diff --git a/README.md b/README.md
index 3b8746d..8456c52 100644
--- a/README.md
+++ b/README.md
@@ -11,70 +11,22 @@ through QC (your validators), reworked until it's in-spec, stamped with a serial
 model call. **The catalog is infinite.**
 
 ```ts
-import { Output, ToolLoopAgent } from "ai"
-import { check, InMemoryThinkingBlockStore, judge, ThinkingBlock } from "thinking-blocks"
-import { z } from "zod"
-
-type FoodInput = { food: string }
-
-// The spec — the shape every finished part must match.
-const NutritionFacts = z.object({
-  serving: z.string(),
-  calories: z.number().nonnegative(),
-  protein: z.number().nonnegative(),
-  carbs: z.number().nonnegative(),
-  fat: z.number().nonnegative(),
-})
-type NutritionFacts = z.infer<typeof NutritionFacts>
-
-const Judgement = z.object({ ok: z.boolean(), reason: z.string() })
-type Judgement = z.infer<typeof Judgement>
+import { check, judge, ThinkingBlock } from "thinking-blocks"
 
-// The machine. Define it once; the catalog it serves is infinite.
-const nutrition = new ThinkingBlock<FoodInput, NutritionFacts>({
+// A machine: an agent makes the part, validators are its QC gates.
+const nutrition = new ThinkingBlock({
   name: "nutrition",
-  store: new InMemoryThinkingBlockStore(),
-  agent: new ToolLoopAgent({
-    model: "openai/gpt-5",
-    output: Output.object({ schema: NutritionFacts }),
-  }),
-  // The serial number: same food -> same part, forever.
-  identity: ({ food }) => food.trim().toLowerCase(),
-  prepareCall: ({ input }) => ({
-    prompt: `Nutrition facts for one typical serving of "${input.food}". Give "serving" as a short human description and protein/carbs/fat in grams.`,
-  }),
-  attempts: { max: 3 },
+  store,        // where finished parts are kept
+  agent,        // a ToolLoopAgent — its output schema is the spec
+  prepareCall: ({ input }) => ({ prompt: `Nutrition facts for ${input.food}` }),
   validators: [
-    // QC, no model — a caliper: the stated calories must reconcile with the
-    // macros (4·protein + 4·carbs + 9·fat), or the part is reworked.
-    check<FoodInput, NutritionFacts>("macros-reconcile", {
-      validate: ({ output }) => {
-        const implied = 4 * output.protein + 4 * output.carbs + 9 * output.fat
-        return Math.abs(implied - output.calories) <= Math.max(40, 0.25 * output.calories)
-          ? { success: true }
-          : { success: false, feedback: `${output.calories} kcal doesn't match the macros (~${Math.round(implied)} kcal) — fix the numbers.` }
-      },
-    }),
-    // QC, with a model — an inspector: is this a realistic serving for the food?
-    judge<FoodInput, NutritionFacts, Judgement>("serving-is-realistic", {
-      agent: new ToolLoopAgent({
-        model: "openai/gpt-5",
-        output: Output.object({ schema: Judgement }),
-      }),
-      schema: Judgement,
-      prepareCall: ({ input, output }) => ({
-        prompt: `Is "${output.serving}" a realistic serving of "${input.food}", with plausible macros?`,
-      }),
-      validate: ({ judgement }) =>
-        judgement.ok ? { success: true } : { success: false, feedback: judgement.reason },
-    }),
+    check("macros-reconcile", { validate }),                       // QC in code
+    judge("serving-realistic", { agent, schema, prepareCall, validate }), // QC by a model
   ],
 })
 
-// Order a finished part: the machine makes it, runs it through QC, keeps it.
-const facts = await nutrition.get({ food: "dragon fruit" })
-// facts.output is { serving, calories, protein, carbs, fat }, validated and kept;
-// the same food -> the same part, instantly, forever, with no model call.
+const part = await nutrition.get({ food: "dragon fruit" })
+if (part.ok) part.output.calories // 60 — typed, validated, kept
 ```
 
 Under the hood a Thinking Block is `function + AI agent + validation + memory +