vajeeda's picture
feat(phase10): ABScriptEnv, ContrastiveReward, A/B rollout fn, 25 tests PASS, gate PASS
cfe83fc
Raw
History Blame Contribute Delete
9.98 kB
from __future__ import annotations
import json
import random
import uuid
from typing import Optional, Tuple
from viral_script_engine.environment.env import ViralScriptEnv
from viral_script_engine.environment.trajectory import Trajectory, TrajectoryType
from viral_script_engine.rewards.contrastive_reward import ContrastiveReward
class ABScriptEnv:
"""
A/B Testing wrapper around ViralScriptEnv.
Each episode runs TWO parallel trajectories from the same starting script:
- Trajectory A (critic_first): forced to act on Critic's top claim in step 1
- Trajectory B (defender_first): forced to act on Defender's concern in step 1
- Steps 2+ are free — the Arbitrator makes its own decisions in both
The Arbitrator observes BOTH trajectories in the state() output.
The contrastive reward fires at episode end based on the delta.
This teaches the Arbitrator: "I could have done X first or Y first.
One led to a better outcome. Learn which one."
"""
def __init__(
self,
scripts_path: str = "data/test_scripts/scripts.json",
cultural_kb_path: str = "data/cultural_kb.json",
max_steps: int = 5,
difficulty: str = "easy",
):
self.env_a = ViralScriptEnv(
scripts_path=scripts_path,
cultural_kb_path=cultural_kb_path,
max_steps=max_steps,
difficulty=difficulty,
use_escalation=False,
use_anti_gaming=False,
)
self.env_b = ViralScriptEnv(
scripts_path=scripts_path,
cultural_kb_path=cultural_kb_path,
max_steps=max_steps,
difficulty=difficulty,
use_escalation=False,
use_anti_gaming=False,
)
self.contrastive_reward_calc = ContrastiveReward()
self._traj_a: Optional[Trajectory] = None
self._traj_b: Optional[Trajectory] = None
self._episode_id: Optional[str] = None
self._step_num: int = 0
self._forced_action_a: Optional[dict] = None
self._forced_action_b: Optional[dict] = None
# ------------------------------------------------------------------
# Public API
# ------------------------------------------------------------------
def reset(self, seed=None, options=None) -> dict:
"""
Reset BOTH environments with the SAME script and seed.
Run step 1 automatically with the forced actions.
Return the state after forced step 1.
"""
if seed is None:
seed = random.randint(0, 2 ** 31)
self._episode_id = str(uuid.uuid4())
self._step_num = 0
obs_a, _ = self.env_a.reset(seed=seed)
obs_b, _ = self.env_b.reset(seed=seed)
return self._run_forced_step_1(obs_a, obs_b)
def reset_from_script_id(self, script_id: str, scripts_path: str) -> dict:
"""Reset both environments to a specific script by ID."""
with open(scripts_path) as f:
all_scripts = json.load(f)
script = next((s for s in all_scripts if s["script_id"] == script_id), None)
if script is None:
raise ValueError(f"Script {script_id!r} not found in {scripts_path}")
self._episode_id = str(uuid.uuid4())
self._step_num = 0
episode_config = {
"script_id": script["script_id"],
"script_text": script["script_text"],
"region": script["region"],
"platform": script["platform"],
"niche": script["niche"],
"difficulty": script.get("difficulty", "hard"),
}
obs_a, _ = self.env_a.reset_from_config(episode_config)
obs_b, _ = self.env_b.reset_from_config(episode_config)
return self._run_forced_step_1(obs_a, obs_b)
def step(self, action: dict) -> Tuple[dict, float, bool, bool, dict]:
"""
Execute the action in BOTH environments simultaneously (step 2+).
Same action applied to both trajectories.
Returns combined observation with both trajectory states.
Terminated when BOTH trajectories have reached max_steps.
"""
if self._traj_a is None or self._traj_b is None:
raise RuntimeError("Call reset() before step()")
if not self._traj_a.terminated:
obs_a, r_a, done_a, _, info_a = self.env_a.step(action)
self._traj_a.current_script = obs_a.get(
"current_script", self._traj_a.current_script
)
self._traj_a.cumulative_reward += r_a
self._traj_a.step_count += 1
self._traj_a.terminated = done_a
self._traj_a.final_reward_components = info_a.get("reward_components")
if not self._traj_b.terminated:
obs_b, r_b, done_b, _, info_b = self.env_b.step(action)
self._traj_b.current_script = obs_b.get(
"current_script", self._traj_b.current_script
)
self._traj_b.cumulative_reward += r_b
self._traj_b.step_count += 1
self._traj_b.terminated = done_b
self._traj_b.final_reward_components = info_b.get("reward_components")
self._step_num += 1
terminated = self._traj_a.terminated and self._traj_b.terminated
episode_reward = 0.0
if terminated:
result = self.contrastive_reward_calc.compute(self._traj_a, self._traj_b)
episode_reward = result.final_reward
return self.state(), episode_reward, terminated, False, {}
def state(self) -> dict:
"""
Returns state showing both trajectories:
{
"trajectory_a": { current_script, reward_components, debate_history,
cumulative_reward, step_count, terminated, trajectory_type },
"trajectory_b": { ... },
"delta": traj_a.cumulative_reward - traj_b.cumulative_reward,
"leading_trajectory": "A" or "B",
"step_num": current step,
"episode_id": ...
}
"""
if self._traj_a is None or self._traj_b is None:
return {}
delta = self._traj_a.cumulative_reward - self._traj_b.cumulative_reward
leading = "A" if delta >= 0 else "B"
return {
"trajectory_a": self._traj_state(self.env_a, self._traj_a),
"trajectory_b": self._traj_state(self.env_b, self._traj_b),
"delta": delta,
"leading_trajectory": leading,
"step_num": self._step_num,
"episode_id": self._episode_id,
}
def reward(self) -> float:
"""Called at episode end — returns the contrastive reward."""
if self._traj_a is None or self._traj_b is None:
return 0.0
result = self.contrastive_reward_calc.compute(self._traj_a, self._traj_b)
return result.final_reward
# ------------------------------------------------------------------
# Internal helpers
# ------------------------------------------------------------------
def _run_forced_step_1(self, obs_a: dict, obs_b: dict) -> dict:
"""
After both envs are reset, run step 1 with forced actions and
initialise the Trajectory objects.
"""
initial_script = obs_a.get("current_script", "")
region = obs_a.get("region", "pan_india_english")
platform = obs_a.get("platform", "Reels")
niche = obs_a.get("niche", "personal finance")
self._traj_a = Trajectory(
trajectory_id=f"{self._episode_id}_A",
trajectory_type=TrajectoryType.CRITIC_FIRST,
initial_script=initial_script,
current_script=initial_script,
)
self._traj_b = Trajectory(
trajectory_id=f"{self._episode_id}_B",
trajectory_type=TrajectoryType.DEFENDER_FIRST,
initial_script=initial_script,
current_script=initial_script,
)
# Run critic and defender once to determine forced actions
critique = self.env_a.critic.critique(
script=initial_script,
region=region,
platform=platform,
niche=niche,
)
defender_out = self.env_a.defender.defend(
script=initial_script,
critic_claims=critique.claims,
region=region,
platform=platform,
)
forced_a = self._traj_a.get_forced_first_action(critique.claims, defender_out)
forced_b = self._traj_b.get_forced_first_action(critique.claims, defender_out)
self._forced_action_a = forced_a
self._forced_action_b = forced_b
# Execute forced step 1 in each environment
obs_a_new, r_a, done_a, _, info_a = self.env_a.step(forced_a)
obs_b_new, r_b, done_b, _, info_b = self.env_b.step(forced_b)
self._traj_a.current_script = obs_a_new.get("current_script", initial_script)
self._traj_a.cumulative_reward = r_a
self._traj_a.step_count = 1
self._traj_a.terminated = done_a
self._traj_a.final_reward_components = info_a.get("reward_components")
self._traj_b.current_script = obs_b_new.get("current_script", initial_script)
self._traj_b.cumulative_reward = r_b
self._traj_b.step_count = 1
self._traj_b.terminated = done_b
self._traj_b.final_reward_components = info_b.get("reward_components")
self._step_num = 1
return self.state()
def _traj_state(self, env: ViralScriptEnv, traj: Trajectory) -> dict:
s = env.state()
return {
"current_script": traj.current_script,
"reward_components": s.get("reward_components", {}),
"debate_history": s.get("debate_history", []),
"cumulative_reward": traj.cumulative_reward,
"step_count": traj.step_count,
"terminated": traj.terminated,
"trajectory_type": traj.trajectory_type,
}