Frontier Signal / Glossary
The words we use.
Short definitions for the terms in the brief, the watchlist and the price index.
Buyers and programmes
- Buyer
- A company that says in public it pays for private data. Some buy for themselves; some sell on to AI labs.
- Own-account buyer
- A buyer that licenses data for its own use, such as training or evaluation, and does not resell it.
- Marketplace
- A buyer that prepares data and licenses it to labs, then pays the owner a share of each sale.
- Intermediary
- A firm that sits between data owners and the labs that use the data.
- Programme
- A buyer’s published offer: what it takes, its terms and, when stated, its price.
- Lab-direct programme
- A data programme run by an AI lab itself rather than by a buyer in between.
- Referral or introduction fee
- Money a programme pays a person who introduces a company that later sells data to it.
- Buyer board
- A public list of buyers with their published terms and figures. Board totals count buyers on that list, not completed deals.
- Reported concerns
- A buyer the board has taken off its main list because of open questions about its claims or terms.
Prices
- Published figure
- A price a buyer or programme states in public. It is not an offer or a completed sale.
- Middle half
- The range that holds the middle half of published prices for one kind of data: a quarter of prices sit below it and a quarter above. Read it as the usual band.
- Ceiling or “up to”
- The top figure a programme advertises. Few sellers reach it.
- Top published figure
- The highest figure any buyer on the board publishes. It is a ceiling across the whole market, not a typical price.
- Deal range
- The spread of prices buyers list for one kind of company, such as law firms or medical practices.
- Unit
- What a price is for: per company, per repository, per hour, per image or per word. Compare prices only in the same unit.
- Revenue share
- Payment as a share of each later sale instead of one upfront price.
- Not published
- The source names no figure. We do not estimate one.
Data types
- Company records
- A company’s working history: email and chat threads, documents, tickets, orders and invoices, usually de-identified.
- Source code
- Private codebases, often with their change history.
- Datasets, tasks and environments
- Prepared exercises and settings that AI systems practise or are tested on, often built from real work.
- Agent trajectories
- Recorded step-by-step runs of an AI agent or a person doing a task.
- Physical-world data
- Video and sensor records of real places, machines and manual work.
- Expert and human work
- Paid time from people with domain skills who write, check or grade AI work.
- De-identification
- Removing names and other details that point to a real person before data changes hands.
- Exclusive licence
- The buyer is the only one allowed to use the data. Non-exclusive means the owner can license it again.
- Headcount floor
- The smallest company, by staff, a buyer will take. Many buyers publish none.
Labs and hiring
- Frontier lab
- A company building the largest general AI models.
- Posting
- One open job listing on a lab’s job board. Postings are not hires.
- Human Data
- The lab teams that source, manage and check the human-made data used to train and evaluate models.
- Over-indexed
- A role or word that shows up more often than you would expect from the rest of the job board.
Reading a cut
- Cut
- One dated reading of every source. Each brief and archive page belongs to one cut.
- Material change
- A change to a buyer, a price, a term or a lab role that the brief calls out as new.
- Holding cut
- A check that found no material change. It keeps the last material cut on the brief.
- Weekly rollup
- The public brief for one week, ending Friday (America/New_York).
- Mention count
- How often a category comes up across tracked sources. Categories overlap, so it is a rough guide, not a count of buyers.
- Coverage gap
- A place where the evidence stops: a page we could not reach, a parked domain or a claim we could not check.