Operational data
for frontier models
Goldspring values, de-identifies and licenses the records companies produce in the course of real work to the labs building models and agents. De-identification runs in your environment, and every release requires your approval.
- Support team8 years of tickets and chat$90K–$240K
- Engineering org5 years of code, reviews and issues$160K–$420K
- Insurance claims deskClaims with notes and outcomes$220K–$560K
- Sales teamCRM, call notes and email$60K–$180K
Ranges from our valuation model for representative datasets. They are not offers; a valuation requires a reviewed sample.
Estimate your dataset
Describe your systems and history. The estimate uses metadata only; a valuation follows review of a sample.
Where does your data live?
Search for every system involved. Nothing personal is saved until you create an account.
Models have read the public web. They have not seen how work gets done.
Training and evaluating agents requires long-horizon records of expert work: the request, the reasoning, the revisions and the result. These records exist almost entirely inside companies.
Conversations
Decisions in context, with the reasoning that produced them.
Slack, Teams, Gmail, call recordings
Code
Changes linked to the review, discussion and issue behind each one.
GitHub, GitLab, Jira, Linear
Documents
Procedures, specifications and institutional knowledge.
Notion, Confluence, Google Drive
Records
Transactions and cases, each with a recorded outcome.
Salesforce, HubSpot, NetSuite
Tickets
Multi-step work from first request to resolution.
Zendesk, ServiceNow, Intercom
Linkage matters. The same unit of work traced across correspondence, tickets and code is worth more than any one source, and longer histories add further value.
From inventory to license
- 01
Inventory
We estimate value from metadata alone: the systems you use, the people who use them and the years of history they hold. No records are shared at this stage.
- 02
De-identification and valuation
A representative sample is de-identified inside your environment and released only after your review. We measure it against the criteria labs use and return a valuation with its reasoning.
- 03
Licensing
Vetted AI labs license the dataset under contract. Each buyer and each term requires your approval, and listings identify your company only with your consent.
Built by a team that builds training data for frontier labs.
For over a year, the team behind Goldspring has built reinforcement learning environments and evaluations for frontier AI labs. Goldspring applies the same standards to data that already exists inside companies.
How your data is handled
These controls are built into the product and set out in the agreement signed before any data is shared.
Security details- No data is used for training or shared with a buyer without a license you sign.
- Records leave your environment only after you review and approve them. Before that, we receive counts and date ranges.
- Names, contact details, identifiers and credentials are removed in your environment. Each release carries a signed, verifiable privacy receipt.
- Buyers see an anonymized description of the dataset, published only with your approval of its wording.
- If you only want a valuation, your sample is deleted 7 days after it is delivered.
- Stored data is encrypted, and every access by our team is logged and visible to you.
What does it cost?
Nothing. The estimate and the valuation are free. Goldspring is paid a share of license fees, only when a dataset is licensed.
Who licenses the data?
AI labs and companies building models and agents, for training, evaluation and agent environments. Every buyer is vetted, and each one requires your approval.
What data qualifies?
Records of real work: correspondence, code and code review, documentation, CRM and financial records, tickets and project history. Longer histories, more connected systems and recorded outcomes all increase value.
What does it require from our team?
An introductory call, then roughly an hour from one engineer to prepare a sample and review what it contains. De-identification is automated and runs in your environment.
How is a dataset valued?
By scale (systems, people and years of history) and by quality, measured on the properties labs pay for: recorded outcomes, multi-step structure and linkage across systems. The estimate uses metadata; the valuation uses a reviewed sample and comes with its reasoning.
How does de-identification work?
An engineer, or a coding agent such as Claude Code, Codex or Cursor, runs our review tool on exports in your environment. It removes personal and sensitive fields and shows every field that would be released. Nothing is sent without your approval.
How is de-identification verified?
Each release carries a signed privacy receipt: what was removed, by type and by field, the anonymizer version, and our independent re-scan of what arrived. Signatures can be checked on our verification page.
Is licensing required?
No. A valuation carries no obligation. If you stop there, your sample is deleted 7 days after the valuation is delivered.
Will buyers know which company the data comes from?
Only with your consent. Listings describe the dataset with an anonymized profile, such as “mid-market US insurer,” and you approve the wording.