Restore the privileged critic: its ablation flips with scale

The 39.3M ablation said the privileged critic hurt (switching it off won 0.5160
[0.505, 0.527]). Head to head at 131M, both sides trained identically, it says the
opposite: off *loses*, 0.4633 [0.453, 0.474]. Against league the critic-on model
scores 0.6094 and the critic-off one 0.5526.

The critic earns its keep once there is enough data to fit it -- at 39.3M the
privileged value trunk is underfit and only adds advantage noise. Defaulted back
on, with the small-scale number kept in the docstring as a warning: an ablation at
a budget you do not intend to ship can invert.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh

docs/plans/lost-cities-classic-3round.md

@@ -490,3 +490,28 @@ Fable은 이것을 "가장 큰 누락 아이디어"로 꼽았다. 실측은 **
 3. **선형 총점 보상** — tanh 대비 0.5859 (사용자 제안)
 4. **매치 관측 결함 수정** — 약 2.5%p
 5. ~~전지적 critic~~ — **해가 된다. 끈다.**
+
+### 정정: 전지적 critic의 가치는 **스케일에 따라 뒤집힌다**
+
+위 A/B는 39.3M learner 액션 규모에서 돌렸다. 실전 규모(131M)에서 동일 설정으로 둘을 직접
+맞붙이면 **결론이 반대로 나온다.**
+
+| 규모 | critic OFF의 승률 (critic ON 상대) | 판정 |
+|---|---|---|
+| 39.3M (A/B 규모) | **0.5160** [0.505, 0.527] | OFF가 낫다 |
+| **131M (실전 규모)** | **0.4633** [0.453, 0.474] | **ON이 낫다** |
+
+**전지적 critic은 그것을 학습시킬 데이터가 충분해질 때 값어치를 한다.** 39.3M에서는 특권
+정보를 쓰는 큰 가치 트렁크를 제대로 못 맞춰서 어드밴티지에 노이즈만 얹었고, 131M에서는
+제대로 맞춰서 분산 감소가 실현된다.
+
+**교훈: 배포할 예산이 아닌 규모에서 ablation을 돌리면 결론이 뒤집힐 수 있다.**
+`privileged_critic` 기본값은 **ON으로 되돌린다.**
+
+### 최종 순위 (진짜 3라운드 클래식, duplicate 8,192판)
+
+| 모델 | learner 액션 | vs league |
+|---|---|---|
+| **매치 스택 (critic ON)** | 131M | **0.6094** (+22.0점) |
+| 매치 스택 (critic OFF) | 131M | 0.5526 (+10.9점) |
+| league (기존 최강, 웹 배포판) | 122.6M | — |

src/lost_cities_jax/match_ppo.py

@@ -89,14 +89,17 @@ N_SEATS = 2
 class Ablation:
     """Switches for measuring what each piece of the match stack is worth."""
 
-    privileged_critic: bool = False
+    privileged_critic: bool = True
     """On: the critic also sees the opponent's hand and the deck in order.
 
-    Defaults OFF because it measured *negative*. Ablated against the full stack
-    over 8192 duplicate matches, switching it off won 0.5160 [0.505, 0.527]. A
-    critic that knows the deck fits V(full state), which is not
-    E[return | masked obs], so the advantage carries a component the actor cannot
-    act on -- noise from its side, not variance reduction.
+    Its worth flips with scale, so do not trust a cheap ablation here. Ablated at
+    39.3M learner actions, switching it OFF *won* 0.5160 [0.505, 0.527] -- the
+    obvious reading being that a critic fitting V(full state) rather than
+    E[return | masked obs] hands the actor advantage noise it cannot act on. But
+    head to head at 131M, both sides trained identically, OFF *loses*: 0.4633
+    [0.453, 0.474]. The privileged critic earns its keep once there is enough data
+    to fit it. Defaults on for that reason, and the small-scale result is kept as
+    a warning about ablating at a budget you do not intend to ship.
     """
 
     both_seats: bool = True

tests/lost_cities_jax/test_match_ppo.py

@@ -13,6 +13,7 @@ from lost_cities_jax.match_obs import (
     match_observation,
 )
 from lost_cities_jax.match_ppo import (
+    Ablation,
     create_match_train_state,
     make_match_rollout_fn,
     make_match_train_iteration,
@@ -54,9 +55,14 @@ def test_the_critic_sees_the_opponents_hand_and_the_actor_does_not():
 
 
 def test_privileged_input_cannot_move_the_policy_logits():
-    """Separate trunks, or the critic's view of the deck leaks into play."""
+    """Separate trunks, or the critic's view of the deck leaks into play.
+
+    The privileged critic is off by default -- it measured negative, see the
+    ablation in the plan -- but the isolation property still has to hold for
+    anyone who switches it on, and for its later reuse as a search evaluator.
+    """
     cfg = _cfg()
-    state = create_match_train_state(cfg, jax.random.PRNGKey(1))
+    state = create_match_train_state(cfg, jax.random.PRNGKey(1), Ablation(privileged_critic=True))
     match = match_reset(jax.random.PRNGKey(2))
 
     obs = match_observation(match, jnp.int32(0))[None, :]