Governed
by execution.
Every AI request checked in-path. Every refusal recorded. Every dollar accounted.
Fire a request through six gates.
Pick a scenario. Watch it pass — or get refused in-path.
Select a scenario above to begin.
Two kinds of governance
Only one is running here. Policy that executes, not paperwork that promises.
Six gates, in-path.
Every AI request passes through six enforcement checkpoints. Disabled model, exceeded budget, rate limit — refused before tokens are generated.
Every call, written.
Operation, purpose, model, tokens, cost, latency, status. Append-only, monotonic, exportable. History cannot be rewritten — only extended.
Registered models only.
If a model is not in the registry, it does not serve. Route revisions are versioned. Fallback chains are explicit. No shadow models.
Accountability, never content.
Live, append-only, exportable. Every request writes an event — succeeded or refused.
Append-only. Every request writes an event.
Disabled model → refused. Exceeded budget → rejected. Missing route → error. Every gate refuses in-path. No tokens, no spend, no exceptions.
Every AI capability is registered as a governed purpose with its own routing, budgets, and audit trail. If a capability is not on this list, it does not reach a model.
Token-level governance tells you what the model cost. Decision-level governance tells you who it helped decide — and that is where hiring law lives.
Every AI touchpoint is a governed purpose.
Parse, match, analyze, draft — the touchpoint list is a query, not an investigation.
AI assistance and human verdict, bound together.
Alongside token events, R1 records hiring decisions with reasoning in the decision audit trail.
Accountability chains end at a person.
The system records assistance; people own verdicts. Which is what regulators and auditors actually ask for.
The ledger records accountability — never content.
Operation and purpose of every call.
Requested model vs. model that served. Token counts — prompt, completion, total. Cost per request, month, principal. Latency and terminal status. Route revision in force. Hashed principal and billable owner.
Prompt text or model responses.
Candidate documents or profiles. Message or conversation content. Un-hashed identifiers. Append-only. Monotonic config revisions. History cannot be rewritten — only extended. Exportable for compliance review.
Stated plainly: what the architecture supports today, and what remains ahead.
Annex III — high-risk.
Employment and recruitment are named high-risk. The obligations — documentation, logging, human oversight — are the artifacts this system produces by executing.
Audit-ready data model.
Bias audits require enumerating every AI touchpoint in hiring. Purpose-bound governance makes that enumeration a query, not a project.
Boundary documented.
Adverse-impact review needs the exact boundary of algorithmic assistance in selection. Purposes draw it explicitly, per operation.
On the roadmap: independent third-party bias audits, provider attestations, gateway content guardrails. A governance page that overstates is itself a governance failure.
Three rules the system is built on.
Governance must execute.
A policy that cannot block a request is a suggestion. Controls run in the request path — enforcement is code, not documentation.
Every token is accounted.
No AI request exists outside the ledger. Unlogged spend is a defect, not an edge case.
Humans decide.
AI ranks, drafts, extracts, analyzes. Hiring decisions remain with people — the system records assistance, never verdicts.
Proved in public
Enterprise dashboards ask you to trust them. R1's infrastructure is published, live, and independently checkable.
Hire(d) better.