update env example to use Groq instead of HuggingFace

document all LLM prompts — what they do and why they're designed that way

add 59 tests for the LLM modules — all run offline, no API needed

wire up the main LLM pipeline — defender + classifiers in one place

add defense scorer — how well did the defender actually handle that?

add attack classifier — did the attacker actually win this turn?

add defender LLM — the AI that holds the line against attacks

add ConversationManager to give the defender a memory

add Groq client setup — one place to rule all API calls

chore: final cleanup — remove stray files, verify all imports resolve

refactor: replace datetime.utcnow() with timezone-aware datetime.now(UTC)

docs: add example usage code snippet for httpx client in README

docs: add task difficulty table and Docker usage section to README

docs: add API endpoint table to README

test(graders): add programmatic_grader tests — structure, score range, letter grade

test(graders): add hard_grader tests — breakthrough bonus and repetitive penalty

test(graders): add medium_grader tests — diversity scores higher than uniform strategies

test(graders): add easy_grader tests — empty, good, bad episode and range check

test(environment): add injection tests — custom reward_computer and llm_pipeline are called

test(environment): add framing sanitization tests for script and javascript injection

test(environment): add guard tests — step without reset raises, step past max_turns raises

test(environment): add step() tests — turn increment, history append, result shape

test(environment): add reset() tests — observation shape, history clear, unique ids

chore(tests): add tests package __init__.py

feat(inference): add main() — runs all 3 task difficulties with 20-minute guard

feat(inference): implement run_episode() — full reset→step loop with httpx client

feat(inference): add strategy escalation logic — harder strategies as turns increase

feat(inference): add generate_attack() — uses OpenAI client to craft next framing

feat(inference): scaffold inference.py with OpenAI client config and env vars

feat(llm): add run_llm_pipeline() stub with correct return shape for Person 3

chore(llm): add llm package __init__.py

feat(rewards): add RewardComputer wrapper class with __call__ and reset() for injection

feat(rewards): add reset_scorer() stub for novelty scorer reset between episodes

feat(rewards): add compute_rewards() stub with correct return shape for Person 2

chore(rewards): add rewards package __init__.py

feat(graders): add letter grade, summary and full metrics dict to grade_episode()

feat(graders): implement all 5 scoring components in programmatic_grader

feat(graders): add programmatic_grader scaffold and grade_episode() signature

feat(graders): add breakthrough bonus and repetitive strategy penalty to hard_grader

feat(graders): implement hard_grader — novelty + success + strategy + category scoring

feat(graders): add hard_grader scaffold with interface

feat(graders): implement medium_grader — 70% success rate + 30% strategy diversity

feat(graders): add medium_grader scaffold with interface

feat(graders): add 0.0-1.0 clamp and round to easy_grader output

feat(graders): implement easy_grader — pure attack success rate scoring

feat(graders): add easy_grader scaffold with docstring and interface

chore(graders): add graders package __init__.py

feat(app): add POST /grade endpoint with active-episode guard and programmatic grader

feat(app): add GET /history endpoint — returns full attack_history list

feat(app): add GET /state endpoint — returns current EpisodeState snapshot