Each person's salary is encrypted in their own browser to a model running in an attested enclave on NEAR AI Cloud. Only pay bands come out, and only for groups of at least four. A NEAR contract checks the enclave's signature before it will publish anything.
The point is what nobody gets to see.
Only ciphertext. Your salary is encrypted before it leaves your browser, to a key that exists only inside the model's enclave. Your role is public so the contract can count groups; your salary is not.
Decrypts every record, computes the 25th percentile, median and 75th percentile per role, and returns only those. Its key is bound to an Intel TDX + NVIDIA attestation you can check yourself.
Refuses a result unless the enclave signed it, and the request was exactly the instruction fixed at the start plus each submission once, as its own message, nothing added. Each published group must have at least four people.
Stated plainly.
Browser encryption to NEAR AI Cloud's attested model key, ciphertexts on NEAR testnet, one end-to-end encrypted request to the model, on-chain ed25519 check of its signature, the k-anonymity rule, and in-browser verification of everything on this page.
The model does the arithmetic. The signature proves which inputs went in and what came out, not that the percentiles are right; a language model can get them wrong. Attestation is checked for a fresh nonce here; the full Intel/NVIDIA quote check is NEAR AI's verifier. Demo salaries are made up.
Run the percentile code itself in an attested enclave (deterministic, no model), coarser bands for small groups, and payroll-signed salary proofs so nobody can inflate their own number.