For AI labs and data vendors
Company working data with provenance you can file.
Codicilla licenses cleaned working data from real companies across industries: email, documents, tickets, CRM and call transcripts, engineering files, and code with its history. It comes from operating companies and from closed or acquired ones. Every dataset ships with a data card and a rights file, and Codicilla signs as the licensor.
What we offer
Cross-linked records
Tickets tied to the email and documents that resolved them, specs tied to the work that shipped. Links are preserved through cleanup and reported on the card.
Long time spans
Years of a team's work in one dataset, across many industries and functions.
Real operating context
Customer issues, sales and service calls, project decisions and engineering work that synthetic data does not reproduce.
Measured cleaning
Residual PII rate, public overlap and exclusions are measured and stated, not assumed. Regulated data is left out by default.
Data card
One page per dataset, before any sample.
Every field is filled. Where something is unknown, the card says so.
- Dataset
- EXAMPLE: B2B scheduling SaaS, engineering and support archive
- Source company
- Anonymized. US B2B SaaS, seed through Series A
- Years covered
- 2019-03 to 2024-11
- Code
- 38 repos, about 1.9M lines, 41,000 commits. TypeScript, Python, Go, SQL
- Docs and wiki
- 2,300 Notion pages, about 1.4M words
- Support tickets
- 61,000 tickets, 190,000 messages, linked to commits where a fix shipped
- Chat
- 820,000 Slack messages in 140 public channels. DMs and private channels excluded
- Seller of record
- Example Co. Liquidating LLC, acting through its ABC assignee
- Chain of title
- Assignment deed on file. IP assignments for all past staff and contractors
- PII handling
- Automated detection plus human review of a 2% sample. Stable pseudonyms. Residual PII in audit sample: 0.03% of messages
- Open source
- Scanned. GPL and AGPL files removed. Permissive licenses kept with notices
- Dedup
- 6.1% of code files matched public repositories and were removed
- Exclusions
- HR, legal and board channels. Secrets. Customer exports. EU and UK data subjects
- License
- Non-exclusive, 3 years, training and internal evaluation only
License terms
Plain terms, set per dataset.
Terms are agreed per license. These are the defaults we start from.
- Exclusivity
- Non-exclusive by default. Exclusive licenses, by field or term, on request.
- Permitted use
- Model training. Evaluation by agreement. No resale or redistribution.
- No regurgitation
- The licensee applies controls against verbatim output of source content.
- Deletion on termination
- Raw data deleted at the end of the term, with written certification.
- Audit rights
- Defined audit rights on both sides, on notice.
- Warranty
- Codicilla warrants title and the stated cleaning process, within agreed limits.
Provenance documentation
The rights file.
Each dataset comes with the documents behind it, available for your counsel's review under NDA.
- Licensor or seller of record and its authority: board or owner approval, assignment deed, or court order
- Lien search results
- Employee and contractor IP assignment confirmation
- Customer contract review and third-party content decisions
- Privacy policy review and any employee notices and opt-outs
- Cleaning log: PII method, secrets scan, license scan, dedup results
How licensing works
- 1
Tell us the need
Domains, languages, content types, volume and timing.
- 2
Data card and sample
Review the card, then a stratified sample under NDA.
- 3
License
Agree scope, term, exclusivity and permitted use.
- 4
Delivery
Checksummed transfer to your storage, in JSONL, Parquet or git bundles.
Tell us what your models need.
We are lining up first companies now and want to hear what buyers need before we sign them.