Skip to content
Outrich Data

46 terms

Data deals, in plain words.

The terms that come up when a company sells or licenses data: the contract language, the escrow and title vocabulary, the privacy concepts and the AI training terms. Each definition is written to be read on its own, and each has its own page with an example.

A

Acceptance checks
Acceptance checks are the tests a delivered dataset must pass before the buyer accepts it and escrowed funds are released. Typical checks cover file integrity, record volume, schema, time span, de-identification and duplicates. Each check should name a property, a method, a pass threshold and the party who runs it.
Agent trajectory
An agent trajectory is a recorded sequence of the steps taken to complete a task: what was observed at each point, what action was taken, which tool was used and what resulted. AI labs use trajectories to train and test agents, meaning models that carry out multi-step work across software systems.
Anonymization
Anonymization is processing personal data so that the people it describes can no longer be identified by any means reasonably likely to be used. Under the GDPR, truly anonymous information falls outside data protection law. The standard is demanding, and data that still has a key linking it to people is not anonymous.

C

Chain of title
Chain of title is the documented path showing how a dataset was created, who has held rights in it since, and on what basis it can be licensed today. Each link is supported by evidence, such as system records, assignment agreements and contracts. A gap in the chain is a rights risk.
Checksum manifest
A checksum manifest is a list of every file in a dataset delivery with its name, size and cryptographic hash, usually SHA-256. The recipient recomputes each hash after transfer and compares it with the list. Matching values show the files arrived complete and unaltered, and identify exactly what was delivered.
Chunking
Chunking is splitting long text into smaller pieces so that each fits within what a model or retrieval system can process at once. Good chunks break at natural boundaries, such as paragraphs or conversation turns, and carry metadata that records which source they came from and their position in it.
Closing statement
A closing statement is the record issued when a data transaction completes. It lists the parties, the dataset delivered, the license signed, the amounts deposited and released, the acceptance result and the date of closing. Each party receives a copy, giving both the same account of what changed hands.
Cure period
A cure period is the time a seller has to fix a dataset delivery that failed an acceptance check before the buyer may reject it finally and recover escrowed funds. It begins when the buyer gives written notice naming the failed check, and it ends with a corrected delivery or the return of funds.

D

Data card
A data card is a short document that describes a dataset for a prospective buyer or user. It states the source, the record types, schema, volume, time span, languages, processing and de-identification steps, known gaps and the license scope offered. A buyer reads the data card before it looks at any sample.
Data contamination
Data contamination occurs when material intended for testing a model has already appeared in its training data, or when a dataset sold as unseen is in fact present in existing models. Contaminated test results overstate ability, because the model may be recalling answers instead of working them out.
Data escrow
Data escrow is an arrangement in which a neutral third party holds the buyer's funds, and sometimes the dataset itself, while a data sale completes. The holder releases money to the seller and data to the buyer only when the conditions both parties signed in the escrow instructions have been met.
Data processing agreement (DPA)
A data processing agreement, or DPA, is a contract between a business that controls personal data and a service provider that processes it on the business's behalf. It limits the provider to the customer's documented instructions. A DPA usually prevents the provider from using that data for its own purposes.
Data subject
A data subject is the identified or identifiable living person that personal data relates to. In company data, data subjects include customers, their staff, the company's own employees and anyone mentioned in records. Privacy laws give data subjects rights, such as access, erasure and objection, over data about them.
Data title search
A data title search is an examination, carried out before a data deal closes, of whether the seller has the right to license the data. It reviews how the data was created, who owns it, and what employee agreements, customer contracts, privacy notices and earlier licenses restrict. Its findings go into a title report.
De-identification
De-identification is the process of removing or altering details in a dataset so that records can no longer reasonably be linked to a specific person or client. It covers direct identifiers such as names and account numbers and indirect ones such as rare combinations of traits. Legal standards for it differ by law.

E

Escrow instructions
Escrow instructions are the written directions, signed by both buyer and seller, that tell an escrow holder exactly what to hold, when to release it and to whom. The holder acts on these instructions alone, so they must state release conditions, the cure period and the return of funds in specific terms.
Evaluation data
Evaluation data is a set of examples used to test a model, not to train it. Each example has an input and a known correct outcome, so the model's answers can be scored. Its value depends on the model never having seen it, which is why unpublished company data is sought after.
Expert Determination
Expert Determination is one of two HIPAA methods for de-identifying health information. A person with appropriate statistical and scientific knowledge analyzes the data and determines that the risk of identifying any individual is very small for the anticipated recipient. The expert documents the methods and results that justify the conclusion.

F

Field of use
A field of use is a limit on the purposes or markets in which a licensee may use licensed data. It can be defined by type of use, industry or kind of product. Field-of-use terms let a seller grant rights, or exclusivity, in one area while keeping its freedom in others.
Fine-tuning
Fine-tuning is further training of an existing model on a smaller, focused dataset so that it performs better in a particular domain, task or style. It starts from a model that already has general ability and adjusts it with examples that show the specific behavior wanted, such as resolved cases.

H

Held-out set
A held-out set is a portion of data deliberately kept separate from training so it can be used to test a model fairly. Because the model has not learned from these examples, its performance on them indicates how well it will handle new cases. Keeping the separation intact is essential.

I

Indemnity
An indemnity is a contractual promise by one party to cover specified losses of the other. In a data license, the seller typically indemnifies the buyer against third-party claims that the data infringes rights or was unlawfully shared. The clause sets scope, notice, control of the defense and any cap.

J

JSONL
JSONL, or JSON Lines, is a text file format in which each line is one complete JSON object. There is no enclosing array and no comma between records. It is the common delivery format for AI training data because files can be read, split, sampled and appended one line at a time.

K

k-anonymity
k-anonymity is a measure of re-identification risk. A dataset is k-anonymous when every combination of its quasi-identifier values is shared by at least k records. A person in the data cannot then be narrowed to fewer than k candidates by those fields. A k of one means some record is unique.

L

l-diversity
l-diversity is a privacy measure that extends k-anonymity. It requires each group of records sharing the same quasi-identifier values to contain at least l different values of a sensitive attribute. This prevents an observer from learning a person's sensitive detail merely by knowing which group they belong to.
Lawful basis
A lawful basis is one of the legal grounds the GDPR requires before personal data may be processed. Article 6 lists six: consent, contract, legal obligation, vital interests, public task and legitimate interests. A controller must identify the basis for each purpose, and a new purpose may need a new assessment.

M

Most favored nation clause
A most favored nation clause, or MFN, is a promise that no other licensee will receive better terms for the same data. If the seller later grants another buyer a lower price or broader rights, the first buyer becomes entitled to the same improvement. Sellers usually limit its scope and duration.

P

Perpetual license
A perpetual license grants rights to use data with no end date. The licensee may keep and use its copy indefinitely within the agreed scope. Perpetual does not mean unlimited: field of use, sublicensing limits and exclusivity terms still apply, and the license can still end for breach if the contract says so.
Personal data
Personal data is any information relating to an identified or identifiable living person. The GDPR uses this term, and it reaches well beyond names to include online identifiers, location data and any detail that can be linked to someone. United States laws use similar terms, such as personal information, with their own definitions.
Protected health information (PHI)
Protected health information, or PHI, is individually identifiable health information held or transmitted by a HIPAA covered entity or its business associate. It includes medical records, billing details and any health data tied to a person. PHI may be used or disclosed only as the HIPAA Privacy Rule permits.
Provenance
Provenance is the documented history of a dataset: where each part originated, who created it, and every processing step between the source system and the delivered files. Good provenance lets a buyer confirm the data is what the seller says it is and supports the buyer's own record keeping.
Pseudonymization
Pseudonymization is replacing identifiers in a dataset with codes or hashes so records cannot be attributed to a person without additional information that is kept separately. It reduces risk and keeps records linkable. Under the GDPR, pseudonymized data is still personal data, because the link to the person can be restored.
Purpose limitation
Purpose limitation is the GDPR principle that personal data must be collected for specified, explicit and legitimate purposes and not further processed in a way incompatible with them. A company that wants to use data for a new purpose must first assess whether that purpose is compatible with the original one.

Q

Quasi-identifier
A quasi-identifier is a field that does not identify a person alone but can do so in combination with others or with outside information. Common examples are date of birth, postal code, job title, employer and gender. Removing names while leaving quasi-identifiers intact often leaves records re-identifiable.

R

Re-identification
Re-identification is working out which person a record describes after identifying details have been removed. It usually happens by linking the remaining fields to another source of information, or by recognizing distinctive content. The risk of re-identification is the central measure of whether de-identification has worked.
Refresh
A refresh is a later delivery of new records under an existing data license, usually on a set schedule such as monthly or quarterly. It keeps the licensed dataset current. Refresh terms set what is delivered, when, at what price, and whether each delivery passes through acceptance checks and escrow.
Release conditions
Release conditions are the facts that must be true before an escrow holder pays out funds or hands over data. In a data sale the main condition is acceptance: the delivered dataset has passed the checks written into the license. Conditions should be objective, finite and tied to a named decision maker.
Representations and warranties
Representations and warranties are formal statements of fact a seller makes in a data license, such as that it owns the data, may license it, and complied with privacy law. If a statement proves untrue, the buyer has a contractual remedy. They allocate the risk of facts the buyer cannot verify itself.
RL environment
An RL environment, short for reinforcement learning environment, is a simulated setting in which an AI agent attempts tasks, takes actions and receives a score for the result. AI labs build environments that imitate real software and workflows so that agents can practice and be measured on realistic work.

S

Safe Harbor method
The Safe Harbor method is one of two ways to de-identify health information under the HIPAA Privacy Rule. It requires removing 18 listed categories of identifiers about the individual and their relatives, employers and household members, and having no actual knowledge that the remaining information could identify the person.
Sublicense
A sublicense is a grant by a licensee of some or all of its licensed rights to a third party. In data licensing, the right to sublicense determines whether a buyer may pass the data to affiliates, contractors or customers. It exists only if the license grants it, and on the terms stated.
Synthetic data
Synthetic data is data generated by a program or a model instead of being recorded from real events. It may imitate the patterns of a real dataset or be created from rules. It is used to add volume, cover rare cases or reduce privacy risk, but it lacks the grounding of real records.

T

Term license
A term license grants rights to use data for a fixed period, after which the rights end unless renewed. At expiry the licensee normally must stop using the delivered files and delete them. A term license cannot undo training: a model that learned from the data during the term keeps that learning.
Title report
A title report is the written result of a data title search. It describes the dataset examined, states the basis on which the seller owns it, lists the contracts, notices and earlier licenses reviewed, and sorts findings into what is clear to license, what is restricted and what must be cured before closing.
Token
A token is the unit of text a language model reads and writes. A tokenizer splits text into tokens, each a short word, part of a word or a punctuation mark. Dataset size, model limits and usage prices are measured in tokens, and counts vary between tokenizers.
Training data
Training data is the collection of examples a machine learning model learns from. During training, the model adjusts its internal parameters to predict or reproduce patterns in those examples. The quality, variety and relevance of the training data largely determine what the model can do and where it fails.

Find out if your data qualifies.

Eleven questions, about five minutes, no call required. You get an answer on screen and a written follow-up.

esc