diff --git a/benchmarks/ai-gateway/ai-gateway.bench.ts b/benchmarks/ai-gateway/ai-gateway.bench.ts
index ac2160b3..b504abe7 100644
--- a/benchmarks/ai-gateway/ai-gateway.bench.ts
+++ b/benchmarks/ai-gateway/ai-gateway.bench.ts
@@ -1,11 +1,12 @@
 /**
- * AI Gateway benchmark, built on @benchsdk/runner's runBenchmark with
- * `groupBy: 'round'`. Fairness is the whole point (see AI_GATEWAYS.md): every
- * gateway's Nth iteration must run at roughly the same point in time as every
- * other gateway's Nth iteration, so no gateway is favored by running during a
- * different network condition. `groupBy: 'round'` provides exactly that — one
- * task per gateway per round, taking turns — replacing the bespoke round-robin
- * loop this file used to hand-roll.
+ * AI Gateway benchmark. Declarative — exports `config` + `task`; `bench run`
+ * owns the entrypoint. Configured with `groupBy: 'round'`. Fairness is the
+ * whole point (see AI_GATEWAYS.md): every gateway's Nth iteration must run at
+ * roughly the same point in time as every other gateway's Nth iteration, so no
+ * gateway is favored by running during a different network condition.
+ * `groupBy: 'round'` provides exactly that — one task per gateway per round,
+ * taking turns — replacing the bespoke round-robin loop this file used to
+ * hand-roll.
  *
  * Declared as two phases (cold, warm); the framework owns the phase boundary
  * and exposes it via `ctx.phase`, so the task branches on phase identity, not
diff --git a/benchmarks/browser/browser-task.ts b/benchmarks/browser/browser-task.ts
index 442db779..749a7dd7 100644
--- a/benchmarks/browser/browser-task.ts
+++ b/benchmarks/browser/browser-task.ts
@@ -2,8 +2,7 @@
  * Shared browser-lifecycle workload for the browser benchmark. One iteration =
  * create a session, connect over CDP, navigate to example.com, then release
  * (always, even on failure). Orchestration (how many, how parallel) is owned by
- * @benchsdk/runner's runBenchmark — this file only describes what one iteration
- * does.
+ * @benchsdk/runner — this file only describes what one iteration does.
  */
 import { chromium } from 'playwright-core';
 import type { BenchmarkTask, TaskContext, TaskResult } from '@benchsdk/runner';
diff --git a/benchmarks/sandbox/dax.bench.ts b/benchmarks/sandbox/dax.bench.ts
index 6f57e46a..28eb63bc 100644
--- a/benchmarks/sandbox/dax.bench.ts
+++ b/benchmarks/sandbox/dax.bench.ts
@@ -1,11 +1,12 @@
 /**
  * Dax benchmark: runs the OpenCode build (scripts/dax-benchmark.sh) inside a
  * freshly created sandbox once per iteration and measures its build phases
- * (prepare / bun-download / bun-unpack / clone / install / typecheck). Config
- * lives here; platform orchestration is owned by @benchsdk/runner's runBenchmark.
- * The phase-parsing + resource-sizing logic is reused from ./dax.ts so the
- * legacy `results/sandbox-dax/` JSON shape is preserved verbatim via the
- * legacy-results adapter (TEMPORARY local-JSON bridge — see legacy-results).
+ * (prepare / bun-download / bun-unpack / clone / install / typecheck).
+ * Declarative — exports `config` + `task`; `bench run` owns the entrypoint and
+ * platform orchestration. The phase-parsing + resource-sizing logic is reused
+ * from ./dax.ts so the legacy `results/sandbox-dax/` JSON shape is preserved
+ * verbatim via the legacy-results adapter (TEMPORARY local-JSON bridge — see
+ * legacy-results).
  *
  * `groupBy: 'round'` is used so the task owns its own measured `latencyMs`
  * (the build's totalMs) and can attach pre-measured phase steps; the local
diff --git a/benchmarks/storage/snapshot-fork-task.ts b/benchmarks/storage/snapshot-fork-task.ts
index 9b93e187..de17881d 100644
--- a/benchmarks/storage/snapshot-fork-task.ts
+++ b/benchmarks/storage/snapshot-fork-task.ts
@@ -3,8 +3,8 @@
  *   seed dataset -> snapshot -> fork(from snapshot) -> fork(from live) ->
  *   read-back-from-fork (verify) -> teardown.
  * Every created resource is torn down in a `finally` so a mid-iteration failure
- * does not leak real storage. Orchestration is owned by @benchsdk/runner's
- * runBenchmark — this file only describes what one iteration does.
+ * does not leak real storage. Orchestration is owned by @benchsdk/runner —
+ * this file only describes what one iteration does.
  */
 import crypto from 'node:crypto';
 import type { Storage } from '@storagesdk/core';
diff --git a/benchmarks/storage/snapshot-fork.bench.ts b/benchmarks/storage/snapshot-fork.bench.ts
index 6548a524..666bdd54 100644
--- a/benchmarks/storage/snapshot-fork.bench.ts
+++ b/benchmarks/storage/snapshot-fork.bench.ts
@@ -1,8 +1,8 @@
 /**
  * Storage snapshot/fork benchmark: per-iteration seed -> snapshot -> fork ->
  * verify, per provider (concurrency 1 = sequential; each iteration creates real
- * snapshots/forks). Config lives here; orchestration is owned by @benchsdk/runner's
- * runBenchmark.
+ * snapshots/forks). Declarative — exports `config` + `task`; `bench run` owns
+ * the entrypoint and orchestration.
  *
  *   bench run benchmarks/storage/snapshot-fork.bench.ts
  *   bench run benchmarks/storage/snapshot-fork.bench.ts --dataset wide --iterations 5 --provider tigris
diff --git a/benchmarks/storage/storage-task.ts b/benchmarks/storage/storage-task.ts
index e6f4966c..9adc69fa 100644
--- a/benchmarks/storage/storage-task.ts
+++ b/benchmarks/storage/storage-task.ts
@@ -2,8 +2,7 @@
  * Shared storage upload/download workload for the storage benchmark. One
  * iteration = upload a buffer, download it, compute throughput, then delete
  * (best-effort cleanup). Orchestration (how many, how parallel) is owned by
- * @benchsdk/runner's runBenchmark — this file only describes what one iteration
- * does.
+ * @benchsdk/runner — this file only describes what one iteration does.
  */
 import type { BenchmarkTask, TaskContext, TaskResult } from '@benchsdk/runner';
 import { TaskError } from '@benchsdk/runner';