fix(orb): stop double-counting PRs already tracked by the own-ledger

src/orb/outcomes.ts

@@ -41,10 +41,16 @@ export interface OrbGlobalStats {
  * The public global aggregate: merged / closed / total terminal PR outcomes across REGISTERED installations
  * only (registered = 1) — an install that hasn't been opted in never contributes to the public counter. SUM over
  * no matching rows is NULL, so each total is nullish-guarded to 0 (fail-safe on an empty/cold table).
+ *
+ * The NOT EXISTS anti-join skips any (repo, pr_number) that already has a `github_app.pr_public_surface_published`
+ * audit event — i.e. a PR the own-ledger disposition query in public-stats.ts already counted. Without it, a PR
+ * reviewed before the self-host cutover (own-ledger) that also has a terminal outcome recorded here (Orb) gets
+ * counted twice. Quantified 2026-07-12: 243 PRs (173 merged + 70 closed), 96% in one repo, inflating the public
+ * "PRs reviewed" counter. That event_type's target_key only ever references the own-ledger's own repos, so this
+ * is a no-op for every other registered installation's outcomes.
  */
 export async function getOrbGlobalStats(env: Env, opts: { excludeAccount?: string } = {}): Promise<OrbGlobalStats> {
-  // excludeAccount de-dups an account already counted by another source — the homepage counts JSONbored's own
-  // repos via cloud review_audit, so it excludes that account here to avoid double-counting. "" = include all.
+  // excludeAccount de-dups an account already counted by another source. "" = include all.
   const exclude = (opts.excludeAccount ?? "").toLowerCase();
   const row = await env.DB.prepare(
     `SELECT
@@ -53,7 +59,12 @@ export async function getOrbGlobalStats(env: Env, opts: { excludeAccount?: strin
        COUNT(*) AS total
      FROM orb_pr_outcomes o
      JOIN orb_github_installations i ON i.installation_id = o.installation_id AND i.registered = 1
-     WHERE ? = '' OR LOWER(COALESCE(i.account_login, '')) <> ?`,
+     WHERE (? = '' OR LOWER(COALESCE(i.account_login, '')) <> ?)
+       AND NOT EXISTS (
+         SELECT 1 FROM audit_events ae
+         WHERE ae.event_type = 'github_app.pr_public_surface_published'
+           AND LOWER(ae.target_key) = LOWER(o.repository_full_name) || '#' || o.pr_number
+       )`,
   )
     .bind(exclude, exclude)
     .first<{ merged: number | null; closed: number | null; total: number | null }>();

src/review/public-stats.ts

@@ -25,14 +25,15 @@
 // GLOBAL: the homepage total folds in every REGISTERED Orb installation's outcomes (getOrbGlobalStats) on top of
 // the own-ledger side, so the counter reflects the whole fleet, not just gittensory's own repos. The own-ledger
 // side (audit_events) is a FROZEN snapshot as of the self-host cutover -- it stops growing the day each repo's
-// live processing moved off this worker -- while orb_pr_outcomes keeps growing in realtime for any repo with the
-// central Orb App installed (including JSONbored's own, which still runs the Orb App for telemetry alongside its
-// self-hosted review engine). The two sources are NOT mutually exclusive by account: JSONbored appears in both,
-// but only overlaps for the few-day window before the cutover froze the own-ledger side, so no excludeAccount is
-// applied here -- a small, bounded double-count for that historical window is preferable to silently dropping
-// all of JSONbored's live post-cutover Orb data (which excluding "jsonbored" used to do, since every REGISTERED
-// installation currently happens to be a JSONbored account). See getOrbGlobalStats for the general-purpose
-// excludeAccount dedup this call deliberately does not use.
+// live processing moved off this worker, and can never grow again now that the old App has been fully deleted --
+// while orb_pr_outcomes keeps growing in realtime for any repo with the central Orb App installed (including
+// JSONbored's own, which still runs the Orb App for telemetry alongside its self-hosted review engine). The two
+// sources overlap by (repo, pr_number), not just by account: earlier reasoning here called that overlap "a small,
+// bounded double-count," but it was measured directly (2026-07-12) at 243 PRs (173 merged + 70 closed, 96% in one
+// repo) -- large enough to move accuracyPct and mislead visitors, not a rounding error. getOrbGlobalStats now
+// excludes exactly those overlapping (repo, pr_number) pairs via a NOT EXISTS anti-join against the same
+// `github_app.pr_public_surface_published` audit events this file's own disposition query reads, so the two sums
+// are over disjoint PR sets and can be added directly.
 import { getOrbGlobalStats } from "../orb/outcomes";
 
 /** FALLBACK estimate of maintainer review/triage time saved per reviewed PR, used ONLY when the real per-PR

test/integration/orb-outcomes.test.ts

@@ -96,4 +96,19 @@ describe("getOrbGlobalStats", () => {
     expect(await getOrbGlobalStats(e, { excludeAccount: "jsonbored" })).toEqual({ merged: 0, closed: 1, total: 1 }); // JSONbored dropped
     expect(await getOrbGlobalStats(e)).toEqual({ merged: 1, closed: 1, total: 2 }); // no exclude → both
   });
+
+  it("excludes a PR already counted by the own-ledger (published-surface audit event), but still counts an unrelated one", async () => {
+    const e = createTestEnv();
+    const db = e.DB as unknown as TestD1Database;
+    await registerInstall(e, 100, 1);
+    // acme/dup#1 was already published (and therefore counted) by the own-ledger disposition query.
+    await db
+      .prepare(
+        "INSERT INTO audit_events (id, event_type, target_key, outcome, created_at) VALUES ('evt-1', 'github_app.pr_public_surface_published', 'acme/dup#1', 'success', '2026-06-24T00:00:00Z')",
+      )
+      .run();
+    await recordOrbPrOutcome(e, "pull_request", closedPr("acme/dup", 1, "2026-06-24T00:00:00Z", 100)); // merged, already own-ledger-counted → must be excluded
+    await recordOrbPrOutcome(e, "pull_request", closedPr("acme/new", 2, null, 100)); // closed, no own-ledger counterpart → must still count
+    expect(await getOrbGlobalStats(e)).toEqual({ merged: 0, closed: 1, total: 1 });
+  });
 });

test/unit/public-stats.test.ts

@@ -34,11 +34,18 @@ function isWeekly(sql: string): boolean {
 function isEffort(sql: string): boolean {
   return sql.includes("reviewEffortMinutes");
 }
+// getOrbGlobalStats's own-ledger anti-join also references `github_app.pr_public_surface_published` (to skip
+// PRs the disposition query already counted) — exclude it here the same way isWeekly/isEffort already are, or
+// every stub that doesn't special-case `orb_pr_outcomes` would wrongly route the orb read through here too.
+function isOrbGlobal(sql: string): boolean {
+  return sql.includes("orb_pr_outcomes");
+}
 function isDispositions(sql: string): boolean {
   return (
     sql.includes("github_app.pr_public_surface_published") &&
     !isWeekly(sql) &&
-    !isEffort(sql)
+    !isEffort(sql) &&
+    !isOrbGlobal(sql)
   );
 }
 // The reversal read is the only one that inspects engine auto-actions (close/merge) against pull_requests state.