diff --git a/test/entity-card-perf.slow.test.ts b/test/entity-card-perf.slow.test.ts
index 831317be1..928bf1647 100644
--- a/test/entity-card-perf.slow.test.ts
+++ b/test/entity-card-perf.slow.test.ts
@@ -12,12 +12,15 @@
  *      loosen in CI only with evidence of runner noise). The protocol DOC
  *      promises this number; the bound is op-layer latency (transport
  *      excluded, as documented).
- *   2. RATIO GUARD (machine-independent) — entity p99 ≤ 50× max(getPage p50,
+ *   2. RATIO GUARD (machine-independent) — entity p99 ≤ 100× max(getPage p50,
  *      1ms) on the same corpus. Calibration: the card is ~7 indexed reads +
- *      a keyword search on the miss path, measured ~21× a 1ms-floored
- *      getPage at 20K pages — an O(N) scan regression lands at 200ms+
- *      (≥200×), far past the ceiling even on a slow runner, while the
- *      2.4× headroom absorbs planner noise.
+ *      a keyword search on the miss path. It measures ~21× a getPage p50 of
+ *      ~2.5ms, but on a fast runner getPage p50 floors to 1ms and normal
+ *      entity p99 (~50ms) reads as ~50×. An O(N) scan regression lands at
+ *      200ms+ (≥200×), far past the ceiling even on a slow runner. The ceiling
+ *      is 100× (not 50×) so the guard is never STRICTER than the 100ms absolute
+ *      budget when getPage floors to 1ms — the earlier 50× tripped on fast
+ *      runners (a p99 tail ÷ a sub-ms median) while p99 stayed well under budget.
  *
  * The 200K-page validation is a documented MANUAL recipe in
  * docs/protocol/MEMORY_VERBS_v1.md — not CI-gated (seed time would dominate).
@@ -40,8 +43,12 @@ const MEASURED = 200;
 const TARGET_ENTITIES = 50; // pages the measured calls rotate over
 
 const P99_BUDGET_MS = 100 * (Number(process.env.GBRAIN_PERF_BUDGET_MULTIPLIER) || 1);
-// entity p99 ≤ 50× max(getPage p50, 1ms) — see the calibration note above.
-const RATIO_CEILING = 50;
+// entity p99 ≤ 100× max(getPage p50, 1ms) — see the calibration note above.
+// (100×, not 50×: at the 1ms getPage floor, 50× would cap p99 at 50ms — stricter
+// than the 100ms absolute budget — and tripped on fast runners where a p99 tail
+// is divided by a sub-ms getPage median. 100× stays far below the ≥200× O(N)
+// regression signal.)
+const RATIO_CEILING = 100;
 
 function percentile(sorted: number[], p: number): number {
   const idx = Math.min(sorted.length - 1, Math.ceil((p / 100) * sorted.length) - 1);