diff --git a/scripts/verify-self-audit-proof.test.ts b/scripts/verify-self-audit-proof.test.ts
index 080fea1..01f03a4 100644
--- a/scripts/verify-self-audit-proof.test.ts
+++ b/scripts/verify-self-audit-proof.test.ts
@@ -865,3 +865,62 @@ test("main() creates diff.txt in retain bundle when reports diverge", () => {
     rmSync(tmp, { recursive: true, force: true });
   }
 });
+
+// ─── PR-mined theme table is live data, not a determinism signal ─────────────
+
+const MINED_TABLE = (rows: string) => `${CLEAN_REPORT}
+## Observed Failure Modes (PR Review Mining)
+
+*Why this matters:* This run analyzed 25 PRs and surfaced 4 recurring rule candidates.
+
+Repo: \`lambda-curry/anvil\`
+PRs analyzed: 25 · Comments reviewed: 29 · Substantive comments: 29 · Candidates: 4
+
+| Theme | Frequency | PR Spread | Severity | Rule Signal Match | Comment Alignment |
+|-------|-----------|-----------|----------|-------------------|------------------|
+${rows}
+`;
+
+test("a new PR-mined theme row is not a determinism failure", () => {
+  // The break that took main red after #41: a code review pushed a fifth theme
+  // over its threshold between a branch build and the merge build of the same
+  // commit, the rows shifted, and "Documentation" was compared against a newly
+  // inserted "Error Handling".
+  const checkedIn = MINED_TABLE(
+    `| Naming | 9 comments | 3 PRs (medium) | low | 🟡 signal match | 100% strong |
+| Documentation | 5 comments | 2 PRs (medium) | low | 🟡 signal match | 100% strong |`,
+  );
+  const fresh = MINED_TABLE(
+    `| Naming | 11 comments | 5 PRs (high) | medium | 🟡 signal match | 100% strong |
+| Error Handling | 3 comments | 3 PRs (medium) | medium | 🟡 signal match | 100% strong |
+| Documentation | 5 comments | 2 PRs (medium) | low | 🟡 signal match | 100% strong |`,
+  );
+
+  expect(compareSelfAuditReports(checkedIn, fresh).failures).toEqual([]);
+});
+
+test("the mined table disappearing entirely is still a failure", () => {
+  // The line held deliberately: rows changing is churn and normalizes away, but
+  // the table vanishing means mining itself broke, and that should stay loud.
+  const checkedIn = MINED_TABLE(
+    `| Naming | 9 comments | 3 PRs (medium) | low | 🟡 signal match | 100% strong |`,
+  );
+  const fresh = MINED_TABLE("");
+
+  expect(
+    compareSelfAuditReports(checkedIn, fresh).failures.length,
+  ).toBeGreaterThan(0);
+});
+
+test("a real scoring change is still caught through the mined table", () => {
+  // The guard: normalizing live PR data must not blind the proof to the
+  // deterministic surface it exists to protect.
+  const checkedIn = MINED_TABLE(
+    `| Naming | 9 comments | 3 PRs (medium) | low | 🟡 signal match | 100% strong |`,
+  );
+  const fresh = checkedIn.replace("Issues found | none", "Issues found | 3");
+
+  expect(
+    compareSelfAuditReports(checkedIn, fresh).failures.length,
+  ).toBeGreaterThan(0);
+});
diff --git a/scripts/verify-self-audit-proof.ts b/scripts/verify-self-audit-proof.ts
index e5894be..f10de01 100644
--- a/scripts/verify-self-audit-proof.ts
+++ b/scripts/verify-self-audit-proof.ts
@@ -53,6 +53,18 @@ type CliOptions = {
   retainDir: string | null;
 };
 
+/** One row of the PR-mined theme table: `| Naming | 9 comments | 3 PRs (medium) | low | … |`. */
+const MINED_THEME_ROW_PATTERN =
+  /^\| .+? \| \d+ comments \| \d+ PRs \([a-z]+\) \| [a-z]+ \|.*$/gm;
+
+const MINED_THEME_ROW = "| <normalized PR-mined theme row> |";
+
+/** Consecutive placeholder rows, so the NUMBER of themes stops mattering too. */
+const COLLAPSE_MINED_THEME_ROWS = new RegExp(
+  `(?:${MINED_THEME_ROW.replaceAll(/[.*+?^${}()|[\]\\]/g, String.raw`\$&`)}\n?)+`,
+  "g",
+);
+
 function normalizeVolatileReportFields(reportText: string): string {
   return (
     reportText
@@ -73,10 +85,16 @@ function normalizeVolatileReportFields(reportText: string): string {
       // already normalized. Anvil mines its own PR history, so these move whenever anyone opens
       // a PR here and the packet needs a hand refresh: 20->22, then 22->24 days later.
       // Normalizing them ends the chore.
-      .replaceAll(
-        /^\| (.+?) \| \d+ comments \| \d+ PRs \([a-z]+\) \| [a-z]+ \|/gm,
-        "| $1 | <normalized> comments | <normalized> PRs (<normalized>) | <normalized> |",
-      )
+      //
+      // The theme SET is volatile too, not only the counts inside each row. Normalizing just the
+      // numbers still broke the moment a code review pushed a fifth theme over its threshold:
+      // the rows shifted, and "Documentation" was compared against a freshly inserted "Error
+      // Handling". Which themes appear is derived wholly from live GitHub comments and can differ
+      // between two runs minutes apart — a branch build and the merge build of the same commit
+      // disagreed exactly that way. So the block collapses to one placeholder. What this proof
+      // exists to protect — scores, stages, checks, every deterministic line — is untouched.
+      .replaceAll(MINED_THEME_ROW_PATTERN, MINED_THEME_ROW)
+      .replace(COLLAPSE_MINED_THEME_ROWS, `${MINED_THEME_ROW}\n`)
   );
 }