diff --git a/.github/bench/README.md b/.github/bench/README.md
index ca456f0b..a9e4d742 100644
--- a/.github/bench/README.md
+++ b/.github/bench/README.md
@@ -51,9 +51,15 @@ entries, and the unit tests pin the semantics.
 
 1. Dispatch **Benchmark A/B** with `mode=aa` (same ref twice) at least 3
    times; each posts a noise report to the run summary and artifact.
-2. B/op floor must be ~0. If it isn't, the reduced profile has too little
-   warmup for C2/escape analysis — raise `-wi` in `JMH_FLAGS` /
-   `JMH_PROFILE` in bench-pr.yml and recalibrate.
+2. B/op floor must be ~0. If it isn't: first pin the fork JVM's
+   allocation ergonomics — adaptive TLAB sizing makes `gc.alloc.rate.norm`
+   count environment-dependent retire waste (measured on PR #116:
+   ±78-81% A/A swings on `PlatedBench.visitorUniverseJson n=4096`
+   unpinned, ±0.0% with `-jvmArgsAppend -XX:-ResizeTLAB` plus a fixed
+   `-Xms`/`-Xmx`; fork counts do NOT fix it — `-f 3` unpinned still
+   swung ±11-41%). Then raise `-wi` in `JMH_FLAGS` / `JMH_PROFILE` in
+   bench-pr.yml and recalibrate. Changing the flags changes the profile —
+   see step 4.
 3. Commit `.github/bench/thresholds.json`, e.g.
    `{"bop_regression_pct": 1.0, "bop_min_delta_bytes": 16}` — quantile
    data from the reports, not guesses.
diff --git a/.github/bench/bench_tools.py b/.github/bench/bench_tools.py
index ecefaf40..b78b2063 100644
--- a/.github/bench/bench_tools.py
+++ b/.github/bench/bench_tools.py
@@ -488,8 +488,11 @@ def q(vals, frac):
         )
     lines += [
         "",
-        "B/op max should be ~0. If it isn't, the reduced profile has too "
-        "little warmup for C2/escape analysis — bump `-wi` before any gate.",
+        "B/op max should be ~0. If it isn't: pin the fork JVM's allocation "
+        "ergonomics first (-jvmArgsAppend -XX:-ResizeTLAB plus a fixed "
+        "-Xms/-Xmx — adaptive TLAB sizing makes B/op count "
+        "environment-dependent retire waste), then bump `-wi` before any "
+        "gate.",
     ]
     return "\n".join(lines)
 
diff --git a/.github/workflows/bench-pr.yml b/.github/workflows/bench-pr.yml
index 93cb4fec..aac5578c 100644
--- a/.github/workflows/bench-pr.yml
+++ b/.github/workflows/bench-pr.yml
@@ -54,8 +54,16 @@ env:
   # The measurement profile. thresholds.json is calibrated against exactly
   # this string — change it and the gate reverts to advisory until an A/A
   # recalibration (delete thresholds.json in the same PR).
-  JMH_PROFILE: "pr:-i3-wi2-f1-t1-gc"
-  JMH_FLAGS: "-i 3 -wi 2 -f 1 -t 1 -foe true -prof gc -rf json"
+  #
+  # Pinned allocation ergonomics (-XX:-ResizeTLAB + fixed heap): adaptive
+  # TLAB sizing makes gc.alloc.rate.norm count environment-dependent retire
+  # waste — PR #116 measured A/A swings of +78-81% B/op on
+  # PlatedBench.visitorUniverseJson n=4096 across identical-source runs;
+  # pinning took the same A/A to ±0.0%. Forks don't fix it (measured: -f 3
+  # unpinned still ±11-41% — it is per-invocation environment state, not
+  # per-fork sampling noise).
+  JMH_PROFILE: "pr:-i3-wi2-f1-t1-gc-rt1g"
+  JMH_FLAGS: "-i 3 -wi 2 -f 1 -t 1 -foe true -prof gc -rf json -jvmArgsAppend -XX:-ResizeTLAB -jvmArgsAppend -Xms1g -jvmArgsAppend -Xmx1g"
 
 jobs:
   bench-ab:
diff --git a/.github/workflows/bench-sweep.yml b/.github/workflows/bench-sweep.yml
index c2d025d1..5b74994a 100644
--- a/.github/workflows/bench-sweep.yml
+++ b/.github/workflows/bench-sweep.yml
@@ -58,8 +58,11 @@ concurrency:
   cancel-in-progress: false
 
 env:
-  JMH_PROFILE: "sweep:-i5-wi3-f3-t1-gc"
-  JMH_FLAGS: "-i 5 -wi 3 -f 3 -t 1 -foe true -prof gc -rf json"
+  # Pinned allocation ergonomics — see bench-pr.yml: adaptive TLAB sizing
+  # makes B/op count environment-dependent retire waste; -XX:-ResizeTLAB +
+  # a fixed heap made the flagged families' A/A B/op deterministic.
+  JMH_PROFILE: "sweep:-i5-wi3-f3-t1-gc-rt1g"
+  JMH_FLAGS: "-i 5 -wi 3 -f 3 -t 1 -foe true -prof gc -rf json -jvmArgsAppend -XX:-ResizeTLAB -jvmArgsAppend -Xms1g -jvmArgsAppend -Xmx1g"
 
 jobs:
   sweep: