Skip to main content
Wardin sits in the request path between your code and your LLM providers, and turns every request into one signed artifact instead of four disconnected ones. Enforcing policy, metering cost, signing the outcome, and attributing it to a team or session all happen to the same request, in the same pass — an ED25519-signed, hash-chained receipt. That one receipt is read by three different audiences:
  • Security / compliance — did governance controls actually run on this request?
  • Finance — what did it cost, and is the number trustworthy?
  • Effectiveness — did this governed spend produce accepted work?
See The Evidence Layer for how those three readers share one record instead of three reconciled reports.

​
The core problem

Most teams deploying AI at scale stitch together a gateway + observability tool + guardrail service + FinOps tool + productivity tool to get four separate, after-the-fact views of the same request. Each sits outside the request path, producing dashboards that lag reality and logs that only describe what allegedly happened — none of them enforces anything, and none of them can prove it. Wardin sits in the request path. That means:
  • Budgets are enforced before a token is spent, not discovered in a report
  • PII is redacted before it leaves your network
  • Policies (model allowlist, injection detection) block non-compliant calls at the source
  • Every enforcement decision, cost, and actor is signed into one hash-chained receipt — not logged separately and reconciled later
  • Every request is attributed to a team, user, and session, so governed spend can be tied to accepted work

​
How it works

You point your existing SDK at https://gw.wardin.ai instead of https://api.anthropic.com. No client code changes. Signing and chaining happen in an asynchronous writer, off the response path — a receipt adds zero latency to your request. See Signed Receipts for how the chain works.
Wardin produces audit-grade, Art-12-style runtime records for gateway-routed traffic — not a compliance certification. No software makes an organization “compliant,” and Wardin doesn’t hold (or confer) SOC 2, ISO 27001, or HIPAA. Receipts today are verifiable over the hot analytics window (7–30 days). Framework-mapped control citations, the one-click Evidence Bundle export, and the offline verifier (spec published; reference CLI, public release in preparation) are shipped today; long-term WORM retention beyond the hot window is still in development — see The Evidence Layer for what’s shipped versus roadmap.

​
Key concepts

The Evidence Layer

The organizing idea: enforce, meter, sign, and attribute converge into one receipt, read by three audiences.

The Request Path

The six-stage rail — the mechanism that produces the evidence layer on every request.

Virtual Keys

Scoped API keys that proxy to your real provider credentials — with per-key budgets, rate limits, and policies.

Budgets

Hard spending limits enforced via Redis atomic counters before a request is forwarded.

Policies

Model allowlists, prompt injection detection, PII redaction — enforced in-path.

Caching

L1 exact-match and L2 semantic caching reduce spend without changing your prompts.

Sessions

Requests grouped into sessions so cost-per-task is a real metric — built for agentic clients.

Signed Receipts

Every model-inference request emits an ED25519-signed, hash-chained receipt — tamper-evident audit evidence.