diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index ca89813..50fec2c 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -40,6 +40,29 @@ jobs:
       - name: Test (unit + property + worker + type-level, coverage-gated)
         run: pnpm test -- --coverage
 
+  # ---------------------------------------------------------------------------
+  # Bench regression lane: a dedicated job (noise-isolated from unit CI) asserting each ecsia
+  # iteration path stays under its committed ns/entity RATIO vs a same-run bitECS control. The ratio
+  # cancels shared-runner drift, so a failure is a real regression, not scheduling noise. Single Node
+  # version (timing, not correctness — that's build-test's job).
+  # ---------------------------------------------------------------------------
+  bench-regression:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: pnpm/action-setup@0e279bb959325dab635dd2c09392533439d90093 # v6.0.8
+      - uses: actions/setup-node@v6
+        with:
+          node-version: 24
+          cache: pnpm
+      - run: pnpm install --frozen-lockfile
+      - name: Build
+        run: pnpm build
+      - name: Bench regression (ratios vs bitECS control)
+        env:
+          BENCH_REGRESSION: '1'
+        run: pnpm vitest run --project bench bench/test/regression.bench.test.ts
+
   # ---------------------------------------------------------------------------
   # Runtime lanes (P3): prove the SHIPPED dist actually runs on each claimed runtime.
   # Each lane builds first (the smoke imports packages/ecsia/dist), then runs the SAME
diff --git a/bench/regression-baseline.json b/bench/regression-baseline.json
new file mode 100644
index 0000000..22bce4e
--- /dev/null
+++ b/bench/regression-baseline.json
@@ -0,0 +1,8 @@
+{
+  "_comment": "CI bench regression ceilings — MAX allowed ns/entity RATIO of each ecsia path vs a SAME-RUN bitECS control (so machine drift cancels). A real regression (e.g. codegen breaking → bindColumns deopts from ~0.72x to ~1.5x) trips the ceiling; ~10% run-to-run noise does not. RATCHET: when a path durably improves, lower its ceiling here. Measured 2026-06-08: bindColumns ~0.72x, eachChunk ~1.08x, each ~7.4x.",
+  "ratiosVsBitecs": {
+    "bindColumns": 0.9,
+    "eachChunk": 1.3,
+    "each": 9.0
+  }
+}
diff --git a/bench/test/regression.bench.test.ts b/bench/test/regression.bench.test.ts
new file mode 100644
index 0000000..3410fb8
--- /dev/null
+++ b/bench/test/regression.bench.test.ts
@@ -0,0 +1,63 @@
+// CI bench REGRESSION lane. Times each ecsia iteration path against a SAME-RUN bitECS control and
+// asserts the ns/entity RATIO stays under a committed ceiling (bench/regression-baseline.json). The
+// ratio cancels machine drift — a noisy shared runner moves both ecsia and bitECS together — so a
+// failure means a genuine regression (e.g. codegen breaking and bindColumns deopting to ~1.5x), not
+// scheduling noise. Gated behind BENCH_REGRESSION=1 so it runs ONLY in its dedicated CI job, never
+// in the default `pnpm test` (where measurement noise would flake unit CI). Ratchet ceilings down in
+// the baseline file when a path durably improves.
+
+import { describe, expect, test } from 'vitest'
+import { readFileSync } from 'node:fs'
+import { fileURLToPath } from 'node:url'
+import { makeEcsiaIter, makeEcsiaCursorIter, makeEcsiaPinnedIter, makeBitEcsIter } from '../iterate.js'
+import type { IterCase } from '../iterate.js'
+
+const ENABLED = process.env['BENCH_REGRESSION'] === '1'
+const N = 50_000
+const WARMUP = 300
+const TIMED = 1500
+const REPS = 3 // best-of-N rounds (each round rebuilds + re-warms) to shake off a single bad schedule
+
+interface CtxIter extends IterCase {
+  step(): void
+}
+
+/** p50 ns/entity over TIMED samples, taking the best (min) p50 across REPS rebuilds. */
+function nsPerEntity(make: (n: number) => CtxIter): number {
+  let best = Infinity
+  for (let rep = 0; rep < REPS; rep++) {
+    const c = make(N)
+    for (let i = 0; i < WARMUP; i++) c.step()
+    const s: number[] = []
+    for (let r = 0; r < TIMED; r++) {
+      const t0 = performance.now()
+      c.step()
+      s.push(performance.now() - t0)
+    }
+    s.sort((a, b) => a - b)
+    const p50 = (s[s.length >> 1] as number) * 1e6 / N
+    if (p50 < best) best = p50
+  }
+  return best
+}
+
+const baseline = JSON.parse(
+  readFileSync(fileURLToPath(new URL('../regression-baseline.json', import.meta.url)), 'utf8'),
+) as { ratiosVsBitecs: Record<string, number> }
+
+describe.skipIf(!ENABLED)('bench regression — ecsia/bitECS ns/entity ratios under ceiling', { timeout: 120_000 }, () => {
+  // ONE bitECS control measured in the same process/run as the ecsia paths below.
+  const bit = nsPerEntity(makeBitEcsIter)
+
+  test.each([
+    ['bindColumns', makeEcsiaPinnedIter as (n: number) => CtxIter],
+    ['eachChunk', makeEcsiaCursorIter as (n: number) => CtxIter],
+    ['each', makeEcsiaIter as (n: number) => CtxIter],
+  ])('%s ratio vs bitECS stays under its ceiling', (name, make) => {
+    const ns = nsPerEntity(make)
+    const ratio = ns / bit
+    const ceiling = baseline.ratiosVsBitecs[name] as number
+    // Report the actual ratio in the assertion message so a CI failure shows the regression size.
+    expect(ratio, `${name}: ${ns.toFixed(2)} ns/e = ${ratio.toFixed(3)}x bitECS (${bit.toFixed(2)} ns/e); ceiling ${ceiling}x`).toBeLessThanOrEqual(ceiling)
+  })
+})