diff --git a/packages/core/canaries.test.ts b/packages/core/canaries.test.ts
new file mode 100644
index 0000000..2fa94c5
--- /dev/null
+++ b/packages/core/canaries.test.ts
@@ -0,0 +1,267 @@
+// =============================================================================
+// Canary tests — semantic invariants on compiled prompt output
+// =============================================================================
+// Pins the executive phrasings the compiler is supposed to produce. If a
+// refactor accidentally drops any of these structures while keeping unit
+// tests green, the corresponding canary fires loudly.
+//
+// Each canary documents what it pins. Refinements per #47 review:
+//   - "must be valid JSON" is load-bearing in the contract narrative; pinned.
+//   - Strategy preamble matches on the strategy NAME appearing, not on a
+//     specific sentence (lower brittleness, same regression-catching power).
+//   - Retry-context section header and previousFailureReason content are
+//     asserted independently — both halves can break independently.
+//   - Output-schema fenced block (```json ... ```) is asserted as a serialisation
+//     contract distinct from the surrounding prose.
+//
+// These canaries are AUTHORED AGAINST CURRENT COMPILER OUTPUT — they codify
+// present behaviour as the contract going forward, not aspirational behaviour.
+// =============================================================================
+
+import { describe, expect, it } from "vitest";
+import { compileStep } from "./compiler.js";
+import type { ExecutionContext, LogicSpec, Step } from "./types.js";
+
+// -----------------------------------------------------------------------------
+// Helpers (match style of compiler.test.ts)
+// -----------------------------------------------------------------------------
+
+function makeSpec(steps: Record<string, Step>, overrides: Partial<LogicSpec> = {}): LogicSpec {
+	return {
+		spec_version: "1.0",
+		name: "canary-spec",
+		steps,
+		...overrides,
+	};
+}
+
+function makeCtx(overrides: Partial<ExecutionContext> = {}): ExecutionContext {
+	return {
+		currentStep: "test",
+		previousOutputs: {},
+		input: null,
+		attemptNumber: 1,
+		branchReason: null,
+		previousFailureReason: null,
+		...overrides,
+	};
+}
+
+// -----------------------------------------------------------------------------
+// Canaries
+// -----------------------------------------------------------------------------
+
+describe("canaries: compiler output semantic invariants", () => {
+	// -- Output contract ------------------------------------------------------
+
+	it("step with output_schema produces a Required Output Format section", () => {
+		const spec = makeSpec({
+			step_a: {
+				output_schema: {
+					type: "object",
+					required: ["result"],
+					properties: { result: { type: "string" } },
+				},
+			},
+		});
+
+		const compiled = compileStep(spec, "step_a", makeCtx());
+
+		expect(compiled.systemPromptSegment).toContain("## Required Output Format");
+	});
+
+	it("output schema produces the load-bearing 'must be valid JSON' phrasing", () => {
+		// Pinned per #47 review: this phrasing carries the contract narrative.
+		// A refactor that softens it (e.g. to "should be JSON-shaped") would
+		// silently weaken the project's central pitch.
+		const spec = makeSpec({
+			step_a: {
+				output_schema: { type: "object" },
+			},
+		});
+
+		const compiled = compileStep(spec, "step_a", makeCtx());
+
+		expect(compiled.systemPromptSegment).toContain("must be valid JSON");
+	});
+
+	it("output schema is rendered inside a fenced ```json block, not inlined as prose", () => {
+		// Pinned per #47 review as a serialisation-contract regression test
+		// distinct from the surrounding "must be valid JSON" prose. If the
+		// schema serialisation is silently switched to plain text or to a
+		// non-JSON fence, this fires.
+		const spec = makeSpec({
+			step_a: {
+				output_schema: {
+					type: "object",
+					properties: { x: { type: "number" } },
+				},
+			},
+		});
+
+		const compiled = compileStep(spec, "step_a", makeCtx());
+
+		expect(compiled.systemPromptSegment).toMatch(/```json\s*\n[\s\S]+?\n```/);
+	});
+
+	// -- Quality gates --------------------------------------------------------
+
+	it("step with quality gates produces a Pre-Response Checklist section", () => {
+		const spec = makeSpec(
+			{
+				step_a: {},
+			},
+			{
+				quality_gates: {
+					pre_output: [{ name: "groundedness", check: "{{ output.cited == true }}" }],
+				},
+			},
+		);
+
+		const compiled = compileStep(spec, "step_a", makeCtx());
+
+		expect(compiled.systemPromptSegment).toContain("## Pre-Response Checklist");
+	});
+
+	it("quality gates produce 'Before responding, verify:' framing", () => {
+		const spec = makeSpec(
+			{
+				step_a: {},
+			},
+			{
+				quality_gates: {
+					pre_output: [{ name: "groundedness", check: "{{ output.cited == true }}" }],
+				},
+			},
+		);
+
+		const compiled = compileStep(spec, "step_a", makeCtx());
+
+		expect(compiled.systemPromptSegment).toContain("Before responding, verify:");
+	});
+
+	it("quality gate items render as Markdown checkboxes (- [ ])", () => {
+		const spec = makeSpec(
+			{
+				step_a: {},
+			},
+			{
+				quality_gates: {
+					pre_output: [
+						{ name: "groundedness", check: "{{ output.cited == true }}", message: "Cite sources" },
+					],
+				},
+			},
+		);
+
+		const compiled = compileStep(spec, "step_a", makeCtx());
+
+		expect(compiled.systemPromptSegment).toMatch(/^- \[ \]/m);
+	});
+
+	// -- Confidence -----------------------------------------------------------
+
+	it("step with confidence config produces explicit threshold language", () => {
+		const spec = makeSpec({
+			step_a: {
+				confidence: { minimum: 0.7, target: 0.85 },
+			},
+		});
+
+		const compiled = compileStep(spec, "step_a", makeCtx());
+
+		expect(compiled.systemPromptSegment).toMatch(/minimum confidence of [0-9.]+/);
+	});
+
+	it("step with escalate_below produces escalation language", () => {
+		const spec = makeSpec({
+			step_a: {
+				confidence: { minimum: 0.7, escalate_below: 0.4 },
+			},
+		});
+
+		const compiled = compileStep(spec, "step_a", makeCtx());
+
+		expect(compiled.systemPromptSegment).toMatch(/falls below [0-9.]+/);
+		expect(compiled.systemPromptSegment).toMatch(/escalate/i);
+	});
+
+	// -- Retry context (split into two independent canaries per #47 review) ---
+
+	it("retry context (attempt > 1) produces a Retry Context section header", () => {
+		// Header presence — independent of whether a previousFailureReason is set.
+		const spec = makeSpec({
+			step_a: {},
+		});
+
+		const compiled = compileStep(spec, "step_a", makeCtx({ attemptNumber: 2 }));
+
+		expect(compiled.systemPromptSegment).toContain("## Retry Context");
+	});
+
+	it("retry context surfaces the previousFailureReason content", () => {
+		// Content presence — independent of the section header. Both halves
+		// can regress independently per #47 review.
+		const reason = "low confidence on step_a output";
+		const spec = makeSpec({
+			step_a: {},
+		});
+
+		const compiled = compileStep(
+			spec,
+			"step_a",
+			makeCtx({ attemptNumber: 2, previousFailureReason: reason }),
+		);
+
+		expect(compiled.systemPromptSegment).toContain(reason);
+	});
+
+	// -- Strategy preamble (per #47 review: match name, not sentence) ---------
+
+	it("reasoning strategy is named in the strategy preamble", () => {
+		// Per #47 review: match on the strategy NAME appearing somewhere in the
+		// segment, not on a specific sentence. Lower brittleness, same
+		// regression-catching power.
+		const spec = makeSpec(
+			{
+				step_a: {},
+			},
+			{
+				reasoning: { strategy: "react" },
+			},
+		);
+
+		const compiled = compileStep(spec, "step_a", makeCtx());
+
+		expect(compiled.systemPromptSegment).toContain("react");
+	});
+
+	// -- Branch routing -------------------------------------------------------
+
+	it("branch context surfaces the routing reason when present", () => {
+		const spec = makeSpec({
+			step_a: {},
+		});
+
+		const compiled = compileStep(
+			spec,
+			"step_a",
+			makeCtx({ branchReason: "previous step output exceeded threshold" }),
+		);
+
+		expect(compiled.systemPromptSegment).toContain("## Branch Context");
+		expect(compiled.systemPromptSegment).toContain("previous step output exceeded threshold");
+	});
+
+	// -- Step identity --------------------------------------------------------
+
+	it("compiled step segment names the current step in a Current Step header", () => {
+		const spec = makeSpec({
+			my_step: {},
+		});
+
+		const compiled = compileStep(spec, "my_step", makeCtx());
+
+		expect(compiled.systemPromptSegment).toContain("## Current Step: my_step");
+	});
+});