Skip to main content
What works today | LLMWhat is built, what evidence stands behind each part, and what that evidence does not establish.LLMstatusllmstatusdeveloperoperatorbuild-agent-systemsdeploy-operate-productsreference

What works today

Every library below is implemented and tested. None of it is released, and none of it has been run against a live provider or a real hosting account. The tests use pinned response bytes, loopback sockets and in-process fakes, and make no paid call.

By capability​

CapabilityCratesEvidence
Neutral turn: port, items, streaming, cancellation, observations, typed failuresllm-coreESS suite and recorded mutations
Bounded HTTP/SSE transportllm-httpRust fixture tests; one recorded mutation (the turn deadline)
Injected secrets, coordinated renewal, file and keychain adaptersllm-credentialsESS suite and recorded mutations; keychain against mock stores only
Bindings and request-time authenticationllm-providersCovered through the routing suite and recorded mutations
TOML catalog, selection and explanationllm-routingESS suite and recorded mutations
Ordered fallbackllm-routingRust tests with scripted models, ESS scenarios in the routing suite, recorded mutations
Chat Completions projection and clientllm-chatESS suite and recorded mutations
Messages projection and clientllm-messagesESS suite and recorded mutations
Responses projectionllm-responsesESS suite and recorded mutations
Usage pricingllm-costESS suite and recorded mutations
Durable spending limits (SQLite)llm-costESS suite over real local storage, recorded mutations
Authenticated single-owner gateway: probes and route inventoryllm-gatewayRust tests over a real loopback socket, recorded mutations; no ESS suite
Hosting lifecycle contractllm-provisionESS suite against FakeProvider, recorded mutations
Runpod vLLM adapterllm-runpod52 Rust tests against EmulatedRunpod, 39 recorded mutations; no ESS suite

What the evidence means​

An ESS suite is a set of written scenarios in the repository's ESS specification, run through an adapter against the real public library functions, three times, with identical counts. A skipped scenario fails the gate.

A recorded mutation is a deliberate break in the production source, a named test or scenario shown to fail because of it, and a byte-for-byte restore. Each crate's set lives in docs/verification/*-falsification.json. It shows that the tests would catch a regression. It says nothing about what a real provider returns.

These are different strengths. The gateway and Runpod adapter have recorded mutations over their own Rust tests but no ESS suite yet; llm-http has its fixture tests and a single mutation.

What no test here establishes​

  • That any provider's live API or subscription answers as the fixtures do.
  • That the Runpod control plane behaves as the emulator does, or that stopping a pod stops billing.
  • That an OS keychain service is available on a given machine.
  • That a provider's usage report is correct. Pricing trusts the observations it is given.

See Limitations for the full trust boundary and Not yet for what is missing.

Planning state​

Lifecycle state lives in the repository's AEP planning store. Of 31 stories, 18 are implemented, one (runtime-contracts: releasing the versioned contract) is active, and 12 are drafts. The drafts are the items on Not yet, plus four stories that add ESS specifications for the gateway, HTTP transport, provider bindings and Runpod adapter.