diff --git a/src/scoring/preview.ts b/src/scoring/preview.ts
index 6ef7de7540..94e5fa2c49 100644
--- a/src/scoring/preview.ts
+++ b/src/scoring/preview.ts
@@ -965,7 +965,11 @@ function inferCredibility(evidence?: ContributorEvidenceRecord | null): number {
   const merged = Number(payload?.mergedPullRequests ?? 0);
   const stale = Number(payload?.stalePullRequests ?? 0);
   const unlinked = Number(payload?.unlinkedPullRequests ?? 0);
-  if (!Number.isFinite(merged)) return 0.8;
+  // The payload is a loosely-typed cache (`Record<string, JsonValue>`), so any count can arrive as a
+  // non-numeric value. A non-finite `stale`/`unlinked` propagates NaN through the arithmetic below —
+  // and `clamp` cannot rescue NaN — poisoning the whole credibility multiplier and score, so guard all
+  // three counts, not just `merged`, falling back to the neutral credibility.
+  if (!Number.isFinite(merged) || !Number.isFinite(stale) || !Number.isFinite(unlinked)) return 0.8;
   return clamp(0.75 + merged * 0.04 - stale * 0.03 - unlinked * 0.02, 0.25, 1);
 }
 
diff --git a/test/unit/scoring.test.ts b/test/unit/scoring.test.ts
index 7592bdbde7..0cfc09e2d7 100644
--- a/test/unit/scoring.test.ts
+++ b/test/unit/scoring.test.ts
@@ -4,7 +4,7 @@ import { DEFAULT_ISSUE_DISCOVERY_SHARE, DEFAULT_SCORING_CONSTANTS, detectActiveM
 import { buildScorePreview, calculateTimeDecay, makeScorePreviewRecord, resolveTimeDecay } from "../../src/scoring/preview";
 import { unmodeledScoringConstantsFingerprint } from "../../src/upstream/unmodeled-scoring-drift";
 import type { ScorePreviewInput } from "../../src/scoring/preview";
-import type { RepositoryRecord, ScoringModelSnapshotRecord } from "../../src/types";
+import type { JsonValue, RepositoryRecord, ScoringModelSnapshotRecord } from "../../src/types";
 import { createTestEnv } from "../helpers/d1";
 
 // A realistic constants.py body — at least MIN_RECOGNIZED_SCORING_CONSTANTS (8) recognized constants — so a
@@ -1213,6 +1213,31 @@ NOVELTY_BONUS_SCALAR = 3
     expect(fallbackCredibility.gates.baseTokenGatePassed).toBe(false);
   });
 
+  it("falls back to neutral credibility when any evidence count is non-finite (not just mergedPullRequests)", () => {
+    const score = (payload: Record<string, JsonValue>) =>
+      buildScorePreview({
+        repo,
+        snapshot,
+        contributorEvidence: { login: "riskdev", generatedAt: "2026-05-23T00:00:00.000Z", payload },
+        input: { repoFullName: repo.fullName, sourceTokenScore: 100, totalTokenScore: 200, sourceLines: 10, openPrCount: 0 },
+      });
+
+    // A malformed `stale` or `unlinked` would NaN-poison the credibility multiplier and the whole
+    // estimated score; each must degrade to the same neutral 0.8 the `merged` guard already produced.
+    for (const malformed of [
+      { mergedPullRequests: 5, stalePullRequests: "n/a", unlinkedPullRequests: 0 },
+      { mergedPullRequests: 5, stalePullRequests: 0, unlinkedPullRequests: "bad" },
+    ] satisfies Record<string, JsonValue>[]) {
+      const preview = score(malformed);
+      expect(preview.gates.credibilityObserved).toBe(0.8);
+      expect(Number.isFinite(preview.scoreEstimate.estimatedMergedScore)).toBe(true);
+    }
+
+    // Well-formed counts still flow through the arithmetic rather than the guard.
+    const wellFormed = score({ mergedPullRequests: 5, stalePullRequests: 2, unlinkedPullRequests: 1 });
+    expect(wellFormed.gates.credibilityObserved).toBeCloseTo(0.87, 5);
+  });
+
   it("refreshes scoring snapshots from upstream fixtures and falls back cleanly", async () => {
     const env = createTestEnv({ GITHUB_PUBLIC_TOKEN: "token" });
     vi.stubGlobal("fetch", async (input: RequestInfo | URL) => {