.
  • Team chat in Slack: public channels, a 5.7 GB export (unpacked) with about 2.18 million messages in 987 channels from May 2015 to 1 October 2026; the owner expects 15 to 20% of them in English. Private channels may be added by agreement.
  • Wikis in Notion and Confluence, work tracking in Jira, customer support conversations in Intercom and the CRM in HubSpot, all kept by the same company.
  • Up to about 30,000 recorded sales and customer-success calls from the company's own teams, up to about 15,900 hours, about 32 minutes a call on those counts. The owner's figure is a bulk count and may be lower; the exact count is being measured.
  • Never licensed, sold or promised to anyone, and no vendor contract claims it. Open to an exclusive license or to non-exclusive licenses.
  • First years and sizes for each tool, and the English share of the other tools, are being gathered from the owner.

Best for

  • Training and evaluating enterprise agents that work across email, chat, wikis, tickets and the CRM of one real company.
  • Sales and customer-success agents: real B2B sales and success calls with the same company's CRM and support desk.
  • Long-horizon evaluation held out from public data: how projects, customer issues and decisions unfold over years.
  • Seed data for realistic workplace environments: chat, ticketing, wiki and CRM replicas for agent training.
  • Call understanding: summarizing sales and success calls and pulling out next steps, for recognition and understanding only.

Before delivery

  • The owner's rights and a privacy review, tool by tool, before delivery; the owner notifies staff first. Records of US residents follow the state privacy route counsel sets (California first); people who opted out of sale or sharing stay out.
  • Personal data of people outside the company (prospects, customers, people on calls) is removed or de-identified to the buyer's standard and the law's before delivery; customer data found in internal channels is screened out, and information customers shared with the company in confidence (account details, screens and files shown on calls, support conversations) is screened out of email, chat, support and calls before delivery.
  • Calls are licensed for recognition, understanding and evaluation; no voice cloning or speech synthesis, and no speaker identification. The consent position of the other party, and whether the recording notice they heard covers licensing for AI training, is checked for each call set.
  • Counts per tool and the exact call hours are measured and listed in a manifest before signing; the call price applies to the hours delivered.
  • Samples only with the owner's agreement, as metadata or masked output, for evaluation only and deleted after the agreed window.
  • Onward licensing terms are agreed with the owner, who may bar named competitors.
  • Once you accept the NDA in Rightmark's buyer portal, Rightmark shares the detailed source list, with staff locations by country. The owner's name and product names are shared only when the owner agrees to an introduction.

Left out

  • Always out: health data, data about under-18s, and payment-card and bank-account numbers.
  • Private channels and direct messages in team chat, unless the owner agrees to add private channels.
  • Data the company's customers keep in its product: a separate lot, offered only with those customers' opt-in.
  • Source code is not part of this lot.
  • Material that belongs to clients or partners. Other exclusions are being agreed with the owner.

Lot HRT-ARCHIVE

Exclusive license available: more than ten years of operating records from a B2B software company of a few hundred staff

The internal record of a B2B software company over more than ten years, to a few hundred staff today: team chat since 2015, email and documents, more than 72,000 support tickets, about 590,000 CRM records and its work tracking, all in English. It shows how a real company builds, sells and supports software across seven systems, in records that cannot be collected from the public web.

of records; team chat for nearly all of them
12 years
staff today (owner's figure)
a few hundred
team chat messages in the last 13 months alone
about4.09M
support tickets
72,000+
CRM records
about590,000

What is inside

  • Seven systems: email, documents and calendar in one workspace (about 3.5 TB of mail and files, about 100,000 calendar records), team chat (about 60 GB), a support desk (more than 72,000 tickets), a CRM (about 590,000 records), work tracking (about 5 GB), error monitoring for 12 projects (recent months only) and design files (about 5 GB).
  • Team chat runs from 2015 to 2026, with about 4.09 million messages in the last 13 months alone; the team's daily stand-up updates, collected by a bot, sit in the chat too. The owner describes its records as chat-first.
  • A few hundred staff today, at a company that builds software products for business customers.
  • Most of its staff are in India; the owner estimates about 70% of the records come from staff in India and 30% from the US. All records are in English.
  • Delivered de-identified: names and contact details are removed or masked before anything is shared. Support tickets and CRM records pass a privacy review first. The owner reports that staff were notified.
  • The company's own records only: its customers' product data and file storage, HR and payroll, a chat workspace with outside members, third-party prospect data and AI chat logs stay out.
  • The company's code is offered as its own lot (286 repositories with full git history) and can be quoted together with this archive. Links between tickets and commits have not been checked yet.
  • Figures come from the owner's data inventory. A measured manifest per system comes before any contract.

Best for

  • Workplace agents that act across email, chat, documents, calendar, CRM, support desk and work tracking, trained or evaluated on one company's real records over 12 years
  • Customer-support agents: more than 72,000 B2B support tickets with the product and engineering context around them
  • Sales and account-management agents: about 590,000 CRM records from a B2B software seller
  • Company replicas and RL environments seeded with a real software company's chat, documents and tickets
  • Held-out evaluation sets from private records that are not on the public web

Before delivery

  • A measured manifest per system (volumes, first and last years, mailbox count) before contract; figures today are the owner's counts.
  • Rights: the records of every group company in scope are covered, with the signer named; the owner's "no vendor" answer on training rights confirmed in writing.
  • De-identification as staff were told: names and contact details removed or masked; personal data in tickets and CRM removed, through the de-identification pilot the owner asked for or the owner's own tooling, then our privacy review.
  • Counsel's view on licensing India-based staff's records under India's data protection law before delivery.
  • Exclusivity scope set against the code: once any buyer licenses the code, a whole-company exclusive is off, so archive and code quotes are compared before anything is signed.

Left out

  • HR and payroll records
  • The product's customer data and file storage, and other client and end-user data
  • A chat workspace that includes outside members
  • Third-party prospect data
  • AI chat logs
  • Client contracts and NDA material
  • Payment data, privileged legal material and credentials
  • Code (its own lot; quoted together on request)
  • Names and contact details (removed or masked before delivery)

Lot MKT-ARCHIVE

Exclusive option: about 11 years of a marketing software company's records, 12 tool kinds, about 200 staff

Email, calendar, team chat, documents, work tracking, CRM, support desk, meeting recordings, knowledge base and QA logs from an operating software company of about 200 people, never licensed before. Offered with or without the same company's measured code (405 repositories, 266,826 commits).

staff (owner's form)
about200
of records, full history kept (owner's form)
11 years
tool kinds, from email to QA and test logs
12
countries the records come from
4
earlier licenses of these records (owner's form)
0

What is inside

  • Tool kinds per the owner: email, calendar, team chat, documents, work management, code, CRM, support desk, meeting recordings, knowledge base, QA and test logs, and custom systems.
  • The code host holds more than 10 years of work according to the owner; the measured history starts in 2011 and is offered as lot MKT-CODE.
  • A support desk and a knowledge base sit beside the CRM, the team chat and the meeting recordings.
  • Language: the owner writes that intra-office communication is in English; his form puts English at about half of all records. The inventory measures the share per tool.
  • Software industry, no regulated data, an operating company open to an exclusive license of the whole archive.
  • Still to be measured: sizes, start years and languages per tool come with the owner's inventory before any contract.

Best for

  • Enterprise agents that work across email, chat, tickets, CRM and documents
  • Customer-support agents trained and evaluated on the company's support desk records (masked; years and volume per the inventory)
  • Seed data for workplace RL environments across 12 tool kinds
  • Meeting understanding from the recorded meetings, subject to a consent review
  • Models of engineering and the business around it, when taken with the code lot

Before delivery

  • A measured inventory per tool (product, years, size, languages), agreed in writing before contract.
  • Records any third party claims rights over are removed or cleared before delivery.
  • Customers' own content (their campaigns, contact lists and their end users' messages) stays out unless customer terms allow it; personal data is masked, and accounts whose contracts forbid use are left out.
  • Signed by the company entity that holds the records, with written assignment from the contractors whose work they are.
  • A privacy review for each of the four countries, and staff notice where the law requires it.
  • Evaluation samples restricted to evaluation, deletion on request; onward transfer and resale terms set in the license; the owner may veto direct competitors.

Left out

  • Customers' own content unless customer terms and opt-in allow it
  • Records under a third party's training-rights claim
  • Personal data (masked) and any regulated data (none per the form)
  • Code, unless the buyer takes the whole-company exclusive (offered as MKT-CODE)

Lot RE-2

About 1.9M outbound emails with 17K labeled replies and 4,100 booked meetings, plus the company's sales scripts and playbooks (owner's count)

One US software company's outbound email record: sends, about 9.7M delivery events, labeled replies and booked meetings, plus campaign and pipeline records and about 6M words of its own SOPs, sales scripts and playbooks. A model learns how real prospects answer cold outreach, with each reply's label as the target. For scale, one open research set holds 1,857 real labeled replies, and a public reply classifier was trained on 5,000 synthetic ones; this record holds about 17,000 by the owner's count.

outbound email sends (owner's count)
About1.9M
labeled replies, from 14K+ repliers (owner's count)
About17K
booked meetings (owner's count)
About4,100
email delivery events, with campaign and pipeline records (owner's count)
About9.7M
internal SOPs, sales scripts and playbooks (owner's count)
About6M words

What is inside

  • Outbound email: about 1.9M sends, more than 14,000 repliers and about 17,000 labeled replies, by the owner's count.
  • The app database: about 9.7M email delivery events plus campaign and pipeline records, by the owner's count.
  • About 4,100 booked meetings, by the owner's count; whether each one links to the email thread that booked it is to be stated.
  • An internal knowledge base of about 6M words: SOPs, sales scripts and playbooks, by the owner's count.
  • Help-desk conversations with the company's customers are also on offer; count and years to be stated.
  • Part of the record is the company's customers' campaign data, offered as far as its customer terms allow and anonymized, after our review of those terms.
  • Prospects' names, addresses and email addresses are masked, and reply text goes through a privacy review before delivery.
  • The same company's consented sales demos are a separate lot (RE-1), so a buyer can pair outreach with live calls.

Best for

  • Classifying and routing real cold-email replies, with the record's own labels
  • Training AI SDR agents to pick the next step after a reply, with booked meetings as the outcome where linked
  • Deliverability and send-pattern models from about 9.7M delivery events
  • Grounding and evaluating sales agents on a real company's SOPs, scripts and playbooks
  • Held-out test sets of real replies for outbound-email tools

Before delivery

  • The owner states years covered, the distinct-message count after templates, the reply label set and who labeled (a person or a model).
  • Which outbound messages a model wrote is stated before delivery, so buyers that exclude model-written text can filter.
  • We and our counsel read the customer-terms clause before any customer campaign data is included.
  • Prospects' and homeowners' names, addresses and email addresses are masked; reply text passes a privacy review; US state privacy rules are checked.
  • Each booked meeting is linked to its thread where the data allows, and the share that links is stated.
  • The company's state registration is confirmed before contract.

Left out

  • Prospects' and homeowners' names, addresses, email addresses and phone numbers (masked)
  • Customer campaign data that the company's customer terms do not cover
  • Replies that fail the privacy review
  • Customers' personal details in help-desk conversations (masked)

2 lots

Audio and video

Recorded calls, interviews and demos

Lot HRT-INTERVIEWS

AI job interviews, audio and video, licensed only where the candidate opts in to AI training: a monthly stream from a live hiring product

Real job candidates answer an AI interviewer on camera. The supplier records each speaker on its own channel, with a transcript and the interview outcome, in English (about 90%) and Indian languages. Only sessions where the candidate ticks a separate, optional box naming AI training are licensed, and buyers can add questions written to their own spec.

interviews a month on the live product (supplier's figure); the opted-in share is measured once the opt-in runs
about50,000
English; the rest in Hindi and other Indian languages (supplier's figures)
about90%
in 21 sample sessions, measured from the files
5.6 hours
per sample session, 16 on average; 720p video at 30 fps
10 to 27 min

What is inside

  • Real job candidates, screened for jobs at employers in India, answer an AI interviewer in a hiring product that is in production today: about 50,000 interviews a month, by the supplier's figure.
  • Each session is audio and video with each speaker on a separate channel, plus a transcript and the interview outcome (supplier's description). The candidate is on camera; the interviewer is an AI voice.
  • Consent: a separate, optional checkbox in which the candidate agrees that the audio, video, transcript and outcome may be licensed to AI developers to train and evaluate AI models; the interview and its result do not depend on it, and it can be withdrawn at any time. The supplier says the box is unticked by default and that every file carries its consent record; Rightmark checks both against the files before any sample.
  • Only opted-in sessions are licensed. Interviews recorded before the opt-in are not offered: their consent covers recording only.
  • Questions written to a buyer's spec can be added to the interviews, so a buyer can order spoken answers to its own question set by the month.
  • Measured sample: 21 sessions, 5.6 hours in all, 10 to 27 minutes each (16 on average), H.264 video at 1280x720 and 30 fps with 2-channel AAC audio at 44.1 kHz. Samples are shared only after each file is matched to its consent record.
  • The supplier reports each withdrawal made after delivery, so the licensee can delete those sessions.

Best for

  • Speech recognition and understanding on English and Indian-language speech from job candidates in India, with transcripts
  • Voice-agent and turn-taking work: real people in a structured conversation with an AI interviewer, on separate channels
  • Evaluation sets for AI interviewing and spoken-answer assessment, with outcomes
  • Audio-visual speech: on-camera answers under an opt-in that names AI training
  • Custom collection: spoken answers to a buyer's own questions, ordered by the month

Before delivery

  • The opt-in is live in production, and every delivered session carries its consent record: time, wording version, tick state.
  • Each sample file is matched to its consent record before any sample is shared.
  • Candidate pay terms seen: the supplier says candidates who opt in are paid.
  • The terms of the vendor behind the AI interviewer's voice checked; its channel masked or excluded if they require.
  • Rights under the employers' contracts (who owns the recordings) confirmed in writing; the signing company's registry checked.
  • License for recognition, understanding and evaluation only, with a no-cloning covenant and audit rights; never to voice-cloning or avatar buyers.
  • Withdrawals passed on after delivery, with deletion by the licensee.
  • Counsel on India's consent rules for the opt-in, on any US candidates' video, and on licensing interview outcomes for model training.

Left out

  • Past interviews recorded before the opt-in (their consent covers recording only)
  • Any session without a matching AI-training opt-in record, and any withdrawn session
  • Speech generation, voice cloning, avatars and likeness synthesis

Lot RE-1

B2B sales demos with US real-estate agents under an AI-training notice from 26 October 2026 (owner's plan): about 60 to 75 hours a month at today's rate, each call with transcript, speaker labels and outcome (owner's count)

Real sales demos between a software company's closers and real-estate agents across the US. From 26 October 2026, by the owner's plan, each one is recorded under an AI-training notice given in the invite and said at the start of the call. Each call carries a machine transcript with speaker labels and its logged outcome, so a model learns how discovery and a screen-shared product walkthrough play out, and how each call ended. For scale, one public research set of real B2B sales calls holds 50 calls.

recorded sales demos, by the owner's count; 10 to 25 more hours of client calls on request
60 to 75 hours a month
per call: two-party video calls, closer and prospect
30 to 60 minutes
where the real-estate-agent prospects are, by the owner's statement
50 states
AI-training notice in the invite and said at the start of the call (owner's plan)
Every call from 26 October 2026
each call's outcome logged with the recording (owner's plan)
1 outcome per call

What is inside

  • Two-party video calls between the company's closers and real-estate-agent prospects across the US, 30 to 60 minutes each: discovery questions, then a screen-shared walkthrough of the product.
  • Each call comes with the notetaker's machine transcript and speaker labels. The audio is one mixed track per call, with no separate speaker tracks.
  • Consent is collected at the source (owner's plan, from 26 October 2026): the invite names AI training, the notice is said at the start of the call, on the recording, and the closers sign a release naming AI training. Calls where anyone declines stay out.
  • Every hour in the lot is recorded from 26 October 2026 on. Demos recorded before the notice started are never offered.
  • Each call's outcome is logged; the label set comes with the samples.
  • Screen shares are kept free of homeowner details where possible and flagged where not; a flagged part is cut, or that call is delivered as audio only.
  • Samples: a mock demo recorded the same way, then 3 to 5 demos recorded under the notice, with transcripts; the owner aims to send them in the week of 2 November 2026.
  • Licensed by the month as it is recorded: per hour delivered, batch by batch, non-exclusive unless the owner agrees to an exclusive.

Best for

  • Speech recognition and speaker diarization tests on long, two-party video-call audio with speaker-labeled transcripts
  • Training and evaluating AI sales agents and sales-coaching models on real discovery calls, with each call's outcome as the label
  • Sales-conversation understanding: outcome prediction, next-step and objection detection, call summaries and CRM notes
  • Held-out evaluation sets for meeting assistants that summarize customer calls

Before delivery

  • The AI-training notice is in every invite and said at the start of every delivered call, from 26 October 2026; the wording in use is on file.
  • Each closer's signed release naming AI training is on file.
  • Calls where anyone declines, and every demo recorded before the notice, are left out.
  • Each batch is screened against state recording laws; some prospects are in all-party-consent states.
  • The notetaker's and the video platform's terms are checked for any claim on recordings and transcripts.
  • The license limits use to recognition, understanding, turn-taking and evaluation, with a no-cloning covenant and audit rights.
  • Samples reviewed and hours measured from the recorder export before any invoice; screen shares with homeowner details cut or delivered as audio only.
  • The company's state registration is confirmed before contract.

Left out

  • Every demo recorded before 26 October 2026 (no AI-training notice): never offered, samples included
  • Any call where a participant declines
  • Screen-share segments that show homeowners' details (cut, or the call goes as audio only)
  • Voice cloning, voice synthesis and likeness uses
  • The company's outbound phone calls by its setters (not offered now; two-sided calls may follow later by the owner's plan)

For data owners

Own data like this? List it with Rightmark.

Start with an indicative appraisal: no files or system access needed. It is free, and nothing is licensed until you sign.

Appraise your data

Or start from company records or source code. Know an owner? Become a Scout.

11 lots

Code

Private codebases with their full history

Lot SE-C1

A B2B sales-software company's in-house code: 153 repositories, 4.75 million lines, 122,367 commits over 12 years

The product code of an established B2B sales-software company: a main application with 12 years of history (75,186 commits, 38,849 merged pull requests, 193 contributors) and 152 services, apps, tools and AI components around it. C#, TypeScript, JavaScript and Python with git history, never licensed before by the owner's form, with AI use measured and disclosed per repository.

lines of code in 153 repositories
4,755,994
commits over about 12 years, to October 2026
122,367
merged pull requests
60,640
of history on the main application, 193 contributors
12 years
of commits by an AI-agent identity (2 repositories)
1.8%

What is inside

  • Owner: A B2B sales-software company, about 12 years old, 50 to 120 staff over the years and about 75 today, incorporated in the US with most staff outside the US. By the owner's form, the code has never been licensed, sold or promised elsewhere.
  • Main application: 2,888,550 lines (C# 58.2%, TypeScript 27.9%, JavaScript 11.3%), 75,186 commits and 38,849 merged pull requests over 12 years and 147 active months, 193 contributors, 737,086 lines of tests in 5,394 test files.
  • Around it: 152 more repositories with 1,867,444 lines and 47,181 commits, among them 23 repositories named for AI and LLM work (590,991 lines).
  • Languages over all 4,755,994 lines: C# 49.6%, TypeScript 22.2%, JavaScript 13.0%, Python 9.7%, Swift 1.7%, SQL 1.5%.
  • Engineering practice: CI in 125 of 153 repositories (GitHub Actions in 122), test frameworks in 116, Dockerfiles in 114; 1,398,033 lines of tests in 9,695 test files; 15,224 bug-fix commits, 2,625 of them changed tests in the same commit.
  • Spread over time: 11 repositories started in 2018 or earlier, 101 between 2022 and 2024 and 21 in 2026 (8 of them between August and October); 54 had commits in the 90 days before the check; 60 have 100 or more merged pull requests; contributors per repository up to 193 (median 7).
  • AI use, stated plainly: 2,201 of 122,367 commits (1.8%) come from an AI-agent identity, in 2 repositories, one of them mostly agent-written; 26 repositories carry AI-assistant settings files (189 files); automation accounts made at least 1,109 listed commits (0.9%). 51.9% of added lines date from 2023 on, a date measure that cannot show who or what wrote them. 127 repositories (692,299 lines) carry no AI settings and no agent commits, and 9,284,273 of the main application's added lines date from before 2023.
  • Measured by the owner's repository check on full clones: English in 87.4% of 118,991 commit subjects checked; git history complete in 152 repositories, 1 squashed.

Best for

  • Training on in-house enterprise C# and TypeScript with history: 49.6% of the lines are C#.
  • Code review and pull-request models: 60,640 merged pull requests with their commits.
  • Bug-fix and test-writing data: 15,224 bug-fix commits, 2,625 that changed tests in the same commit.
  • Repository-scale context: one application of 2,888,550 lines built over 12 years by 193 contributors, beside 152 services and tools.
  • Software-evolution research: 147 active months of history on one product.

Before delivery

  • Licensed once the owner's IP pack is in: the employment and contractor agreements that assign the code to the company, reviewed by counsel before signing.
  • Every live credential found in the code is rotated before any copy moves, and the owner confirms the rotation in writing by kind and date.
  • The owner confirms in writing which repositories, if any, were ever public; any that were leave the lot or move to the partly-public rate.
  • Personal data in code (2,632 email addresses in 42 repositories) is masked, and data files committed to repositories are removed by path, before delivery.
  • Delivered as git history per repository after these steps; the line count is final after a re-measure without third-party folders and committed data.

Left out

  • Repositories that did not clear the provenance review, including public projects, third-party installs and client work.
  • Second copies of a repository's history: each history is priced once.
  • Third-party files inside the repositories (copyright notices in 9 repositories, 1 copyleft file), removed by path.
  • Customer configurations named for outside companies, and personal data in code (masked); every live key is rotated first.

Lot SE-C2

Code rated strong for task building: 17 repositories with 13,164 bug-fix commits, 2,379 of them with their tests, from one B2B software company

Seventeen repositories the owner's repository check rates strong for building coding tasks: 6 pass all five of its checks and 11 pass four. Their 2,379 bug fixes that changed tests in the same commit are the raw material for verified tasks in C#, TypeScript, JavaScript and Python, from code the owner states has never been licensed.

repositories rated strong for building coding tasks
17
bug-fix commits; 2,379 changed tests in the same commit
13,164
lines of tests in 7,869 test files
1,118,963
run CI on GitHub Actions
16 of 17
have no AI-agent commits
15 of 17

What is inside

  • Owner: A B2B sales-software company, about 12 years old, 50 to 120 staff over the years and about 75 today, incorporated in the US with most staff outside the US.
  • The 17: the main application (2,888,550 lines) and 16 more from 11,122 to 342,489 lines, 3,749,326 lines in all: C# 49.4%, TypeScript 24.5%, JavaScript 13.6%, Python 9.9%.
  • Task readiness by the tool's checks: all 17 rated strong; 6 pass all five checks and 11 pass four (the misses: history depth in 7, CI in 1, fixes with tests in 3).
  • Raw material for tasks: 13,164 bug-fix commits, 2,379 of which changed tests in the same commit (1,505 in the main application, 874 in the other 16).
  • History: 94,069 commits and 45,588 merged pull requests; 2 to 193 contributors per repository (median 13); 15 of 17 had commits in the 90 days before the check.
  • Build and test: CI in 16, test frameworks in 17, Dockerfiles in 15; 3 need private packages to build.
  • AI use, stated plainly: 2 of the 17 carry commits by an AI-agent identity (2,201 commits; 1 repository's history is mostly agent commits); 11 carry AI-assistant settings files; 51.5% of added lines date from 2023 on. Without those 2, the subset is 15 repositories, 3,300,390 lines and 1,801 fixes with tests; 6 repositories (186,520 lines) carry no AI settings and no agent commits.
  • English in 87.1% of 91,167 commit subjects checked (lowest repository 74.4%).

Best for

  • SWE-style task sets in C#, TypeScript, JavaScript and Python; SWE-bench draws its 2,294 problems from 12 Python repositories.
  • RL environments for bug-fixing agents, built from fixes that changed tests in the same commit.
  • Data for writing tests and repairing CI: 1,118,963 lines of tests beside the code they test.
  • Evaluation sets from code that, by the owner's form, has not been licensed before.

Before delivery

  • Licensed once the owner's IP pack is in: the employment and contractor agreements that assign the code to the company, reviewed by counsel before signing.
  • Every live credential found in the code is rotated before any copy moves, and the owner confirms the rotation in writing by kind and date.
  • The owner confirms in writing which repositories, if any, were ever public; any that were leave the lot or move to the partly-public rate.
  • Personal data in code (2,632 email addresses in 42 repositories) is masked, and data files committed to repositories are removed by path, before delivery.
  • Each task is priced once it passes the clean-container check; the owner's CI logs and a task-builder pass come first.

Left out

  • Repositories that did not clear the provenance review, including public projects, third-party installs and client work.
  • Second copies of a repository's history: each history is priced once.
  • Third-party files inside the repositories (copyright notices in 9 repositories, 1 copyleft file), removed by path.
  • Customer configurations named for outside companies, and personal data in code (masked); every live key is rotated first.
  • The other 136 repositories of the code lot, offered as the full set (SE-C1).

Lot SE-C3

One B2B software company's code with its engineering record: 153 repositories, issue tracker, engineering chat and support-desk bug reports

The full code lot offered with the records its engineers worked in: the issue tracker (Jira), the engineering channels of a team chat that runs from 2015 to October 2026, and bug reports from the support desk. Where commits and pull requests cite issue keys and issues link the support conversation, a buyer can follow a customer's report to an issue, a fix and a merged pull request; the share that links is measured before contract.

lines of code in 153 repositories, about 12 years to October 2026
4,755,994
merged pull requests; the share citing an issue key is measured before contract
60,640
bug-fix commits
15,224
team chat history; engineering channels selected for this lot
2015 to October 2026
of record in one lot: code, issue tracker, team chat, support desk
4 kinds

What is inside

  • Owner: A B2B sales-software company, about 12 years old, 50 to 120 staff over the years and about 75 today, incorporated in the US with most staff outside the US. By the owner's form, never licensed before.
  • Code: the 153 repositories of the full code lot (SE-C1), with 122,367 commits and 60,640 merged pull requests; the main application alone has 75,186 commits over 12 years.
  • Issue tracker: Jira, the company's tracker (to be confirmed).
  • Team chat: an export of 5.7 GB unpacked with history from 2015 to October 1, 2026; the engineering channels are selected for this lot (to be confirmed).
  • Support desk: the conversations that report bugs (to be confirmed).
  • Linkage: commits and pull requests join to issues by the issue keys they cite, and bug reports join where an issue links the conversation (to be confirmed).
  • AI use in the code, as in SE-C1: 1.8% of commits come from an AI-agent identity and 51.9% of added lines date from 2023 on; 127 repositories carry no AI settings and no agent commits.

Best for

  • Agents that work across engineering tools: from a reported bug to an issue, a fix and a merged pull request.
  • Bug triage and root-cause models trained on reports joined to the fixes that closed them.
  • Engineering-workflow environments that mirror a working software company's tracker, chat and repositories.
  • Process data on how one product team planned, discussed and shipped over more than ten years.

Before delivery

  • Licensed once the owner's IP pack is in: the employment and contractor agreements that assign the code to the company, reviewed by counsel before signing.
  • Every live credential found in the code is rotated before any copy moves, and the owner confirms the rotation in writing by kind and date.
  • The owner confirms in writing which repositories, if any, were ever public; any that were leave the lot or move to the partly-public rate.
  • Personal data in code (2,632 email addresses in 42 repositories) is masked, and data files committed to repositories are removed by path, before delivery.
  • Records are de-identified to the buyer's standard and the law's before delivery; customer and prospect data inside chat and support conversations is removed.
  • The owner notifies staff before team chat is shared.

Left out

  • Repositories that did not clear the provenance review, including public projects, third-party installs and client work.
  • Second copies of a repository's history: each history is priced once.
  • Third-party files inside the repositories (copyright notices in 9 repositories, 1 copyleft file), removed by path.
  • Customer configurations named for outside companies, and personal data in code (masked); every live key is rotated first.
  • Private channels and direct messages in team chat, and channels outside engineering.
  • Support conversations other than bug reports, and data the company's customers keep in its product.

Lot 02-B

Task-grade code: 14 repositories with CI, test suites and full git history

JavaScript and TypeScript code from one B2B software company, with CI configured in all 14 repositories, more than 6,200 test files and 233 fixes that changed a test in the same commit. Each repository comes with its tests, its CI configuration and its full history: more than 44,000 commits and more than 31,000 merged pull requests in all.

Repositories, CI configured in all 14
14
Original lines with full git history
3.4M+
test files
6,200+
Merged pull requests, counted from merge commits
31,000+
Fix commits that changed a test in the same commit
233

What is inside

  • Full git history in every repository: more than 44,000 commits, first commits 2019 to 2026, none squashed.
  • The largest has more than 2 million lines, more than 17,000 commits, more than 10,000 merged pull requests and 155 people in its history, first commit 2020. 11 of the 14 have more than 500 merged pull requests, the fewest has 54, and 13 of the 14 have 10 or more people in their history.
  • CI: GitHub Actions in all 14, Azure Pipelines in 8, AWS CodeBuild in 5. Frameworks found for the tests: Mocha in 5, Vitest in 4, Jest in 3, Playwright in 1.
  • More than 1.9 million of the lines are in test files, about two thirds of them in the largest repository.
  • The owner reports a test run for each of the 14 with 0 failing and more than 96,000 passing tests in all; three were reported as CI pull-request runs. Rightmark has not seen the run logs.
  • More than 5,600 commits are marked as fixes; 233 of them changed a test file and code in the same commit.
  • Languages by lines: JavaScript 77.7%, TypeScript 22.3%.
  • Disclosed: 72% of commits date from 2023 on, 7 of the 14 commit AI coding-assistant configuration files, and one was built for a client of the owner.

Best for

  • SWE-bench-style environments: a repository at a commit, a failing test and a grader that runs the suite
  • Models that generate and repair tests, on real suites in Mocha, Vitest, Jest and Playwright
  • CI-repair and build agents across GitHub Actions, Azure Pipelines and AWS CodeBuild
  • Agent evaluation on a codebase of more than 2 million lines with 155 people in its history
  • Repository-level code models with full commit history

Before delivery

  • The owner reports that every credential found in the code or its history has been rotated. Credential strings still in the code or its history are removed or shown to be dead before delivery.
  • Any repository found to have been public is left out.
  • The owner's rights are reviewed lot by lot.
  • Once you accept the NDA in Rightmark's buyer portal, Rightmark shares repository-level metrics (names withheld) and the detailed source list. Code samples come only with the owner's agreement.
  • The owner's name, product names and repository names are shared only when the owner agrees to an introduction.
  • One of the 14 was built for a client of the owner. The owner says it may license it; Rightmark has not seen the client contract.
  • 5 of the 14 use private package registries; those packages ship with the code so the repositories can build in a clean container.
  • An exclusive license needs the owner's written OK and takes these 14 out of lot A's non-exclusive sales for its term.

Left out

  • Repositories with an open-source license file (26 in all) and any repository found to have been public
  • Vendored code (more than 110,000 lines in these 14), not counted
  • Customer data in the product's databases and file storage: no code lot carries it

Lot 02-A

A software company's code estate: 286 repositories, 2016 to 2026, with full git history

One B2B software company's git repositories: more than 5 million original lines, more than 88,000 commits and more than 60,000 merged pull requests from 2016 to 2026, with every repository's history intact. 61% of the commits predate 2023, and 28 repositories commit AI coding-assistant files, so the same company's code spans the years before and after AI assistants.

Repositories
286
Original lines
5.0M+
Commits
88,000+
Merged pull requests, counted from merge commits
60,000+
First commits
2016 to 2026

What is inside

  • Full git history in every repository; none squashed.
  • 61% of commits predate 2023, and 229 repositories have every commit before 2023.
  • Languages by lines: JavaScript 76.5%, TypeScript 20.7%, Python 2.2%.
  • The largest has more than 2 million lines, more than 17,000 commits, more than 10,000 merged pull requests and 155 people; 16 repositories are over 50,000 lines and 156 are under 1,000.
  • CI is configured in 94 repositories (GitHub Actions in 51, AWS CodeBuild in 36, Azure Pipelines in 27, CircleCI in 8), with more than 7,300 test files; Docker files in 120.
  • More than 11,000 commits are marked as fixes; 293 changed a test file and code in the same commit.
  • Disclosed: AI coding-assistant configuration files appear in 28 repositories, and 2 were built for clients of the owner.

Best for

  • Continued pre-training on one company's code from 2016 to 2026, with full history
  • Repository-level and cross-repository understanding across one company's code
  • Commit-message and code-change models over more than 88,000 commits
  • Studying how one company's code changed after AI coding assistants arrived
  • Mining SWE tasks: 293 fix commits that changed a test, more than 11,000 marked as fixes

Before delivery

  • The owner reports that every credential found in the code or its history has been rotated. Credential strings still in the code or its history are removed or shown to be dead before delivery.
  • Any repository found to have been public is left out.
  • The owner's rights are reviewed lot by lot.
  • Once you accept the NDA in Rightmark's buyer portal, Rightmark shares repository-level metrics (names withheld) and the detailed source list. Code samples come only with the owner's agreement.
  • The owner's name, product names and repository names are shared only when the owner agrees to an introduction.
  • Two repositories, one in lot B and one in lot D, were built for clients of the owner. The owner says it may license them; Rightmark has not seen the client contracts.
  • Two repositories with copyleft license headers stay out until the rights review clears them.
  • An exclusive license needs the owner's written OK and cannot overlap a whole-company exclusive that includes the code.

Left out

  • 26 repositories with an open-source license file, and any repository found to have been public
  • 2 repositories with copyleft license headers (more than 660,000 lines), until the rights review
  • Vendored code (more than 2.4 million lines), not counted
  • Customer data in the product's databases and file storage

Lot 02-C

Code from before 2023: 19 repositories with every commit made between 2016 and 2022

Every commit in these 19 repositories predates 2023, and none commits AI coding-assistant files. More than 34,000 commits and more than 23,000 merged pull requests from one B2B software company, with up to 41 people in one repository's history.

Repositories, each over 10,000 lines and 100 commits
19
Original lines
890,000+
Commits
34,000+
Merged pull requests, counted from merge commits
23,000+
First commits
2016 to 2022

What is inside

  • Every commit in every one of these repositories predates 2023. Each has over 10,000 lines and at least 100 commits.
  • All JavaScript, by each repository's main language.
  • 15 of the 19 have 10 or more people in their history, and 9 have more than 1,000 merged pull requests.
  • The largest has more than 100,000 lines, more than 3,500 commits and 28 people in its history.
  • CI is configured in 18 (AWS CodeBuild in 16, GitHub Actions in 5), with 407 test files; Docker files in 12.
  • More than 4,800 commits are marked as fixes; 26 changed a test file and code in the same commit.
  • 3 of the 19 are also in lot B. Another 210 repositories in lot A also have every commit before 2023; they are smaller or have fewer than 100 commits.

Best for

  • Pre-training and evaluation sets of code written before 2023, with no AI coding-assistant configuration files
  • A baseline for studies of AI-assisted code, against the same company's 43 repositories with every commit from 2023 on (lot A)
  • JavaScript repositories with pull-request history for code-change models

Before delivery

  • The owner reports that every credential found in the code or its history has been rotated. Credential strings still in the code or its history are removed or shown to be dead before delivery.
  • Any repository found to have been public is left out.
  • The owner's rights are reviewed lot by lot.
  • Once you accept the NDA in Rightmark's buyer portal, Rightmark shares repository-level metrics (names withheld) and the detailed source list. Code samples come only with the owner's agreement.
  • The owner's name, product names and repository names are shared only when the owner agrees to an introduction.

Left out

  • Repositories with any commit from 2023 on (they are in lots A, B and D)
  • Vendored code (more than 12,000 lines in these 19), not counted

Lot 02-D

16 large codebases, each over 50,000 lines: nearly 4 million lines with CI in all 16

16 codebases of one B2B software company, each over 50,000 lines and with full history: more than 54,000 commits, more than 40,000 merged pull requests, CI configured in every one and more than 6,000 test files. Sized for agents that must navigate and change large, multi-file code.

Codebases, each over 50,000 lines
16
Original lines
3.9M+
Commits
54,000+
Merged pull requests, counted from merge commits
40,000+
First commits
2017 to 2026

What is inside

  • Each is over 50,000 lines; the largest has more than 2 million lines.
  • 7 have every commit before 2023, and 5 were started in 2023 or later.
  • CI configured in all 16, with more than 6,000 test files; Docker files in 9.
  • Languages by lines: JavaScript 78.9%, TypeScript 21.1%.
  • More than 6,900 commits are marked as fixes; 210 changed a test file and code in the same commit.
  • 7 are also in lot B and 7 in lot C.
  • Disclosed: AI coding-assistant configuration files appear in 7, and one was built for a client of the owner.

Best for

  • Long-context coding agents that navigate and edit large repositories
  • Codebase programs that require 50,000+ lines and a container build
  • Multi-file refactoring and migration tasks, with existing tests to check them

Before delivery

  • The owner reports that every credential found in the code or its history has been rotated. Credential strings still in the code or its history are removed or shown to be dead before delivery.
  • Any repository found to have been public is left out.
  • The owner's rights are reviewed lot by lot.
  • Once you accept the NDA in Rightmark's buyer portal, Rightmark shares repository-level metrics (names withheld) and the detailed source list. Code samples come only with the owner's agreement.
  • The owner's name, product names and repository names are shared only when the owner agrees to an introduction.
  • One of the 16 was built for a client of the owner. The owner says it may license it; Rightmark has not seen the client contract.
  • 4 of the 16 use private package registries; those packages ship with the code.

Left out

  • Vendored code (more than 120,000 lines in these 16), not counted
  • Customer data in the product's databases and file storage

Lot 02-E

Code with its context: 30 repositories plus the same company's support desk, engineering chat, work tracking and error monitoring

One B2B software company's code from 2016 to 2026 together with the systems around it: more than 72,000 support tickets, the engineering channels of its team chat, its work-tracking records and recent error monitoring. Agents built for software teams can learn from the code and from the tickets, discussions and work items around it, all from the same company.

Repositories (lots B and C)
30
Original lines
4.2M+
Support tickets
about70,000
Merged pull requests, counted from merge commits
51,000+
Systems beside the code: support desk, engineering chat, work tracking, error monitoring
4

What is inside

  • Code: the 30 repositories of lots B and C, more than 73,000 commits, more than 9,900 commits marked as fixes and 247 that changed a test, first commits 2016 to 2026.
  • Support desk: about 70,000 tickets, by the owner's inventory.
  • Team chat: the engineering channels of a workspace of about 60 GB, with about 4.09 million messages in the last 13 months alone; chat history from 2015.
  • Work tracking: a few GB of records; 8 of the 14 task-grade repositories carry pipeline configuration for the same platform.
  • Error monitoring: recent errors only, more than 10,000 in the last 30 days across 12 projects; the owner expects about the last 30 to 90 days are kept. Event payloads, user context, IP addresses, emails and request bodies stay out.
  • Links between tickets, work items and commits have not been counted yet.
  • Delivery only after a rights and privacy review that covers personal data in the tickets, chat and work-tracking records.

Best for

  • Agents that work across an engineering team's tools: ticket, chat, work item, pull request, CI and errors
  • SWE tasks with real problem statements, once a fix is traced to its ticket or work item
  • Support-to-engineering triage and escalation models
  • Error triage with the code and history behind it
  • Evaluating long-horizon engineering work inside one real company's systems

Before delivery

  • The owner reports that every credential found in the code or its history has been rotated. Credential strings still in the code or its history are removed or shown to be dead before delivery.
  • Any repository found to have been public is left out.
  • The owner's rights are reviewed lot by lot.
  • Once you accept the NDA in Rightmark's buyer portal, Rightmark shares repository-level metrics (names withheld) and the detailed source list. Code samples come only with the owner's agreement.
  • The owner's name, product names and repository names are shared only when the owner agrees to an introduction.
  • A rights and privacy review before delivery: personal data in the tickets, chat and work-tracking records is removed before anything is delivered.
  • The engineering channel list is agreed with the owner before delivery; other channels stay in lot F.
  • Lot E's support desk, chat, work-tracking and error data are also part of lot F. If lot F is licensed exclusively, they are carved out before it signs, or lot E waits for that term to end.

Left out

  • HR and payroll, the product's customer data and file storage, a chat workspace that includes outside members, third-party prospect data and AI chat logs
  • Error events' payloads, user context, IP addresses, emails and request bodies
  • Chat channels outside engineering (they are in lot F)

Lot MKT-CODE

A marketing software platform's code and change history since 2011: 405 repositories, 266,826 commits

Its largest repository holds 92,019 commits and 21,559 merged pull or merge requests from 126 contributors over nearly 15 years, and 91% of its added lines date from before 2023. With the 404 repositories around it, the lot carries 266,826 commits and 53,486 fix commits: the record of how a working product was built, reviewed and repaired.

commit history of the largest repository: 14.9 years, with commits in 180 different months
2011 to 2026
original lines of code in 405 repositories (PHP 58%, JavaScript 17%, TypeScript 17%)
7,842,097
commits, with an estimated 75,210 merged pull or merge requests counted from merge commit messages
266,826
fix commits (the subject opens with fix, bug, patch, revert or a similar word); 1,404 changed tests along with the code
53,486
estimated tokens of change history, plus about 104 million tokens of source code
1.89B

What is inside

  • The largest repository: 1,141,718 original lines (54% JavaScript, 45% PHP), 92,019 commits, 21,559 merged pull or merge requests, 126 human contributors, 17,387 fix commits, CI configuration and a PHPUnit suite (217 test files).
  • 101 repositories of 10,000 lines or more hold 7,219,329 lines (92% of the code); 83 of them also have 3 or more contributors, and 25 have 50,000 lines or more.
  • Tests in 179 repositories: 8,586 test files and 923,593 lines of tests; 61 repositories carry CI configuration.
  • 163 repositories (2,185,898 lines) have 10% or less of their added lines from 2023 on, and 33 have 7 or more years of history.
  • 147 repositories had commits in the 90 days before measurement; each repository was measured with its own report on 21 October 2026.
  • Git history is kept in full in 403 of the 405 repositories; the 2 with squashed history are flagged. The owner's form records no earlier license of these records.
  • Set aside before pricing: 15 repositories with open-source license files, 5 that look like imported third-party code and 9 duplicate uploads. The 405 are the cleared candidates among 425 distinct repositories measured.

Best for

  • Pretraining and mid-training on licensed product code with its full change history
  • Code-review and commit-message models from 75,210 merged pull or merge requests
  • Bug localization and repair from 53,486 fix commits in PHP, JavaScript and TypeScript
  • Long-horizon repository understanding: one 1,141,718-line repository followed across nearly 15 years
  • Held-out evaluation sets from product code kept out of public crawls (after the never-public check)

Before delivery

  • Licensed by the company entity that holds the code, with written IP assignments from staff and contractors covering the whole history back to 2011.
  • Any repository that is or was public, or was published as a package, is removed before delivery after a check against public code hosts and package registries.
  • Possible secrets found in the current code are rotated by the owner before delivery; keys are never tested.
  • Files carrying copyleft license headers are removed file by file; email addresses in code are masked.
  • A per-repository disclosure ships with the lot: the share of lines added from 2023 on, and where AI-assistant settings files are committed.
  • Non-exclusive per sale by default; an exclusive license covers the whole cleared set for a stated term.
  • Evaluation samples are restricted to evaluation: no training on them, no retention past the agreed window, deletion on request. Onward transfer and resale terms are set in the license, and the owner may veto direct competitors.

Left out

  • 15 repositories with open-source license files (8 MIT, 5 GPL, 2 Apache)
  • 5 repositories that look like imported third-party code (2,693,131 lines, 11 to 428 commits) until the owner confirms whose code they are
  • 9 duplicate uploads
  • Any repository found to be or have been public
  • Files with copyleft headers, secrets and personal data

Lot RE-3

About 480 AI coding-agent sessions from a working product team, with repositories holding about 3,800 commits and 1,700 pull requests (owner's count)

Sessions in which a software company's team used AI coding agents on its own product, offered with the company's repositories. For contrast, a 258,000-trajectory open release from February 2026 was generated by one model on public-repository tasks; these sessions come from the team's own product work, by the owner's account. Priced on review once the tools, years and session contents are stated.

AI coding-agent sessions (owner's count)
About480
commits (owner's count)
About3,800
pull requests (owner's count)
About1,700

What is inside

  • About 480 AI coding-agent sessions from the company's engineering work, by the owner's count; which agent tools, how far back, and whether each session holds its diffs and test runs are to be stated.
  • Repositories with about 3,800 commits and 1,700 pull requests, by the owner's count; each repository goes through Repo Check on the owner's machine before pricing.
  • Code written by AI agents is labeled as such before any sample, so buyers that exclude AI-written code can filter it.
  • The same company's outbound email record (RE-2) and consented sales demos (RE-1) are separate lots.
  • The owner states all of the engineering data is the company's own.

Best for

  • Coding-agent trajectory training and analysis on real product work, where sessions carry diffs and test runs
  • Repository-level code tasks built from real commits and pull requests, after Repo Check
  • Evaluation of coding agents against how a working team actually used them

Before delivery

  • Repo Check run on every repository, with secrets at HEAD and in history rotated or removed.
  • Ownership and IP assignments from everyone who wrote the code are in hand; public or third-party code is left out.
  • The agent tools' terms are checked for any claim on session logs.
  • Sessions and repositories are labeled for AI-written code before any sample.
  • The company's state registration is confirmed before contract.

Left out

  • Secrets, credentials and customer data found in code or sessions
  • Public, open-source or third-party code
  • Any repository the company does not own outright

Lot VOICE-CODE

Code from an AI voice-agent product, 2025 to 2026: 8 private repositories with full history

One company's own product code for AI voice and chat assistants, written from 2025 on: the agent engine, its platform services, session management, an admin portal and the customer apps. More than 300,000 original lines, 2,447 commits and 2,072 merged pull requests, with 8 to 17 people in each repository's history.

Private repositories, each passing all four repository gates
8
Original lines
300,000+
Commits
2,447
Merged pull requests
2,072
First and last commits
2025 to 2026
People in each repository's history
8 to 17

What is inside

  • The agent engine, in Python: more than 44,000 lines, 539 commits and 284 merged pull requests, with 16 people in its history.
  • A platform-services monorepo, in TypeScript: more than 58,000 lines, 425 commits, 172 merged pull requests and 17 people.
  • Six more TypeScript repositories: a voice and web applications monorepo, an assistant session-management backend, an embedded assistant front end, an admin portal for assistant configuration, a chatbot data-collection service and a generative-analytics customer portal.
  • Test files in 6 of the 8. Some tests were written in September and October 2026 (72 in the agent engine, 25 in the assistant front end), and their dates come with the lot.
  • Measured by Repo Check on 19 October 2026, repository by repository.

Best for

  • Coding-agent training and evaluation on a current AI-assistant product: an agent engine, its services and its apps
  • Graded training tasks from the agent engine and the platform services, once their tests run offline
  • Code-change models: more than 2,000 merged pull requests from 2025 and 2026

Before delivery

  • Each contributor's signed IP assignment is reviewed before signing.
  • One repository began as a copy of public projects; that code was removed and its history rebuilt in October 2026, and it comes with that disclosure.
  • Credentials found in the histories were rotated, the owner reports (20 October 2026); the old values remain in the history.
  • A customer's name, which the owner says is not confidential, stays in older commits of two repositories.
  • Task prices apply once the tests of the agent engine and the platform services run green offline.
  • Once you accept the NDA in Rightmark's buyer portal, Rightmark shares repository-level metrics (names withheld) and the detailed source list. Code samples come only with the owner's agreement.

Left out

  • The company's other repositories, which the owner kept out of this lot.

2 lots

Training tasks

Graded tasks built from real fixes and their tests

Lot 02-B-TASKS

SWE tasks from 233 real fixes that changed a test, with an evaluation split held back for one buyer

Each candidate is a real repository at a real commit, the change a developer made to fix it and the test that changed with it: the shape of SWE-bench-style tasks, where a grader runs the tests. Tasks are checked in a clean container before delivery, and a split can be kept out of every training license as one buyer's private evaluation set.

Candidate tasks: fix commits that changed a test
233
Source repositories, CI configured in all 14
14
Commits marked as fixes in the same repositories
5,600+
Candidates in the largest repository (more than 2 million lines, 155 people)
131

What is inside

  • A candidate is a commit whose subject marks a fix (fix, bug, patch, hotfix, revert, resolve, correct, regression) and that changed both a test file and code.
  • A graded task is the repository at the commit before the fix, a task statement, the developer's change as the reference solution and the tests that grade it.
  • A task is delivered only if its tests fail before the fix and pass after it in a clean container.
  • Spread: 131 candidates in the largest repository, 22 in a TypeScript repository of more than 480,000 lines, 19 in a 79,000-line JavaScript repository; 2 of the 14 have none.
  • The owner reports 0 failing tests in the latest run of each suite; Rightmark has not seen the logs.
  • Disclosed: 204 of the 233 sit in repositories that commit AI coding-assistant configuration files, and 12 sit in the repository built for a client, held until the client consents.

Best for

  • RL training for coding agents on fail-to-pass bug-fix tasks
  • A held-out evaluation set on code that is not in public training data, after the never-public check
  • Grader and reward-model work with the developer's own fix as the reference
  • Comparing agent patches with the change a developer actually made, and its tests

Before delivery

  • The owner reports that every credential found in the code or its history has been rotated. Credential strings still in the code or its history are removed or shown to be dead before delivery.
  • Any repository found to have been public is left out.
  • The owner's rights are reviewed lot by lot.
  • Once you accept the NDA in Rightmark's buyer portal, Rightmark shares repository-level metrics (names withheld) and the detailed source list. Code samples come only with the owner's agreement.
  • The owner's name, product names and repository names are shared only when the owner agrees to an introduction.
  • Each task passes the clean-container check before delivery: its tests fail before the fix and pass after it.
  • The evaluation split is chosen with the buyer, delivered to that buyer only and kept out of every training license of these tasks.
  • Tasks from the client-built repository wait for the client's consent.
  • 5 of the 14 source repositories use private package registries; those packages ship with the tasks.

Left out

  • Candidates that fail the clean-container check
  • 12 candidates in the client-built repository, until the client consents
  • The full git history of the 14 repositories (that is lot B)

Lot MKT-TASKS

24 test-backed repositories with 1,052 fix-plus-test commits, for graded coding tasks

Every repository here carries its own test suite (5,254 test files and 613,518 lines of tests over 2,158,360 lines of mostly PHP and TypeScript), 18 have CI configuration, and 1,052 past fixes changed tests along with the code. Each such fix is a candidate task that the project's own tests can grade.

repositories with their own test suites (16 on PHPUnit, 18 with CI configuration)
24
test files (613,518 lines of tests)
5,254
fix commits that changed tests with the code, of 13,001 fixes
1,052
original lines (58.0% PHP, 37.7% TypeScript)
2,158,360
tests passed when one of the 24 was run offline (owner-reported; 38 failed)
2,538

What is inside

  • Main languages: 15 PHP, 7 TypeScript and 2 Python repositories. Frameworks found for the tests: PHPUnit (16 repositories), Jest (4), Vitest (4), Playwright (3) and pytest (2).
  • 15,351 merged pull or merge requests and 48,683 commits; 22 of the 24 have 10,000 lines or more, 21 have 10,000 lines or more and 3 or more contributors, and 11 have 50,000 lines or more.
  • Each repository passed a task-readiness screen: a test suite of at least 5 test files, with test lines at 10% or more of code lines, on a mainstream stack, plus at least two of CI configuration, 10 or more fixes that changed tests, and 2 or more active years with 3 or more core contributors.
  • Disclosed per repository: 18 of the 24 have half or more of their added lines from 2023 on, and 21 commit AI-assistant settings files.
  • Offline runs so far, reported by the owner on 24 October, cover 2 of the 24: a 67,830-line TypeScript repository passed 2,538 Jest tests and failed 38, run on public dependencies (7 of 323 suites failed to load), and a 6,589-line PHP repository passed 9 of 12 with platform requirements ignored, so that one is indicative.
  • 21 of the 24 had commits in the 90 days before measurement (21 October 2026).

Best for

  • SWE-bench-style task building: the issue, the failing test, the fix
  • RL environments for coding agents graded by each project's own suite
  • Regression-test generation from fixes that changed tests with the code
  • Agent evaluation on PHP and TypeScript product services

Before delivery

  • Priced per repository until each suite runs green offline; sold per graded task after each task passes a clean-container check (the reference fix passes, doing nothing fails).
  • The same title, never-public, secrets, copyleft and masking conditions as the code lot.
  • A per-repository disclosure of when the code was written and where AI-assistant settings are committed.
  • Evaluation samples restricted to evaluation, deletion on request.

Left out

  • Any repository whose suite does not run
  • The 2 test-backed repositories under 10,000 lines (offered with the code lot)
  • Everything held out of the code lot

2 lots

Bundles

Lots from more than one company, licensed in one order

Lot BND-2-CODE

Code from two B2B software companies: 439 repositories, more than 9.7 million lines, with tickets for 30 of them

Two stacks in one order: a sales-software company's code, about half of it C# (153 repositories; the main application runs about 12 years, to October 2026, with 193 contributors) and a second software company's JavaScript and TypeScript code (286 repositories, 30 of them delivered with their support tickets, engineering chat and work tracking). More than 210,000 commits and more than 120,000 merged pull requests, non-exclusive.

repositories from two software companies (286 and 153)
439
lines of code by the owners' automated reports (5,019,487 and 4,755,994)
more than9.7M
merged pull requests, with more than 210,000 commits
more than120,000
repositories delivered with their support tickets, engineering chat, work tracking and error data
30

What is inside

  • second software company: 286 repositories, 76% JavaScript, 20% TypeScript and 2% Python by lines, earliest commits from 2016 to 2026; 94 run CI and 7,347 test files are committed.
  • 30 of those repositories, 4,229,247 lines, come with the support desk, engineering chat channels, work-tracking records and error-monitoring data; links between tickets and commits have not been checked yet.
  • Sales-software company: 153 repositories after the owner's fixes, with C# the main language in 92 of them (49.6% of all lines are C#); the main application has 2,888,550 lines over about 12 years to October 2026, 75,186 commits, 38,849 merged pull requests and 193 contributors.
  • 54 of the 153 changed in the last 90 days, 60 have 100 or more merged pull requests, and 17 carry a training-task range; the second software company adds 14 task-grade repositories.
  • AI assistance is disclosed per repository: AI-assistant files are committed in 28 of the second company's repositories and 26 of the sales-software company's, and in 2 of the latter an AI coding agent is a commit author.
  • Keys, passwords and personal emails found in code are removed before delivery.
  • Figures come from automated reports the owners ran on each repository, which share metrics and no code; generated and vendored code is not counted, and nothing has been measured by Rightmark yet.

Best for

  • Training and evaluating coding agents on multi-year company codebases in two stacks (C#; JavaScript and TypeScript)
  • Building software-engineering tasks from real commits and merged pull requests (31 task-grade repositories flagged)
  • Issue-to-fix and ticket-to-commit research on the 30 repositories delivered with their tickets
  • Code review and pull request models

Before delivery

  • Each owner confirms its rights to the delivered repositories; the sales-software code is offered only once its owner holds IP assignments from its engineers and contractors.
  • Repositories that were ever public, open-source and third-party code, second copies, client material and generated or vendored code are left out.
  • Keys, passwords and personal emails found in code are rotated or removed before delivery.
  • AI-assistant files and AI-agent commits are disclosed per repository, so a buyer can leave them out.
  • Ticket-to-commit links on the 30 linked repositories are measured and the share stated before delivery.
  • Each owner signs its own license; the buyer gets one card and one order. Delivery starts once two members have cleared title.

Left out

  • Open-source and third-party repositories, second copies, client material and keys, generated or vendored code.
  • Two second company's repositories whose files carry copyleft license headers, until a rights review clears them.
  • Any repository confirmed as ever public.
  • Personal data in the linked tickets and chat, masked per the privacy review.

Lot BND-3-RECORDS

How two B2B software companies work: more than ten years each of mail, chat, wikis, tickets and CRM

One license covers two companies' records: a sales-software company (about 350 mailboxes over the years, team chat from 2015 to October 2026, Jira, Confluence, Notion, Intercom and HubSpot) and a second software company of a few hundred staff (more than 3.5 TB of mail and files, more than 72,000 support tickets, more than 590,000 CRM records). Exclusive for 24 months or non-exclusive.

of records at each company: one sales-software company and one other B2B software company
more than10 years
of mail, documents and chat at the second software company, with about 4.09 million chat messages in the last 13 months
more than3.5 TB
CRM records and more than 72,000 support tickets at the second software company
more than590,000
mailboxes over the years at the sales-software company, with team chat from 2015 to October 2026 (public channels, 5.7 GB unpacked)
about350
at the sales-software company: a mail and documents suite, team chat, two wikis, work tracking, a support desk, a CRM and a call recorder
8 tools

What is inside

  • Sales-software company: email and documents, team chat, wikis, work tracking, customer support and CRM.
  • About 300 people have worked there, 50 to 120 at a time, about 75 today; most staff work outside the US.
  • second software company, more than ten years: mail and documents (3.53 TB), team chat (about 60 GB, from 2015), 72,566 support tickets, 591,074 CRM records and about 100,000 calendar records.
  • A few hundred staff; the owner estimates that about 70% of the records come from staff in India and 30% from the US. Its records are in English.
  • Recorded calls (up to 15,900 hours at the sales-software company) and both companies' code can be added; each is priced on its own card.
  • Delivery only after a rights and privacy review; samples come only with each owner's agreement, as metadata or masked output.

Best for

  • Enterprise environments and agent training on real cross-tool work: mail, chat, wiki, tickets and CRM at two companies
  • Testing agents across two company shapes and regions
  • Long-horizon customer support models on more than ten years of tickets and support chat
  • Sales and customer-success workflows inside a sales-software company's own CRM and support desk

Before delivery

  • Each company's records pass a rights and privacy review before delivery; personal data is masked per that review.
  • Each owner states its exclusions in writing; always out: health data, under-18s, payment-card and bank-account numbers.
  • One exclusive per company: if a company's records are licensed exclusively elsewhere before this bundle is signed, that company leaves the bundle.
  • Code and calls already licensed non-exclusively are listed as carve-outs before signing.
  • Volumes per tool, and the year each tool's records start, are measured and stated before contract.
  • Each owner signs its own license; the buyer gets one card and one order. Delivery starts once two members have cleared title.

Left out

  • Always out: health data, data about under-18s, payment-card and bank-account numbers.
  • Code and recorded calls, priced on their own cards; either can be added.
  • Private chat channels at the sales-software company, unless its owner adds them.
  • Any records a tool vendor claims rights over.

How to quote

Name the lot and the license

Quote at the ask for an exclusive license, a non-exclusive license or an exclusive window, propose your own terms, or pass.

Samples under NDA

Samples and detailed source lists come through Rightmark's buyer portal once you accept its NDA, and only with the owner's agreement.

Rights and privacy, lot by lot

Each owner's rights are reviewed before delivery, with a privacy review for every lot, and the counts are measured and listed in a manifest before signing.

Write to us

Write to hello@rightmark.ai and name the lot, or send a request.

For data buyers

Need data that is not listed?

Tell us what you need and we reply with scope and timing.

Request data

For data owners

Own data like this? List it with Rightmark.

Start with an indicative appraisal: no files or system access needed. It is free, and nothing is licensed until you sign.

Appraise your data

Or start from company records or source code. Know an owner? Become a Scout.