From 1bc2c39ca8d47b68c26302dfb0052d277702f6d3 Mon Sep 17 00:00:00 2001
From: Patrick Buckley <eous@users.noreply.github.com>
Date: Wed, 29 Jul 2026 07:36:28 -0700
Subject: [PATCH] fix(eval): stop pinning temperature and reasoning effort on
 the wire
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The eval CLI defaulted temperature to 0.7 and reasoning effort to
medium, so every sweep sent code-chosen sampling knobs the house
assignment scheme forbids — the wire should omit the fields and let
the alias / stored setting / serving default apply, as production
does. Both flags now default to unset and the harnesses plumb None
through to model_turn, whose provider layer already omits absent
knobs. Absolute numbers from earlier sweeps were collected under the
pinned values; contrasts were at least uniform under the same pin.
---
 turnstone/eval/cli.py    | 16 ++++++++++++----
 turnstone/eval/core.py   | 20 ++++++++++----------
 turnstone/eval/nudges.py | 12 ++++++------
 3 files changed, 28 insertions(+), 20 deletions(-)

diff --git a/turnstone/eval/cli.py b/turnstone/eval/cli.py
index e1f0db16..32e2526e 100644
--- a/turnstone/eval/cli.py
+++ b/turnstone/eval/cli.py
@@ -184,8 +184,12 @@ def main() -> None:
     parser.add_argument(
         "--temperature",
         type=float,
-        default=0.7,
-        help="Sampling temperature (default: 0.7)",
+        default=None,
+        help=(
+            "Sampling temperature. Omitted from the wire by default so "
+            "the alias / stored setting / serving default applies — the "
+            "same assignment scheme production runs"
+        ),
     )
     parser.add_argument(
         "--max-tokens",
@@ -195,9 +199,13 @@ def main() -> None:
     )
     parser.add_argument(
         "--reasoning-effort",
-        default="medium",
+        default=None,
         choices=["low", "medium", "high"],
-        help="Reasoning effort (default: medium)",
+        help=(
+            "Reasoning effort. Omitted from the wire by default — a "
+            "code-chosen token is unvetted and can flip thinking on for "
+            "lanes the operator never engaged"
+        ),
     )
     parser.add_argument(
         "--context-window",
diff --git a/turnstone/eval/core.py b/turnstone/eval/core.py
index 2018839d..de4b5bfe 100644
--- a/turnstone/eval/core.py
+++ b/turnstone/eval/core.py
@@ -238,10 +238,10 @@ class HeadlessSession(ChatSession):
         model: str,
         system_prompt_override: str | None = None,
         instructions: str | None = None,
-        temperature: float = 0.7,
+        temperature: float | None = None,
         max_tokens: int = 32768,
         tool_timeout: int = 30,
-        reasoning_effort: str = "medium",
+        reasoning_effort: str | None = None,
         context_window: int = 131072,
         compact_max_tokens: int = 32768,
         auto_compact_pct: float = 0.8,
@@ -651,9 +651,9 @@ def _run_single_test(
     model: str,
     system_prompt: str,
     case: dict[str, Any],
-    temperature: float,
+    temperature: float | None,
     max_tokens: int,
-    reasoning_effort: str,
+    reasoning_effort: str | None,
     context_window: int,
     verbose: bool = False,
     log_prefix: str = "",
@@ -1100,9 +1100,9 @@ def _run_iteration_parallel(
     system_prompt: str,
     cases: list[dict[str, Any]],
     n_runs: int,
-    temperature: float,
+    temperature: float | None,
     max_tokens: int,
-    reasoning_effort: str,
+    reasoning_effort: str | None,
     context_window: int,
     test_timeout: int,
     parallel: int,
@@ -1251,9 +1251,9 @@ def _run_iteration(
     system_prompt: str,
     cases: list[dict[str, Any]],
     n_runs: int,
-    temperature: float,
+    temperature: float | None,
     max_tokens: int,
-    reasoning_effort: str,
+    reasoning_effort: str | None,
     context_window: int,
     verbose: bool = False,
     test_timeout: int = 300,
@@ -1469,9 +1469,9 @@ def run_skill_adherence(
     model: str,
     cases: list[dict[str, Any]],
     n_runs: int,
-    temperature: float,
+    temperature: float | None,
     max_tokens: int,
-    reasoning_effort: str,
+    reasoning_effort: str | None,
     context_window: int,
     test_timeout: int = 300,
     parallel: int = 1,
diff --git a/turnstone/eval/nudges.py b/turnstone/eval/nudges.py
index 5b562b70..d7292ac1 100644
--- a/turnstone/eval/nudges.py
+++ b/turnstone/eval/nudges.py
@@ -517,9 +517,9 @@ class CoordinatorHeadlessSession(HeadlessSession):
         coord_client: CoordinatorClient,
         ws_id: str,
         user_id: str,
-        temperature: float,
+        temperature: float | None,
         max_tokens: int,
-        reasoning_effort: str,
+        reasoning_effort: str | None,
         context_window: int,
     ) -> None:
         super().__init__(
@@ -981,9 +981,9 @@ def _run_single_nudge(
     model: str,
     case: dict[str, Any],
     arm: str,
-    temperature: float,
+    temperature: float | None,
     max_tokens: int,
-    reasoning_effort: str,
+    reasoning_effort: str | None,
     context_window: int,
     max_turns: int,
     test_timeout: int,
@@ -2019,9 +2019,9 @@ def run_nudge_response(
     model: str,
     cells: list[dict[str, Any]],
     n_runs: int = 10,
-    temperature: float = 0.7,
+    temperature: float | None = None,
     max_tokens: int = 8192,
-    reasoning_effort: str = "medium",
+    reasoning_effort: str | None = None,
     context_window: int = 131072,
     max_turns: int = 8,
     test_timeout: int = 300,