MS-P1: Every Layer Is Yours

Spec: course/milestones/MS-P1.toml. Requires a fresh pass of your own for every core Pass 1 module: M03.1, L0.0, lang.04 to lang.07, L10.0, gw.00, dep.00, obs.00, ops.00 (the drill, graded by ol drill end), and craft.02. MS-P1 also reruns the smoke steps of MS-P0. The C primer lang.03 is optional practice.

What it runs

Step	Where	Passes when
train-bigram	local	{tinyllm} train bigram on the fixture corpus reaches NLL at most 3.2004 nats per byte (the count model gives 3.199389)
greedy-from-your-checkpoint	local	greedy generation of 32 tokens from “Once” equals the fixture ids exactly
engine-conforms-openapi-v0	local	ol conform openapi:v0:engine against your engine
gateway-conforms-openapi-v0	local	ol conform openapi:v0:gateway against your gateway in front of your engine
stream-through-your-gateway	local	a keyed streaming request through your gateway yields at least 32 content chunks
adr-0001-present, engine-crashloop-runbook-present	full	the documents exist in your repo
stream-through-the-kind-nodeport	kind	the same stream through the gateway NodePort 30080
one-trace-spans-gateway-and-engine	kind	Jaeger holds one trace whose gateway.proxy span is the parent of the engine’s POST /v1/completions

Run it

practice/bin/ol milestone MS-P1 --smoke   # local processes only: what PR CI runs
practice/bin/ol milestone MS-P1           # adds the documents and the kind steps

The runner first runs your [build].steps, then starts your engine and gateway on allocated ports and waits for their health URLs. Your system.toml declares, per spec/cli-roles.md:

[build]
steps = [["cargo", "build", "--release", "--manifest-path", "rust/Cargo.toml", "-p", "tl-serve"],
         ["go", "build", "-C", "go", "-o", "bin/gateway", "./cmd/gateway"],
         ["uv", "run", "--project", "python", "python", "python/tinyllm/__main__.py",
          "train", "bigram", "--data", "{fixture:MS-P1/corpus.txt}", "--out", "artifacts/models/bigram"]]

[entry]
tinyllm = ["uv", "run", "--project", "python", "python", "python/tinyllm/__main__.py"]
engine  = ["rust/target/release/tl-serve", "--model-dir", "artifacts/models/bigram", "--port", "{port}", "--health-port", "{health_port}"]
gateway = ["go/bin/gateway", "--port", "{port}", "--health-port", "{health_port}", "--upstream", "http://127.0.0.1:{engine.port}"]

[services.engine]
health = "http://127.0.0.1:{health_port}/healthz"

[services.gateway]
health = "http://127.0.0.1:{health_port}/readyz"
after  = ["engine"]

[endpoints]
api_base    = "http://127.0.0.1:{gateway.port}/v1"
api_key_env = "TL_API_KEY"                # export a key in it before you run

[deploy]
kube_context = "kind-<system>"
namespace    = "<system>"
gateway_url  = "http://127.0.0.1:30080"
traces       = "http://127.0.0.1:30686"
services     = { engine = "deploy/<system>-engine", gateway = "deploy/<system>-gateway" }

A kind step without a reachable cluster is skipped with the reason and the verdict is incomplete, not a pass. --ref-deps runs the milestone while some required module is still red and records the verdict as assisted; your entry points always run as they are.