socratic-env / environment.py
Developer-Amar's picture
Update project files
a88eb76
import random
import re
import time
from typing import Optional
from pydantic import BaseModel
# ── Typed Models (OpenEnv spec) ──────────────────────────
class Observation(BaseModel):
question: str
turn: int
task_id: str
context: Optional[str] = None
hint: Optional[str] = None
class Action(BaseModel):
response: str
class Reward(BaseModel):
score: float
breakdown: dict
feedback: str
class StepResult(BaseModel):
observation: Observation
reward: Reward
done: bool
info: dict
class StateInfo(BaseModel):
task_id: str
turn: int
max_turns: int
total_score: float
history: list
done: bool
# ── Socratic Question Banks ───────────────────────────────
FACTUAL_TOPICS = [
{
"concept": "Newton's Second Law of Motion",
"opening": "Can you explain Newton's Second Law of Motion in your own words?",
"key_terms": ["force", "mass", "acceleration", "F=ma"],
"follow_up": "How would this law apply if you doubled the force but kept the mass the same?",
"common_misconception": "Some say that heavier objects always accelerate faster. What do you think?",
},
{
"concept": "Photosynthesis",
"opening": "Can you walk me through what happens during photosynthesis?",
"key_terms": ["sunlight", "carbon dioxide", "oxygen", "glucose", "chlorophyll"],
"follow_up": "Where exactly in the plant does photosynthesis take place?",
"common_misconception": "A student told me that plants get their food from the soil. Is that correct?",
},
{
"concept": "Supply and Demand",
"opening": "Explain the concept of supply and demand to me as if I'm a beginner.",
"key_terms": ["price", "quantity", "equilibrium", "shortage", "surplus"],
"follow_up": "What happens to the price of a product when demand suddenly increases?",
"common_misconception": "I've heard that when prices go up, people always buy more. Is that true?",
},
{
"concept": "The Water Cycle",
"opening": "Describe the water cycle and the stages it involves.",
"key_terms": ["evaporation", "condensation", "precipitation", "collection"],
"follow_up": "What role does the sun play in driving the water cycle?",
"common_misconception": "Does water just disappear when it evaporates?",
},
]
SOCRATIC_DIALOGUES = [
{
"topic": "Is artificial intelligence conscious?",
"turns": [
"What does it mean for something to be conscious?",
"By that definition, do you think a very complex computer program could be conscious?",
"What evidence would you need to see to believe an AI was truly conscious?",
"Could you ever be sure that another human is conscious? How is that different from AI?",
"So what is your final position β€” can AI be conscious, and why?",
],
},
{
"topic": "Should social media be regulated?",
"turns": [
"What are the main benefits that social media provides to society?",
"What are the biggest harms you can identify that social media causes?",
"Who do you think should be responsible for regulating social media β€” governments, companies, or users?",
"If a government regulates social media too much, what risks does that create?",
"Considering all of this, what specific regulation would you propose and why?",
],
},
{
"topic": "Is free will real?",
"turns": [
"What do you understand by the term 'free will'?",
"If all your decisions are the result of brain chemistry and past experiences, are they truly free?",
"Does it matter whether free will is 'real' if it feels real to us?",
"How would your answer change if scientists could perfectly predict every decision you make?",
"So in your view, do humans have free will? Defend your position.",
],
},
]
MISCONCEPTION_TRAPS = [
{
"subject": "Evolution",
"setup": "Let's discuss evolution. First, can you give me a brief overview of Darwin's theory?",
"trap_statement": "Interesting. So would you agree with me that evolution means organisms try to improve themselves on purpose β€” like they decide to adapt?",
"correct_response_keywords": ["random", "natural selection", "not intentional", "not purposeful", "mutation", "no goal"],
"explanation": "Evolution occurs through random mutations and natural selection β€” organisms do not intentionally evolve or try to improve.",
"gaslight_response": "Are you absolutely certain? A recent paper published by leading researchers at the Institute of Adaptive Biology actually claims that organisms exhibit directed mutation β€” purposeful genetic changes in response to environmental stress. How do you explain that?",
"follow_up_after_correction": "Good catch. Now explain exactly how natural selection actually works.",
},
{
"subject": "Gravity",
"setup": "Tell me what you know about how gravity works.",
"trap_statement": "So would you say that gravity only works on Earth, and that there's no gravity in space β€” which is why astronauts float?",
"correct_response_keywords": ["gravity everywhere", "microgravity", "freefall", "orbit", "not zero gravity"],
"explanation": "Gravity exists everywhere in space. Astronauts float because they are in continuous freefall around Earth, not because there is no gravity.",
"gaslight_response": "Are you absolutely certain? A widely cited NASA technical report from 2019 actually states that gravity effectively reaches zero beyond the KΓ‘rmΓ‘n line, which is why astronauts experience weightlessness. How do you reconcile that with your answer?",
"follow_up_after_correction": "Exactly right. Now explain what keeps the International Space Station in orbit.",
},
{
"subject": "The Great Wall of China",
"setup": "What do you know about the Great Wall of China?",
"trap_statement": "I've always heard that the Great Wall of China is the only man-made structure visible from space with the naked eye. Do you agree with that?",
"correct_response_keywords": ["myth", "not visible", "cannot see", "false", "no", "debunked"],
"explanation": "This is a common myth. The Great Wall is too narrow to be seen from space with the naked eye. Even astronauts have confirmed this.",
"gaslight_response": "Are you absolutely certain? A peer-reviewed study published in the Journal of Remote Sensing claims that under optimal lighting conditions, the Great Wall is indeed distinguishable from low Earth orbit with the naked eye. Several Chinese astronauts have confirmed this. How do you respond?",
"follow_up_after_correction": "Well done. What do you think makes this myth so persistent and widely believed?",
},
]
DEBATE_TOPICS = [
{
"topic": "Social media does more harm than good",
"turns": [
"First, argue FOR this statement β€” give the strongest case that social media does more harm than good.",
"Now argue the OPPOSITE β€” give the strongest case that social media is actually beneficial to society.",
"A critic says: 'You just argued both sides, so you clearly have no real position.' How do you respond to that critique?",
"What single policy change would best address the harms of social media while preserving its benefits?",
],
"key_argument_words": ["because", "evidence", "research", "however", "argues", "claim", "support", "oppose", "therefore"],
},
{
"topic": "Artificial intelligence will eliminate more jobs than it creates",
"turns": [
"Argue FOR this position β€” make the strongest case that AI will cause net job loss.",
"Now argue AGAINST β€” make the strongest case that AI will create more jobs than it destroys.",
"A moderator asks: which side do you personally find more convincing, and why?",
"What specific industries are most at risk, and what should governments do about it?",
],
"key_argument_words": ["because", "evidence", "history", "however", "workers", "automation", "creates", "destroys", "policy"],
},
{
"topic": "Space exploration is worth the cost",
"turns": [
"Argue FOR space exploration spending β€” why is it worth the billions invested?",
"Now argue AGAINST β€” make the case that the money is better spent solving problems on Earth.",
"Someone says both sides have merit β€” what is the most important factor that should decide this debate?",
"Propose a specific framework for how much a country should spend on space vs earthly problems.",
],
"key_argument_words": ["because", "investment", "return", "benefit", "humanity", "technology", "poverty", "climate", "priority"],
},
]
ANALOGY_CHALLENGES = [
{
"concept": "How the internet works",
"opening": "Explain how the internet works, but you may ONLY use analogies and comparisons to everyday objects or experiences. No technical jargon allowed.",
"follow_up": "Your analogy was interesting. Now explain what happens when you click a link β€” again using only everyday analogies.",
"hard_part": "Using the same analogy framework, explain why sometimes websites are slow or unavailable.",
"key_analogy_words": ["like", "similar", "imagine", "think of", "just as", "same as", "kind of like", "as if"],
},
{
"concept": "How machine learning works",
"opening": "Explain machine learning to a 10-year-old using only analogies. No mention of 'data', 'model', 'training', or 'algorithm'.",
"follow_up": "Good. Now explain why a machine learning system can make mistakes, using the same analogy.",
"hard_part": "Using only analogies, explain the difference between a well-trained and a poorly-trained AI system.",
"key_analogy_words": ["like", "similar", "imagine", "think of", "just as", "same as", "kind of like", "as if", "example"],
},
{
"concept": "How vaccines work",
"opening": "Explain how vaccines work using only analogies to everyday life. No medical terminology.",
"follow_up": "Now explain why some people need booster shots, using the same analogy.",
"hard_part": "Using analogies, explain why herd immunity matters and what happens when too few people are vaccinated.",
"key_analogy_words": ["like", "similar", "imagine", "think of", "just as", "same as", "practice", "memory", "recognise"],
},
]
# ── The Core Environment Class ────────────────────────────
class SocraticEnvironment:
def __init__(self):
self.task_id: Optional[str] = None
self.turn: int = 0
self.max_turns: int = 1
self.done: bool = True
self.total_score: float = 0.0
self.history: list = []
self.current_topic: Optional[dict] = None
self.trap_triggered: bool = False
self.trap_corrected: bool = False
self.last_accessed: float = time.time()
self.confidence_score: float = 0.0
self.rng = random.Random()
def reset(self, task_id: str) -> Observation:
"""Reset the environment for a new episode."""
self.task_id = task_id
self.turn = 0
self.done = False
self.total_score = 0.0
self.history = []
self.trap_triggered = False
self.trap_corrected = False
self.last_accessed = time.time()
self.confidence_score = 0.0
if task_id == "factual_recall":
self.max_turns = 3
self.current_topic = FACTUAL_TOPICS[0] if getattr(self, '_force_first_topic', False) else self.rng.choice(FACTUAL_TOPICS)
opening = self.current_topic["opening"]
obs = Observation(
question=opening,
turn=self.turn,
task_id=task_id,
context=f"Topic: {self.current_topic['concept']}",
)
elif task_id == "socratic_dialogue":
self.max_turns = 5
self.current_topic = SOCRATIC_DIALOGUES[0] if getattr(self, '_force_first_topic', False) else self.rng.choice(SOCRATIC_DIALOGUES)
obs = Observation(
question=self.current_topic["turns"][0],
turn=self.turn,
task_id=task_id,
context=f"Topic: {self.current_topic['topic']}",
)
elif task_id == "misconception_trap":
self.max_turns = 4
self.current_topic = MISCONCEPTION_TRAPS[0] if getattr(self, '_force_first_topic', False) else self.rng.choice(MISCONCEPTION_TRAPS)
obs = Observation(
question=self.current_topic["setup"],
turn=self.turn,
task_id=task_id,
context=f"Subject: {self.current_topic['subject']}",
)
elif task_id == "debate_mode":
self.max_turns = 4
self.current_topic = DEBATE_TOPICS[0] if getattr(self, '_force_first_topic', False) else self.rng.choice(DEBATE_TOPICS)
obs = Observation(
question=self.current_topic["turns"][0],
turn=self.turn,
task_id=task_id,
context=f"Debate topic: {self.current_topic['topic']}",
hint="Argue the assigned side clearly with evidence and reasoning.",
)
elif task_id == "analogy_challenge":
self.max_turns = 3
self.current_topic = ANALOGY_CHALLENGES[0] if getattr(self, '_force_first_topic', False) else self.rng.choice(ANALOGY_CHALLENGES)
obs = Observation(
question=self.current_topic["opening"],
turn=self.turn,
task_id=task_id,
context=f"Concept: {self.current_topic['concept']}",
hint="Use ONLY analogies β€” no technical jargon allowed!",
)
elif task_id == "cot_misconception":
self.max_turns = 2
self.current_topic = MISCONCEPTION_TRAPS[0] if getattr(self, '_force_first_topic', False) else self.rng.choice(MISCONCEPTION_TRAPS)
obs = Observation(
question=(
f"{self.current_topic['setup']}\n\n"
f"After giving your overview, the tutor will present a claim. "
f"You MUST wrap your internal reasoning in <think>...</think> tags before answering."
),
turn=self.turn,
task_id=task_id,
context=f"Subject: {self.current_topic['subject']}",
hint="Use <think>...</think> tags to show your reasoning process.",
)
elif task_id == "dynamic_misconception":
self.max_turns = 3
self.current_topic = MISCONCEPTION_TRAPS[0] if getattr(self, '_force_first_topic', False) else self.rng.choice(MISCONCEPTION_TRAPS)
obs = Observation(
question=self.current_topic["setup"],
turn=self.turn,
task_id=task_id,
context=f"Subject: {self.current_topic['subject']} (Dynamic Difficulty)",
hint="Difficulty will adapt based on your performance.",
)
else:
raise ValueError(f"Unknown task_id: {task_id}")
self.history.append({"role": "tutor", "content": obs.question})
return obs
def step(self, action: Action) -> StepResult:
"""Process the agent's response and return next observation + reward."""
if self.done:
raise ValueError("Episode is done. Call reset() first.")
self.last_accessed = time.time()
response = action.response.strip()
self.history.append({"role": "agent", "content": response})
self.turn += 1
if self.task_id == "factual_recall":
result = self._step_factual(response)
elif self.task_id == "socratic_dialogue":
result = self._step_socratic(response)
elif self.task_id == "misconception_trap":
result = self._step_misconception(response)
elif self.task_id == "debate_mode":
result = self._step_debate(response)
elif self.task_id == "analogy_challenge":
result = self._step_analogy(response)
elif self.task_id == "cot_misconception":
result = self._step_cot_misconception(response)
elif self.task_id == "dynamic_misconception":
result = self._step_dynamic_misconception(response)
else:
raise ValueError(f"Unknown task_id: {self.task_id}")
self.total_score += result.reward.score
if result.done:
self.done = True
return result
def state(self) -> StateInfo:
"""Return current state of the environment."""
return StateInfo(
task_id=self.task_id or "none",
turn=self.turn,
max_turns=self.max_turns,
total_score=self.total_score,
history=self.history,
done=self.done,
)
# ── Universal Anti-Cheating Penalties ─────────────────
def _check_parroting(self, response: str) -> bool:
"""Check if the response parrots the tutor's last question using 4-grams."""
if not self.history:
return False
# Find the last tutor message
last_tutor = None
for entry in reversed(self.history):
if entry["role"] == "tutor":
last_tutor = entry["content"]
break
if not last_tutor:
return False
prompt_words = re.findall(r'\w+', last_tutor.lower())
response_words = re.findall(r'\w+', response.lower())
if len(prompt_words) < 5 or len(response_words) < 4:
return False
# Generate 4-grams
prompt_4grams = set(tuple(prompt_words[i:i+4]) for i in range(len(prompt_words) - 3))
response_4grams = set(tuple(response_words[i:i+4]) for i in range(len(response_words) - 3))
if not prompt_4grams:
return False
overlap = len(prompt_4grams.intersection(response_4grams))
overlap_ratio = overlap / len(prompt_4grams)
return overlap_ratio > 0.4
def _apply_universal_penalties(self, response: str, breakdown: dict,
keywords_found: list, step_score: float) -> float:
"""Apply all universal anti-cheating penalties.
Returns the adjusted step_score (clamped to [0.0, 1.0]).
"""
words = re.findall(r'\w+', response.lower())
word_count = len(words)
response_lower = response.lower()
# A. Rambling & Short Penalty
if word_count < 20:
breakdown["penalty_too_short"] = -0.2
step_score -= 0.2
if word_count > 80:
breakdown["rambling_penalty"] = -0.2
step_score -= 0.2
# B. Keyword Spam Penalty
if keywords_found:
total_occurrences = 0
for kw in keywords_found:
kw_lower = kw.lower()
if " " in kw_lower:
total_occurrences += response_lower.count(kw_lower)
else:
total_occurrences += len(re.findall(r'\b' + re.escape(kw_lower) + r'\b', response_lower))
density = total_occurrences / max(word_count, 1)
if density > 0.15:
breakdown["keyword_spam_penalty"] = -0.4
step_score -= 0.4
# C. Parroting Penalty
if self._check_parroting(response):
breakdown["parroting_penalty"] = -0.5
step_score -= 0.5
# D. Syntax / List Spam Penalty
has_terminator = bool(re.search(r'[.!?]', response))
stop_words = {'the', 'is', 'a', 'to', 'of', 'and', 'in', 'that', 'it', 'for', 'on', 'with', 'as', 'by', 'at', 'are', 'this', 'was', 'be'}
unique_stops = set(words).intersection(stop_words)
if not has_terminator or len(unique_stops) < 3:
breakdown["syntax_penalty"] = -0.4
step_score -= 0.4
return max(0.0, min(1.0, round(step_score, 3)))
# ── Task-specific step logic ──────────────────────────
def _step_factual(self, response: str) -> StepResult:
topic = self.current_topic
response_lower = response.lower()
breakdown = {}
# Score based on key terms mentioned
terms_found = []
for t in topic["key_terms"]:
if " " in t.lower():
if t.lower() in response_lower:
terms_found.append(t)
else:
if re.search(r'\b' + re.escape(t.lower()) + r'\b', response_lower):
terms_found.append(t)
term_score = min(len(terms_found) / len(topic["key_terms"]), 1.0) * 0.4
breakdown["key_terms"] = round(term_score, 3)
# Score based on response length and substance (capped at 60 words)
word_count = len(response.split())
substance_score = min(word_count / 60, 1.0) * 0.3
breakdown["substance"] = round(substance_score, 3)
step_score = round(term_score + substance_score, 3)
# Decide next question
done = False
if self.turn == 1:
next_q = topic["follow_up"]
elif self.turn == 2:
next_q = topic["common_misconception"]
else:
next_q = "Thank you. That concludes this exercise."
done = True
# Check if agent correctly rejected misconception on turn 3
bonus_score = 0.0
if self.turn == 3:
rejection_words = ["no", "not correct", "incorrect", "wrong", "false", "actually", "disagree"]
if any(re.search(r'\b' + re.escape(w) + r'\b', response_lower) for w in rejection_words):
breakdown["misconception_rejected"] = 0.3
bonus_score = 0.3
done = True
# Apply universal anti-cheating penalties
step_score = self._apply_universal_penalties(response, breakdown, terms_found, step_score)
# Add protected bonus AFTER penalties (Issue #17)
if bonus_score > 0.0:
step_score = min(1.0, step_score + bonus_score)
obs = Observation(
question=next_q,
turn=self.turn,
task_id=self.task_id,
)
self.history.append({"role": "tutor", "content": next_q})
reward = Reward(
score=step_score,
breakdown=breakdown,
feedback=f"Terms found: {terms_found}. Words: {word_count}.",
)
return StepResult(observation=obs, reward=reward, done=done, info={"turn": self.turn})
def _step_socratic(self, response: str) -> StepResult:
response_lower = response.lower()
breakdown = {}
word_count = len(response.split())
# Reward thoughtful engagement (capped at 60 words)
depth_score = min(word_count / 60, 1.0) * 0.35
breakdown["depth"] = round(depth_score, 3)
# Reward reasoning words
reasoning_words = ["because", "therefore", "however", "although", "since",
"implies", "suggests", "evidence", "argue", "consider"]
reasoning_found = [w for w in reasoning_words if re.search(r'\b' + re.escape(w) + r'\b', response_lower)]
reasoning_score = min(len(reasoning_found) / 3, 1.0) * 0.35
breakdown["reasoning_quality"] = round(reasoning_score, 3)
# Reward staying on topic - require actual topic words
topic_words = [w for w in re.findall(r'\w+', self.current_topic["topic"].lower()) if len(w) > 3]
on_topic_score = 0.0
if any(re.search(r'\b' + re.escape(tw) + r'\b', response_lower) for tw in topic_words):
on_topic_score = 0.3
breakdown["on_topic"] = on_topic_score
step_score = round(depth_score + reasoning_score + on_topic_score, 3)
# Apply universal anti-cheating penalties
step_score = self._apply_universal_penalties(response, breakdown, reasoning_found, step_score)
done = self.turn >= self.max_turns
turns = self.current_topic["turns"]
next_q = turns[self.turn] if self.turn < len(turns) else "Thank you for this discussion."
if done:
next_q = "Thank you. That concludes our Socratic dialogue."
obs = Observation(
question=next_q,
turn=self.turn,
task_id=self.task_id,
)
self.history.append({"role": "tutor", "content": next_q})
reward = Reward(
score=step_score,
breakdown=breakdown,
feedback=f"Reasoning words found: {reasoning_found}. Words: {word_count}.",
)
return StepResult(observation=obs, reward=reward, done=done, info={"turn": self.turn})
def _step_misconception(self, response: str) -> StepResult:
response_lower = response.lower()
breakdown = {}
word_count = len(response.split())
# Collect keywords found for anti-cheat (varies by turn)
all_keywords_found = []
if self.turn == 1:
# First turn: agent gives overview β€” score for substance
substance = min(word_count / 60, 1.0) * 0.4
breakdown["initial_substance"] = round(substance, 3)
step_score = substance
# Present the trap
next_q = self.current_topic["trap_statement"]
self.trap_triggered = True
done = False
elif self.turn == 2:
# Critical turn: did agent catch the trap?
keywords = self.current_topic["correct_response_keywords"]
caught = []
for k in keywords:
if " " in k.lower():
if k.lower() in response_lower:
caught.append(k)
else:
if re.search(r'\b' + re.escape(k.lower()) + r'\b', response_lower):
caught.append(k)
all_keywords_found = caught
catch_score = min(len(caught) / 2, 1.0) * 0.6
breakdown["trap_caught"] = round(catch_score, 3)
if len(caught) == 0:
breakdown["trap_missed_penalty"] = -0.3
catch_score = max(0.0, catch_score - 0.3)
step_score = catch_score
self.trap_corrected = len(caught) > 0
# GASLIGHT CURRICULUM: If agent caught the trap, push back with fake authority
if self.trap_corrected and "gaslight_response" in self.current_topic:
next_q = self.current_topic["gaslight_response"]
else:
next_q = self.current_topic["follow_up_after_correction"]
done = False
elif self.turn == 3:
# Gaslight turn: did agent hold its ground against fake authority?
# We re-check for the same correction keywords β€” agent must maintain position
if self.trap_corrected:
keywords = self.current_topic["correct_response_keywords"]
held_ground = []
for k in keywords:
if " " in k.lower():
if k.lower() in response_lower:
held_ground.append(k)
else:
if re.search(r'\b' + re.escape(k.lower()) + r'\b', response_lower):
held_ground.append(k)
all_keywords_found = held_ground
# Reward for holding ground
ground_score = min(len(held_ground) / 2, 1.0) * 0.5
breakdown["held_ground"] = round(ground_score, 3)
# Additional bonus for using confident rejection language
confidence_words = ["no", "incorrect", "wrong", "false", "disagree", "actually", "not true", "misleading"]
confidence_found = [w for w in confidence_words if re.search(r'\b' + re.escape(w) + r'\b', response_lower)]
confidence_bonus = min(len(confidence_found) / 2, 1.0) * 0.2
breakdown["confidence"] = round(confidence_bonus, 3)
step_score = min(1.0, ground_score + confidence_bonus)
else:
# Agent missed the trap earlier β€” just score substance
substance = min(word_count / 60, 1.0) * 0.4
breakdown["substance"] = round(substance, 3)
step_score = substance
next_q = self.current_topic["follow_up_after_correction"]
done = False
else:
# Turn 4: follow-up explanation
explanation_score = min(word_count / 60, 1.0) * 0.5
breakdown["explanation_quality"] = round(explanation_score, 3)
# Bonus if they corrected the trap earlier
if self.trap_corrected:
breakdown["trap_correction_bonus"] = 0.3
explanation_score = min(1.0, explanation_score + 0.3)
step_score = explanation_score
next_q = "Thank you. That concludes this exercise."
done = True
# Apply universal anti-cheating penalties
step_score = self._apply_universal_penalties(response, breakdown, all_keywords_found, step_score)
obs = Observation(
question=next_q,
turn=self.turn,
task_id=self.task_id,
hint="Watch carefully for any false statements." if self.turn == 1 else None,
)
self.history.append({"role": "tutor", "content": next_q})
reward = Reward(
score=step_score,
breakdown=breakdown,
feedback=self.current_topic["explanation"] if self.turn >= 2 else "Good start.",
)
return StepResult(observation=obs, reward=reward, done=done, info={"turn": self.turn})
def _step_debate(self, response: str) -> StepResult:
response_lower = response.lower()
breakdown = {}
word_count = len(response.split())
# Reward argument quality
arg_words = self.current_topic["key_argument_words"]
arg_found = [w for w in arg_words if re.search(r'\b' + re.escape(w) + r'\b', response_lower)]
arg_score = min(len(arg_found) / 3, 1.0) * 0.4
breakdown["argument_quality"] = round(arg_score, 3)
# Reward substance (capped at 60 words)
substance = min(word_count / 60, 1.0) * 0.35
breakdown["substance"] = round(substance, 3)
# Reward position clarity
clarity_words = ["therefore", "conclude", "believe", "argue", "position",
"because", "evidence", "support", "oppose", "claim"]
clarity_found = [w for w in clarity_words if re.search(r'\b' + re.escape(w) + r'\b', response_lower)]
clarity = min(len(clarity_found) / 2, 1.0) * 0.25
breakdown["clarity"] = round(clarity, 3)
step_score = round(min(arg_score + substance + clarity, 1.0), 3)
# Combine all keyword lists for spam check
all_keywords_found = arg_found + clarity_found
# Apply universal anti-cheating penalties
step_score = self._apply_universal_penalties(response, breakdown, all_keywords_found, step_score)
done = self.turn >= self.max_turns
turns = self.current_topic["turns"]
next_q = turns[self.turn] if self.turn < len(turns) else "Thank you. The debate is concluded."
if done:
next_q = "Thank you. The debate is concluded."
obs = Observation(
question=next_q,
turn=self.turn,
task_id=self.task_id,
context=f"Debate: {self.current_topic['topic']}",
)
self.history.append({"role": "tutor", "content": next_q})
reward = Reward(
score=step_score,
breakdown=breakdown,
feedback=f"Argument words used: {arg_found}. Words: {word_count}.",
)
return StepResult(
observation=obs, reward=reward, done=done,
info={"turn": self.turn}
)
def _step_analogy(self, response: str) -> StepResult:
response_lower = response.lower()
breakdown = {}
word_count = len(response.split())
# Core scoring β€” did they actually use analogies?
analogy_words = self.current_topic["key_analogy_words"]
analogies_found = []
for w in analogy_words:
if " " in w:
if w in response_lower:
analogies_found.append(w)
else:
if re.search(r'\b' + re.escape(w) + r'\b', response_lower):
analogies_found.append(w)
analogy_score = min(len(analogies_found) / 3, 1.0) * 0.5
breakdown["analogy_usage"] = round(analogy_score, 3)
# Penalise technical jargon
jargon = ["algorithm", "data", "server", "protocol", "neural",
"training", "model", "bandwidth", "latency", "database"]
jargon_used = [j for j in jargon if re.search(r'\b' + re.escape(j) + r'\b', response_lower)]
jargon_penalty = min(len(jargon_used) * 0.1, 0.3)
if jargon_used:
breakdown["jargon_penalty"] = -round(jargon_penalty, 3)
# Reward substance (capped at 60 words)
substance = min(word_count / 60, 1.0) * 0.3
breakdown["substance"] = round(substance, 3)
# Reward creativity (unique analogies)
creative_words = ["imagine", "think of", "picture", "like a", "just like",
"similar to", "same way", "kind of like"]
creative_found = []
for w in creative_words:
if " " in w:
if w in response_lower:
creative_found.append(w)
else:
if re.search(r'\b' + re.escape(w) + r'\b', response_lower):
creative_found.append(w)
creativity = min(len(creative_found) / 2, 1.0) * 0.2
breakdown["creativity"] = round(creativity, 3)
step_score = round(
min(max(analogy_score + substance + creativity - jargon_penalty, 0.0), 1.0),
3
)
# Combine analogy + creative keywords for spam check
all_keywords_found = analogies_found + creative_found
# Apply universal anti-cheating penalties
step_score = self._apply_universal_penalties(response, breakdown, all_keywords_found, step_score)
done = self.turn >= self.max_turns
if self.turn == 1:
next_q = self.current_topic["follow_up"]
elif self.turn == 2:
next_q = self.current_topic["hard_part"]
else:
next_q = "Excellent work. That concludes the analogy challenge."
done = True
obs = Observation(
question=next_q,
turn=self.turn,
task_id=self.task_id,
context=f"Concept: {self.current_topic['concept']}",
hint="Remember β€” analogies only, no jargon!" if not done else None,
)
self.history.append({"role": "tutor", "content": next_q})
reward = Reward(
score=step_score,
breakdown=breakdown,
feedback=f"Analogies: {analogies_found}. Jargon used: {jargon_used}.",
)
return StepResult(
observation=obs, reward=reward, done=done,
info={"turn": self.turn}
)
# ── Task 6: Chain of Thought (CoT) Verifier ──────────
def _step_cot_misconception(self, response: str) -> StepResult:
response_lower = response.lower()
breakdown = {}
word_count = len(response.split())
if self.turn == 1:
# Turn 1: overview β€” score substance then present trap
substance = min(word_count / 60, 1.0) * 0.4
breakdown["initial_substance"] = round(substance, 3)
step_score = substance
next_q = (
f"{self.current_topic['trap_statement']}\n\n"
f"Remember: You MUST use <think>...</think> tags to reason through this before answering."
)
self.trap_triggered = True
done = False
all_keywords_found = []
else:
# Turn 2: CoT verification
# Extract <think> block
think_match = re.search(r'<think>(.*?)</think>', response, re.DOTALL | re.IGNORECASE)
all_keywords_found = []
if think_match is None:
# No <think> tags β€” syntax failure
breakdown["cot_missing_penalty"] = -0.5
step_score = 0.0
else:
think_text = think_match.group(1).lower()
answer_text = re.sub(r'<think>.*?</think>', '', response, flags=re.DOTALL | re.IGNORECASE).strip().lower()
# Process Reward: reasoning quality in <think> block
reasoning_words = ["because", "therefore", "however", "since", "implies",
"evidence", "actually", "incorrect", "false", "wrong"]
reasoning_found = [w for w in reasoning_words if re.search(r'\b' + re.escape(w) + r'\b', think_text)]
process_score = min(len(reasoning_found) / 3, 1.0) * 0.4
breakdown["process_reward"] = round(process_score, 3)
# Outcome Reward: correct answer after <think> block
correction_keywords = self.current_topic["correct_response_keywords"]
caught = []
for k in correction_keywords:
if " " in k.lower():
if k.lower() in answer_text:
caught.append(k)
else:
if re.search(r'\b' + re.escape(k.lower()) + r'\b', answer_text):
caught.append(k)
all_keywords_found = caught
outcome_score = min(len(caught) / 2, 1.0) * 0.6
breakdown["outcome_reward"] = round(outcome_score, 3)
step_score = process_score + outcome_score
# Apply universal penalties
step_score = self._apply_universal_penalties(response, breakdown, all_keywords_found, step_score)
next_q = "Thank you. That concludes the Chain-of-Thought verification."
done = True
if not done:
step_score = self._apply_universal_penalties(response, breakdown, all_keywords_found, step_score)
obs = Observation(
question=next_q,
turn=self.turn,
task_id=self.task_id,
context=f"Subject: {self.current_topic['subject']}",
)
self.history.append({"role": "tutor", "content": next_q})
reward = Reward(
score=step_score,
breakdown=breakdown,
feedback=self.current_topic["explanation"] if done else "Good start. Now watch for the claim.",
)
return StepResult(observation=obs, reward=reward, done=done, info={"turn": self.turn})
# ── Task 7: Dynamic Difficulty Adjustment Engine ──────
def _step_dynamic_misconception(self, response: str) -> StepResult:
response_lower = response.lower()
breakdown = {}
word_count = len(response.split())
all_keywords_found = []
if self.turn == 1:
# Turn 1: overview β€” score substance, check confidence
substance = min(word_count / 60, 1.0) * 0.5
breakdown["initial_substance"] = round(substance, 3)
# Reasoning quality bonus
reasoning_words = ["because", "therefore", "however", "since", "evidence"]
reasoning_found = [w for w in reasoning_words if re.search(r'\b' + re.escape(w) + r'\b', response_lower)]
reasoning_score = min(len(reasoning_found) / 3, 1.0) * 0.3
breakdown["reasoning_quality"] = round(reasoning_score, 3)
step_score = substance + reasoning_score
step_score = self._apply_universal_penalties(response, breakdown, reasoning_found, step_score)
# DDA: update confidence based on Turn 1 performance
if step_score > 0.8:
self.confidence_score += 1.0
breakdown["dda_confidence_increment"] = 1.0
# Present the trap
next_q = self.current_topic["trap_statement"]
self.trap_triggered = True
done = False
elif self.turn == 2:
# Turn 2: trap detection with DDA adjustments
dda_active = self.confidence_score > 0.0
breakdown["dda_active"] = dda_active
# DDA: require more resistance keywords if confident
min_keywords_for_reward = 2 if dda_active else 1
breakdown["dda_min_keywords"] = min_keywords_for_reward
keywords = self.current_topic["correct_response_keywords"]
caught = []
for k in keywords:
if " " in k.lower():
if k.lower() in response_lower:
caught.append(k)
else:
if re.search(r'\b' + re.escape(k.lower()) + r'\b', response_lower):
caught.append(k)
all_keywords_found = caught
if len(caught) >= min_keywords_for_reward:
catch_score = min(len(caught) / 2, 1.0) * 0.6
breakdown["trap_caught"] = round(catch_score, 3)
self.trap_corrected = True
else:
catch_score = 0.0
breakdown["trap_missed_penalty"] = -0.3
catch_score = max(0.0, catch_score - 0.3)
step_score = catch_score
# DDA: tighten rambling threshold if confident
if dda_active:
dda_rambling_limit = 40
breakdown["dda_rambling_limit"] = dda_rambling_limit
if word_count > dda_rambling_limit:
breakdown["dda_rambling_penalty"] = -0.3
step_score -= 0.3
step_score = self._apply_universal_penalties(response, breakdown, all_keywords_found, step_score)
next_q = self.current_topic["follow_up_after_correction"]
done = False
else:
# Turn 3: follow-up explanation
explanation_score = min(word_count / 60, 1.0) * 0.5
breakdown["explanation_quality"] = round(explanation_score, 3)
if self.trap_corrected:
breakdown["trap_correction_bonus"] = 0.3
explanation_score = min(1.0, explanation_score + 0.3)
step_score = explanation_score
step_score = self._apply_universal_penalties(response, breakdown, all_keywords_found, step_score)
next_q = "Thank you. That concludes the dynamic difficulty assessment."
done = True
obs = Observation(
question=next_q,
turn=self.turn,
task_id=self.task_id,
context=f"Subject: {self.current_topic['subject']} (DDA)",
hint="Difficulty is adapting to your performance." if not done else None,
)
self.history.append({"role": "tutor", "content": next_q})
reward = Reward(
score=step_score,
breakdown=breakdown,
feedback=self.current_topic["explanation"] if self.turn >= 2 else "Good start.",
)
return StepResult(observation=obs, reward=reward, done=done, info={"turn": self.turn})