diff --git a/src/cerberus.test.ts b/src/cerberus.test.ts
index 28eef6b..4f70ae6 100644
--- a/src/cerberus.test.ts
+++ b/src/cerberus.test.ts
@@ -88,6 +88,42 @@ describe("mergeSarif", () => {
     expect(results).toBe(3);
     expect(sarif.version).toBe("2.1.0");
   });
+
+  it("keeps only referenced rules and rewrites ruleIndex", () => {
+    // Scanners embed their whole catalogue; only the matched rules may ship.
+    const log = {
+      runs: [{
+        tool: { driver: { name: "semgrep", rules: [
+          { id: "unused-1", help: { markdown: "x".repeat(5000) } },
+          { id: "hit-a", properties: { "security-severity": "9.1" }, help: { markdown: "x".repeat(5000) } },
+          { id: "unused-2" },
+          { id: "hit-b", helpUri: "https://example.com/b" },
+        ] } },
+        results: [
+          { ruleId: "hit-b", ruleIndex: 3, message: { text: "b" } },
+          { ruleId: "hit-a", ruleIndex: 1, message: { text: "a" } },
+          { ruleId: "hit-a", ruleIndex: 1, message: { text: "a again" } },
+        ],
+      }],
+    };
+    const { sarif } = mergeSarif([log]);
+    const run = sarif.runs[0] as { tool: { driver: { rules: Array<Record<string, unknown>> } }; results: Array<Record<string, unknown>> };
+
+    expect(run.tool.driver.rules.map((r) => r.id)).toEqual(["hit-b", "hit-a"]);
+    // The severity source survives; the prose does not.
+    expect(run.tool.driver.rules[1]!.properties).toEqual({ "security-severity": "9.1" });
+    expect(run.tool.driver.rules[1]!.help).toBeUndefined();
+    // Indices now point at the pruned array.
+    expect(run.results.map((r) => r.ruleIndex)).toEqual([0, 1, 1]);
+    expect(JSON.stringify(sarif)).not.toContain("xxxxx");
+  });
+
+  it("drops a ruleIndex that cannot be resolved rather than mispointing it", () => {
+    const log = { runs: [{ tool: { driver: { rules: [{ id: "a" }] } }, results: [{ ruleIndex: 7, message: { text: "?" } }] }] };
+    const { sarif } = mergeSarif([log]);
+    const run = sarif.runs[0] as { results: Array<Record<string, unknown>> };
+    expect(run.results[0]!.ruleIndex).toBeUndefined();
+  });
 });
 
 describe("evaluateGate", () => {
diff --git a/src/merge.ts b/src/merge.ts
index da890ee..134ca38 100644
--- a/src/merge.ts
+++ b/src/merge.ts
@@ -1,21 +1,89 @@
 /** Merge per-scanner SARIF logs into one 2.1.0 log (runs are concatenated —
  *  the backend normalizes and dedups per run). */
 
-interface SarifLike {
-  runs?: unknown[];
+interface SarifRule {
+  id?: string;
+  [key: string]: unknown;
+}
+
+interface SarifResult {
+  ruleId?: string;
+  ruleIndex?: number;
+  [key: string]: unknown;
+}
+
+interface SarifRun {
+  tool?: { driver?: { rules?: SarifRule[]; [key: string]: unknown }; [key: string]: unknown };
+  results?: SarifResult[];
+  [key: string]: unknown;
+}
+
+/** Rule fields the backend uses (severity, docs) or that stay cheap. Everything
+ *  else — `help`, `fullDescription`, `relationships` — is prose we would ship
+ *  by the megabyte. */
+const KEPT_RULE_FIELDS = ["id", "name", "shortDescription", "helpUri", "properties", "defaultConfiguration"];
+
+function slimRule(rule: SarifRule): SarifRule {
+  const out: SarifRule = {};
+  for (const field of KEPT_RULE_FIELDS) {
+    if (rule[field] !== undefined) out[field] = rule[field];
+  }
+  return out;
+}
+
+/**
+ * Keep only the rules a run's results actually reference, and rewrite their
+ * `ruleIndex` to match.
+ *
+ * Scanners embed their whole catalogue: a Semgrep or Checkov report carries
+ * thousands of rule definitions with long help texts, so a scan with a handful
+ * of findings still weighs tens of megabytes and the upload is rejected.
+ */
+function pruneRun(run: SarifRun): SarifRun {
+  const rules = run.tool?.driver?.rules;
+  if (!Array.isArray(rules) || rules.length === 0) return run;
+
+  const results = Array.isArray(run.results) ? run.results : [];
+  const kept: SarifRule[] = [];
+  const newIndexById = new Map<string, number>();
+
+  const keep = (rule: SarifRule | undefined): number | undefined => {
+    if (!rule) return undefined;
+    const id = typeof rule.id === "string" ? rule.id : undefined;
+    if (id && newIndexById.has(id)) return newIndexById.get(id);
+    const index = kept.length;
+    kept.push(slimRule(rule));
+    if (id) newIndexById.set(id, index);
+    return index;
+  };
+
+  const newResults = results.map((result) => {
+    const byIndex = typeof result.ruleIndex === "number" ? rules[result.ruleIndex] : undefined;
+    const byId = typeof result.ruleId === "string" ? rules.find((r) => r?.id === result.ruleId) : undefined;
+    const index = keep(byIndex ?? byId);
+    // Drop a stale ruleIndex rather than let it point at the wrong rule.
+    const { ruleIndex: _old, ...rest } = result;
+    return index === undefined ? rest : { ...rest, ruleIndex: index };
+  });
+
+  return {
+    ...run,
+    tool: { ...run.tool, driver: { ...run.tool?.driver, rules: kept } },
+    results: newResults,
+  };
 }
 
 export function mergeSarif(logs: unknown[]): { sarif: { version: string; $schema: string; runs: unknown[] }; results: number } {
   const runs: unknown[] = [];
   let results = 0;
   for (const log of logs) {
-    const l = (log ?? {}) as SarifLike;
+    const l = (log ?? {}) as { runs?: unknown[] };
     if (!Array.isArray(l.runs)) continue;
-    for (const run of l.runs) {
-      if (!run || typeof run !== "object") continue;
-      runs.push(run);
-      const r = (run as { results?: unknown[] }).results;
-      if (Array.isArray(r)) results += r.length;
+    for (const runRaw of l.runs) {
+      if (!runRaw || typeof runRaw !== "object") continue;
+      const run = runRaw as SarifRun;
+      runs.push(pruneRun(run));
+      if (Array.isArray(run.results)) results += run.results.length;
     }
   }
   return {