diff --git a/packages/core/src/trace-projection.test.ts b/packages/core/src/trace-projection.test.ts
index 4a34d38..d27ee39 100644
--- a/packages/core/src/trace-projection.test.ts
+++ b/packages/core/src/trace-projection.test.ts
@@ -224,4 +224,49 @@ describe('projectTrace (V9.1)', () => {
     const toolStep = trace.steps.find((s) => s.kind === 'tool' && s.tool?.id === 'live-run');
     expect(toolStep?.status).toBe('running');
   });
+
+  it('projects the benchmark case prompt as the trace input, never the answer', () => {
+    const caseDef: EvaluationCase = {
+      id: 'case-1',
+      name: 'NVDA research',
+      category: 'tool-selection',
+      difficulty: 'golden',
+      input: { prompt: 'Research NVDA' },
+      expected: {},
+      tags: [],
+      source: 'hand-authored',
+    };
+    const evalRun: EvaluationRun = {
+      id: 'eval-run-1',
+      experimentId: 'exp-1',
+      caseId: 'case-1',
+      datasetId: 'ds-1',
+      status: 'completed',
+      startedAt: 1000,
+      completedAt: 9000,
+      answer: 'NVDA looks fine',
+      toolCalls: [],
+      failureModes: [],
+    };
+    const trace = projectTrace({ evaluationRun: evalRun, evaluationCase: caseDef });
+    expect(trace.input).toBe('Research NVDA');
+    expect(trace.input).not.toBe(evalRun.answer);
+  });
+
+  it('leaves the trace input empty when no authoritative prompt was recorded', () => {
+    const evalRun: EvaluationRun = {
+      id: 'eval-run-2',
+      experimentId: 'exp-1',
+      caseId: 'case-2',
+      datasetId: 'ds-1',
+      status: 'completed',
+      startedAt: 1000,
+      completedAt: 9000,
+      answer: 'NVDA looks fine',
+      toolCalls: [],
+      failureModes: [],
+    };
+    const trace = projectTrace({ evaluationRun: evalRun });
+    expect(trace.input).toBe('');
+  });
 });
diff --git a/packages/core/src/trace-projection.ts b/packages/core/src/trace-projection.ts
index 6997bba..10bc573 100644
--- a/packages/core/src/trace-projection.ts
+++ b/packages/core/src/trace-projection.ts
@@ -291,7 +291,11 @@ export function projectTrace(input: TraceProjectionInput): FolioTrace {
     startedAt,
     completedAt,
     latencyMs,
-    input: run?.input ?? evaluationRun?.answer ?? '',
+    // The evaluation run record carries no input of its own; the benchmark
+    // case prompt is the authoritative input for that run (spec §9). Falling
+    // back to `answer` would project the OUTPUT as the input, which is the
+    // exact reconstruction-from-run-state the projection must never do.
+    input: run?.input ?? evaluationCase?.input.prompt ?? '',
     answer,
     error: run?.error?.message ?? evaluationRun?.error?.message,
     completeness: deriveCompleteness(Boolean(run || evaluationRun), tools, traceEvents, traceRefForOutput),