Every number that touches someone's employment is deterministic arithmetic you can recompute from its inputs. The language model only writes prose — it never scores, ranks or decides. Governance is enforced in the API, not described in a PDF.
Same payroll, same people. The raw average hides what a controlled regression finds — and ATLAS HR reports both side by side, with the caveat that controlling for level understates cumulative disparity.
No number appears without its decomposition.
Attrition risk and candidate scores return the weight each factor contributed — recomputable from inputs, never a black box.
Applies the four-fifths rule, then withdraws its own verdict when too few decisions exist for the ratio to mean anything.
Reject a candidate without a written reason and you get a 422 citing the policy clause and GDPR Article 22 — enforced, not documented.
Policy answers carry clause citations and refuse what the corpus can't ground. A model that can never say "I don't know" is just a confident one.
Every inference and automated decision is logged and tamper-evident, with inputs hashed rather than stored verbatim.
Workers, D1 and Workers AI. No third-party model provider; employee data never leaves Cloudflare's network.
No servers to manage.