Compose a pipeline with the weakest component at each position

Composition does not imply agency. A claim checker needs four kinds of component — a plain function, a typed model call, a membership check and an agent — and each one is used only where the weaker ones cannot decide. This page shows the load-bearing parts of the canonical examples/ch40-composition/pipeline.ts, run on Pi 1.0.4 with a scripted provider.

Run it

The two deterministic pieces need no model at all — and one of them is the check the model’s output must pass:

// 1. A plain function. No model. Deterministic, free, testable with no script.
export function retrieve(corpus: Corpus, claim: string): string[] {
	const words = claim.toLowerCase().split(/\W+/).filter((w) => w.length > 3);
	return Object.values(corpus).filter((text) => words.some((w) => text.toLowerCase().includes(w)));
}

// 2. A plain function again: the check the model's output must pass.
// A citation the model invented is not in the evidence, so it fails here.
export function citationsAreReal(a: EvidenceAssessment, evidence: string[]): boolean {
	return a.citations.every((c) => evidence.includes(c));
}

The check runs immediately after every model call and before anything downstream:

	const checked = (a: EvidenceAssessment, given: string[]) => {
		if (!citationsAreReal(a, given)) throw new Error("assessment cites evidence that was not provided");
	};
	checked(assessment, evidence);

The escalation agent exists for one reason — an insufficient verdict — and is bounded so a model that keeps searching cannot run away:

		const agent = makeResearchAgent(models, model, corpus);
		let turns = 0;
		let overran = false;
		const limit = ESCALATION_TURNS;
		const unsubscribe = agent.subscribe((e) => {
			if (e.type === "turn_end" && ++turns >= limit) {
				overran = true;
				agent.abort();
			}
		});
		try {
			await agent.prompt(`Find more evidence about: ${claim}`);
		} finally {
			unsubscribe();

Expected outcome

Every row is observed: pipeline.test.ts (Pi 1.0.4, scripted responses) asserts it, and the request counts are exact for that control flow.

The agent is the one component that can run away, so only it carries a bound — stated in one constant, ESCALATION_TURNS.

Mechanism and limitations

Documented: everything the components rely on — one-request assess(), the Agent loop with tools, and step()’s maxAttempts/maxTurns (pi-ai and pi-agent-core 1.0.4). Observed: the eight tests in pipeline.test.ts, under scripted responses. Proposed: the design — weakest component that works, checks after the model, extract evidence rather than trust prose, and the bound value.

A membership test catches a citation that is not in the evidence; it cannot catch a conclusion that is wrong while every citation is real. The request counts are not a prediction: a real model may take a different path.

Full source and test .