What is your data worth to AI?

Who can sign, who owns the records, whether health, children's or personal financial data is inside, how many identifiers the exports hold, credentials left in code. These hold, cap or exclude a part; they are never averaged into a score.

People and years of records, measured where the exports show them, and which kinds of records exist together: conversations, documents and knowledge, tickets and projects, code with its history, customer records, product and telemetry data, calls and meetings, finance records, engineering and design files, orders and inventory. Buyers price a company by which of these it can show together.

Repeats, automated messages and gaps in time lower it. Records that show a result, records that state a reason and records that link to another system raise it: a ticket joined to the change that closed it and the conversation around it is what buyers repeat most often as the reason they pay more.

The anchor is the formula three buyers share for company records: $140,000 plus $2,333 a person plus $11,667 a year, times a location factor. Realised sales sit well below buyers' calculators, so the middle is 50% of it until our own closed sales replace that figure.

The range runs 4x either side of the middle on answers alone (public estimators for the same company differ about twentyfold), 3x once the exports are measured and 2x with a checked sample.

The same records sold exclusively run about 4 to 5 times a non-exclusive licence; an evaluation-only licence 0.2 to 0.5 times. They are shown side by side, never blended.

Buyers turn code into tasks: a reviewed change, the issue behind it and the tests that prove it. So a codebase is priced three ways and the strongest counts: by the token ($100 a million tokens of code, the floor), as graded repositories ($5,000 to $20,000 each by size, reviews, tests and tickets), or by the tasks its history yields (reviewed changes that touch tests, times the chance it builds in a container, times 50% surviving validation, at about $866 a task, half of it the owner's). Then languages, code written before AI assistants, ownership and whether any of it was ever public. Public code is worth little: it is in every model already.

A task is a problem with a way to check the answer. Priced per task by field ($500 for software to $1,400 for hard engineering), then by how answers are checked, whether each runs in a container, whether current models have been run on it, and whether any was published. A marketplace keeps 20%. Tasks age as models improve, so a set is worth most in the months it is made.

Video and footage, Audio and speech, Images, Text and archives, Simulated environments, Robot and physical-world data. A unit price somebody published (an hour of footage, an image, a million words, an environment), times the size you give, tapering as size grows, then adjusted for what moves the price: releases, consent, language, kind of recording. 975 public prices sit behind the tool, each linked to where it was read.

When the middle of the range is at least $50,000, at least one buyer's published rules leave room for the data, someone who can sign is asking, and the rights and any sensitive part can be settled first. Below that a sale with several buyers costs more than it adds, and the result says where to go instead.