diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 5a9a35e..9460f5c 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -10,7 +10,11 @@ env:
 
 jobs:
   test:
-    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-latest]
+    runs-on: ${{ matrix.os }}
 
     steps:
       - name: Checkout
@@ -44,6 +48,41 @@ jobs:
       - name: Run tests
         run: cargo test --verbose
 
+  msrv:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
+
+      - name: Read minimum supported Rust
+        id: msrv
+        run: |
+          python3 - <<'PYTHON'
+          import os
+          import tomllib
+          with open("Cargo.toml", "rb") as manifest:
+              version = tomllib.load(manifest)["package"]["rust-version"]
+          with open(os.environ["GITHUB_OUTPUT"], "a") as output:
+              output.write(f"version={version}\n")
+          PYTHON
+
+      - name: Set up minimum supported Rust
+        uses: dtolnay/rust-toolchain@631a55b12751854ce901bb631d5902ceb48146f7 # stable
+        with:
+          toolchain: ${{ steps.msrv.outputs.version }}
+
+      - name: Resolve dependencies using the declared MSRV
+        run: cargo generate-lockfile
+
+      - name: Build all targets on MSRV
+        run: cargo check --locked --all-targets --all-features
+
+      - name: Run tests and doctests on MSRV
+        run: cargo test --locked --all-features
+
+      - name: Verify packaged crate on MSRV
+        run: cargo package --locked
+
   # Check documentation builds
   docs:
     runs-on: ubuntu-latest
diff --git a/Cargo.toml b/Cargo.toml
index 22e77af..5ff62a8 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -2,11 +2,14 @@
 name = "braintrust-sdk-rust"
 version = "0.1.0-alpha.2"
 edition = "2021"
+resolver = "3"
 rust-version = "1.86.0"
 authors = ["Braintrust Team <team@braintrust.dev>"]
 description = "Rust SDK for Braintrust logging and tracing"
 repository = "https://github.com/braintrustdata/braintrust-sdk-rust"
 license = "Apache-2.0"
+readme = "README.md"
+include = ["src/**", "tests/**", "examples/**", "README.md", "LICENSE", "braintrust-logo.svg"]
 keywords = ["braintrust", "logging", "tracing", "observability"]
 categories = ["api-bindings", "asynchronous", "development-tools"]
 
diff --git a/README.md b/README.md
index 0fa848c..7dbe3f1 100644
--- a/README.md
+++ b/README.md
@@ -1,4 +1,4 @@
-[![Braintrust](./braintrust-logo.svg)](https://www.braintrust.dev/)
+[![Braintrust](https://raw.githubusercontent.com/braintrustdata/braintrust-sdk-rust/main/braintrust-logo.svg)](https://www.braintrust.dev/)
 
 # braintrust-sdk-rust
 
@@ -8,55 +8,180 @@ Rust SDK for [Braintrust](https://braintrust.dev) logging and tracing.
 
 ## Installation
 
-Add this to your `Cargo.toml`:
+The API below is unreleased. To use this checkout, add it as a path dependency:
 
 ```toml
 [dependencies]
-braintrust-sdk-rust = "0.1.0-alpha.1"
+braintrust-sdk-rust = { path = "../braintrust-sdk-rust" }
 ```
 
-## Usage
+## Rust version
 
-```rust
-use braintrust_sdk_rust::{BraintrustClient, BraintrustClientConfig, SpanLog};
+The minimum supported Rust version is **1.86.0**. This repository uses Cargo
+resolver 3 to prefer dependency versions compatible with its declared MSRV.
+
+Consuming applications control their own resolver and lockfile. For Rust 1.86
+builds, select resolver 3 in the application's root manifest, for example:
+
+```toml
+[workspace]
+resolver = "3"
+```
+
+Regenerate an existing incompatible lockfile with the compatible resolver;
+changing the resolver alone does not replace versions already locked. See
+[Cargo's dependency resolver documentation](https://doc.rust-lang.org/cargo/reference/resolver.html#resolver-versions).
+
+## Logging and tracing
+
+Set `BRAINTRUST_API_KEY`, or pass a key with `.api_key(...)`. A client owns its
+runtime by default, including when built inside an existing Tokio runtime.
+
+```rust,no_run
+use braintrust_sdk_rust::{BraintrustClient, SpanLog};
 use serde_json::json;
+use std::time::Duration;
 
 #[tokio::main]
 async fn main() -> anyhow::Result<()> {
-    // Create a client pointing to your Braintrust instance
-    let client = BraintrustClient::new(
-        BraintrustClientConfig::new("https://api.braintrust.dev")
-    )?;
+    let client = BraintrustClient::builder()
+        .default_project("my-project")
+        .build()
+        .await?;
 
-    // Create a span for logging
-    let span = client
-        .span_builder("your-api-token", "your-org-id")
-        .project_name("my-project")
-        .build();
-
-    // Log input/output using span.log() with the builder pattern
-    span.log(
+    let span = client.span_builder().await?.name("greeting").build();
+    let write = span.log(
         SpanLog::builder()
             .input(json!({"prompt": "Hello, world!"}))
             .output(json!({"response": "Hi there!"}))
-            .build()
-            .expect("Failed to build SpanLog")
-    ).await;
+            .build()?,
+    );
+    write.await?;
+    span.end().await?;
+    client.shutdown(Duration::from_secs(10)).await?;
+    Ok(())
+}
+```
+
+`log()` submits immediately. `end()` finishes the span and returns a completion
+for delivery of its final snapshot. Later mutations of a finished span are
+ignored with a diagnostic. Create child spans for separate operations; use
+`update_span()` or `log_feedback()` to patch an existing row explicitly.
+
+## Submission and completion
+
+Methods that schedule work return `Completion<T, E>`. Calling the method submits
+the operation; awaiting its token observes the result. Inspect `&*completion`
+for the immediate admission `Result<(), E>`, such as a full or closed queue.
+An admitted operation can still fail later. Dropping the token leaves accepted
+work running.
+
+```rust,no_run
+use braintrust_sdk_rust::{SpanHandle, SpanLog};
+
+async fn log(span: &SpanHandle) -> anyhow::Result<()> {
+    let completion = span.log(SpanLog::default());
+    if let Err(error) = &*completion {
+        eprintln!("submission rejected: {error}");
+    }
+    completion.await?;
+    Ok(())
+}
+```
+
+Writes are batched and submitted automatically. `client.flush()` captures queued
+and active writes through the point of the call and observes their delivery.
+A terminal batch failure fails its write completions and covering flush, then
+retires that batch. Later flushes process new work independently. Failures are
+not retained for replay.
+
+Await `client.shutdown(timeout)` before exiting to close admission and drain
+accepted work. A timeout fails remaining completions and discards pending SDK
+state. Concurrent shutdown calls share the first call's deadline and outcome.
+
+## Synchronous applications and runtime selection
 
-    // Flush the span data to Braintrust
-    span.flush().await?;
-    client.flush().await?;
+The same API works on a plain thread without creating a Tokio runtime. Use a
+future executor to wait for values when needed; for example, with the `futures`
+crate:
 
+```rust,no_run
+use braintrust_sdk_rust::BraintrustClient;
+use std::{future::IntoFuture, time::Duration};
+
+fn main() -> anyhow::Result<()> {
+    let client = futures::executor::block_on(
+        BraintrustClient::builder().build().into_future(),
+    )?;
+    futures::executor::block_on(client.shutdown(Duration::from_secs(10)).into_future())?;
     Ok(())
 }
 ```
 
-## Features
+To select your own Tokio runtime, pass `.runtime(current_runtime())` while
+inside it, or `.runtime(Some(handle))`. With an explicit handle, keep that
+runtime alive and driving work until completions finish. `current_runtime()`
+returns `None` outside Tokio, which selects an SDK-owned runtime.
+
+## Evaluations
+
+An `EvalDefinition` contains reusable task and scorer callbacks. Each run supplies
+its own case source and options. Local runs need no client, credentials or
+serialization:
+
+```rust
+use braintrust_sdk_rust::eval::{
+    self, Case, EvalDefinition, EvalOpts, EvalRequest, ExactMatch, SyncFnTask,
+};
+
+fn main() -> anyhow::Result<()> {
+    let definition = EvalDefinition::new(
+        Box::new(SyncFnTask::new(|input: &String, _hooks| Ok(input.to_uppercase()))),
+        vec![Box::new(ExactMatch::new())],
+    )?;
+    let options = EvalOpts::builder()
+        .name("uppercase".into())
+        .definition(definition)
+        .source(vec![Case {
+            input: "hello".to_string(),
+            expected: Some("HELLO".to_string()),
+            ..Default::default()
+        }].into())
+        .quiet(true)
+        .build();
+    let completion = eval::run(EvalRequest::local(options));
+    assert!(completion.is_ok());
+    let summary = futures::executor::block_on(async { completion.await })?;
+    assert_eq!(summary.successful_cases, 1);
+    assert_eq!(summary.results[0].output.as_deref(), Some("HELLO"));
+    Ok(())
+}
+```
+
+Use `EvalRequest::recorded(options, client)` to record an experiment on that
+client's runtime. Recorded inputs and outputs must implement `Serialize`.
+`DatasetCaseSource` reads Braintrust datasets and preserves dataset provenance
+when recording. Run failures retain partial summaries and structured causes.
+
+Use `task(...)` and `scorer(...)` for async callbacks, or `SyncFnTask` and
+`SyncFnScorer` for blocking callbacks. Attach a terminal, JSONL, silent or custom
+reporter with `EvalRequest::with_reporter(...)`. Reporter failures preserve
+computation and recording results.
+
+`BtEvalRunner` adapts the same execution engine to `bt eval`, including discovery,
+filters and output transports. See the
+[CLI example](https://github.com/braintrustdata/braintrust-sdk-rust/blob/main/examples/eval_async.rs)
+and [synchronous example](https://github.com/braintrustdata/braintrust-sdk-rust/blob/main/examples/eval_sync.rs).
+
+## API organization
 
-- **Span-based logging**: Create spans to track LLM calls and other operations
-- **Usage metrics extraction**: Built-in extractors for OpenAI and Anthropic usage metrics
-- **Async-first**: Built on Tokio for high-performance async operations
-- **Background submission**: Logs are submitted in the background to minimize latency
+- `BraintrustClient` creates concrete dataset, experiment and span handles with
+  shared credentials and lifecycle.
+- `eval` separates definitions, case sources, execution, recording and reporting.
+- `api::Client` exposes direct endpoint operations with one HTTP attempt per
+  call. Use it when your application owns delivery and retry policy.
+- Usage extractors support OpenAI and Anthropic responses; streaming helpers
+  support provider response assembly.
 
 ## Preserving existing span provenance
 
@@ -112,4 +237,4 @@ let usage = extract_anthropic_usage(&anthropic_response);
 
 ## License
 
-This project is licensed under the [Apache License, Version 2.0](LICENSE).
+This project is licensed under the [Apache License, Version 2.0](https://github.com/braintrustdata/braintrust-sdk-rust/blob/main/LICENSE).
diff --git a/src/lib.rs b/src/lib.rs
index 504c5b0..fc64fdd 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -1,3 +1,4 @@
+#![doc = include_str!("../README.md")]
 #![deny(private_bounds, private_interfaces)]
 
 pub mod api;