fix: require six times P99 for sustained signals (#3204)

convex/lib/publisherAbuseScoring.test.ts

@@ -449,24 +449,27 @@ describe("publisher abuse scoring", () => {
     expect(score.reasonCodes).toContain("temporal_download_spike_flat_installs");
   });
 
-  it("flags sustained high downloads with flat installs", () => {
+  it("flags sustained flat-install traffic one download above six times platform P99", () => {
     const todayDay = 100;
     const score = computeCurrentSkillTemporalAbuseScore({
       todayDay,
       benchmark: temporalBenchmark({
-        downloads30dP95: 3_000,
-        downloads30dP99: 3_000,
+        downloads30dP95: 284,
+        downloads30dP99: 600,
         spikeMultiplier7dP95: 20,
         spikeMultiplier7dP99: 50,
       }),
-      dailyStats: dailyRange(71, 30, { downloads: 120, installs: 0 }),
+      dailyStats: [
+        ...dailyRange(71, 29, { downloads: 120, installs: 0 }),
+        { day: 100, downloads: 121, installs: 0 },
+      ],
     });
 
     expect(score.spike).toBe(false);
     expect(score.sustained).toBe(true);
-    expect(score.recent30Downloads).toBe(3_600);
+    expect(score.recent30Downloads).toBe(3_601);
     expect(score.recent30Installs).toBe(0);
-    expect(score.downloadInstallRatio30).toBe(3_600);
+    expect(score.downloadInstallRatio30).toBe(3_601);
     expect(score.downloads30dCohortBand).toBe("p99");
     expect(score.reasonCodes).toContain("temporal_sustained_downloads_flat_installs");
   });
@@ -486,17 +489,17 @@ describe("publisher abuse scoring", () => {
     expect(score.downloads30dCohortBand).toBeUndefined();
   });
 
-  it("keeps sub-3000 P99 download traffic below the absolute review floor", () => {
+  it("keeps downloads at exactly six times platform P99 below the sustained threshold", () => {
     const score = computeCurrentSkillTemporalAbuseScore({
       todayDay: 100,
       benchmark: temporalBenchmark({
-        downloads30dP95: 1_000,
-        downloads30dP99: 2_000,
+        downloads30dP95: 284,
+        downloads30dP99: 600,
       }),
-      dailyStats: dailyRange(71, 30, { downloads: 90, installs: 0 }),
+      dailyStats: dailyRange(71, 30, { downloads: 120, installs: 0 }),
     });
 
-    expect(score.recent30Downloads).toBe(2_700);
+    expect(score.recent30Downloads).toBe(3_600);
     expect(score.sustained).toBe(false);
     expect(score.downloads30dCohortBand).toBeUndefined();
   });
@@ -611,8 +614,8 @@ describe("publisher abuse scoring", () => {
   it("finds historical spike and sustained windows for backfill scans", () => {
     const score = computeHistoricalSkillTemporalAbuseScore({
       benchmark: temporalBenchmark({
-        downloads30dP95: 3_000,
-        downloads30dP99: 10_000,
+        downloads30dP95: 284,
+        downloads30dP99: 1_000,
         spikeMultiplier7dP95: 5,
         spikeMultiplier7dP99: 25,
       }),

convex/lib/publisherAbuseScoring.ts

@@ -128,7 +128,7 @@ const TEMPORAL_SUSTAINED_DAYS = 30;
 const TEMPORAL_MAX_SPIKE_INSTALLS = 2;
 const TEMPORAL_MAX_SUSTAINED_INSTALLS = 5;
 const TEMPORAL_MIN_SPIKE_7_DOWNLOADS = 2_000;
-const TEMPORAL_MIN_SUSTAINED_30_DOWNLOADS = 3_000;
+const TEMPORAL_SUSTAINED_DOWNLOADS_P99_MULTIPLIER = 6;
 const TEMPORAL_MIN_BASELINE_7_DOWNLOADS = 100;
 const TEMPORAL_MIN_NEAR_CONVERSION_7_DOWNLOADS = 500;
 const TEMPORAL_MIN_NEAR_CONVERSION_30_DOWNLOADS = 1_000;
@@ -437,7 +437,8 @@ export function classifySkillTemporalAbuseScore(
     score.spikeMultiplier / Math.max(1, benchmark.spikeMultiplier7dP95);
   const downloads30dCohortBand =
     score.recent30Installs <= TEMPORAL_MAX_SUSTAINED_INSTALLS &&
-    score.recent30Downloads >= TEMPORAL_MIN_SUSTAINED_30_DOWNLOADS
+    score.recent30Downloads >
+      benchmark.downloads30dP99 * TEMPORAL_SUSTAINED_DOWNLOADS_P99_MULTIPLIER
       ? p99Band({ value: score.recent30Downloads, p99: benchmark.downloads30dP99 })
       : undefined;
   const spikeMultiplierCohortBand =

convex/publisherAbuseTemporalScan.test.ts

@@ -655,8 +655,8 @@ describe("scheduled temporal publisher abuse scan", () => {
       sampleSize: 1_000,
       downloads30dAverage: 180,
       downloads30dMedian: 45,
-      downloads30dP95: 900,
-      downloads30dP99: 3_000,
+      downloads30dP95: 284,
+      downloads30dP99: 600,
       spikeMultiplier7dP95: 4,
       spikeMultiplier7dP99: 12,
     };
@@ -667,7 +667,7 @@ describe("scheduled temporal publisher abuse scan", () => {
     });
     const candidate = temporalCandidate(
       "skills:anysearch" as Id<"skills">,
-      temporalScore({ recent30Downloads: 3_370, recent30Installs: 4 }),
+      temporalScore({ recent30Downloads: 3_700, recent30Installs: 4 }),
     );
     const runQuery = vi
       .fn()

specs/security-moderation.md

@@ -94,10 +94,10 @@ See also: [acceptable-usage.md](./acceptable-usage.md) for the marketplace polic
   Partial scans must not archive signals or present their top-download slice as
   a platform percentile.
 - Flat-install temporal review signals are deliberately high-confidence:
-  sustained volume must exceed the platform P99, reach at least 3,000 downloads
-  in 30 days, and have at most 5 installs; a spike must exceed the platform P99,
-  reach at least 2,000 downloads in 7 days, and have at most 2 installs. These
-  signals indicate anomalous traffic for manual review, not publisher attribution.
+  sustained volume must exceed six times the platform 30-day download P99 and
+  have at most 5 installs; a spike must exceed the platform P99, reach at least
+  2,000 downloads in 7 days, and have at most 2 installs. These signals indicate
+  anomalous traffic for manual review, not publisher attribution.
 - Publisher abuse scoring must skip staff-linked and official publishers before
   nominations are created. Publisher abuse autoban must process pending
   `potential_ban_candidate` pressure nominations without waiting for the score