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.

Records linked across systems, how much of the company the records hold (its years and its people), records that show a result, clean records written by people, specialised work and English. Each counts as often as buyers name it in their own words, and each is scored from your files where they were measured, from your answers where not, and in the middle where neither says.

Every record of what a whole company's records sold for, were offered or were estimated at (109 today) counts, unless the buyer's own published rules would not take your company. Each is moved to your company with slopes fitted from the buyers' own estimators (each doubling of the head count multiplies the price by 1.47 times; a closed company fetches 0.65 times an operating one), and counts by what it is (a sale, a buyer's offer and an intermediary's figure count 1, 0.5 and 0.25), by its age and by how far it had to be moved. Offers are turned into expected sales by how sales compare with offers among them.

The pessimistic case is the price a tenth of the weighted records fall below, the optimistic case the price a tenth lie above, both kept within the sales among them. The base sits between the two, placed by the criteria buyers name in their own words, each scored from your files or your answers: on answers alone, about halfway. Public estimators for the same company differ about twentyfold, which is why the cases are far apart.

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. A codebase is priced the way they price it, per repository, per reviewed change, per task and as a company's whole code base, from the 25 code prices on record, each times how many of its unit the code holds, and left out where a buyer's own published rules would not take the code (Turing's minimum size, and its rule on code written by people). Where the base lands follows what code buyers name in their own words: tests in CI, reviewed changes tied to their tickets, code written by people, never public, size. Credentials in the code, code written for clients and public code hold it at nothing until settled.

A task is a problem with a way to check the answer. A set is priced from the 36 prices per task on record (what labs, vendors and marketplaces pay for one task, and what task builders are paid), those in its own field when there are enough, times how many tasks it holds. Where the base lands follows what task buyers name: graded automatically, running as an environment, never published. 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. 1,022 public prices sit behind the tool, each linked to where it was read.

When the base case is at least $26,000, what a small company's data sold or was priced at on average across the 8 such records in our price ledger, 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. When only the optimistic case reaches it, measuring your exports narrows the range and decides. When even that is below, the result says where to go instead.