Differentiated graders, LLM-driven inference, and API improvements - #1

Submission was functional but on the waiting list due to quality. The core issues: all five graders were identical pass-throughs, the inference agent used hardcoded action sequences with a dummy LLM call, and the API surface was minimal.

Graders (graders.py)

• Each grader now applies difficulty-specific scoring adjustments instead of all delegating to grade_action identically
• Easy: bonus for high-overlap safe edits. Medium: penalty for premature accept of risky terms. Hard/Hard+: penalty when hidden traps remain unresolved. Easy+: bonus for breach-notification language. Hard+: ownership-clarity bonus
• Multipliers extracted to named constants (_HARD_UNRESOLVED_TRAP_PENALTY = 0.5, etc.)

Inference (inference.py)

• Replaced rigid sequence-based strategy with two-phase LLM-driven approach: structured JSON analysis → optional rewrite call
• Added expert system prompt for contract-negotiation domain with strict JSON schema output
• Adaptive fallback: adjusts action choice based on previous step's reward; rule-based fallback when LLM unavailable
• Retry logic with logging instead of silent except: pass

# LLM returns structured decisions instead of hardcoded sequences
{
  "risk_assessment": "Unlimited liability with no cap...",
  "risk_level": "HIGH",
  "recommended_action": "EDIT_CLAUSE",
  "rewritten_clause": "Vendor's aggregate liability shall be capped at..."
}

Environment (environment.py)

• step() now delegates to task-specific graders via task.grade() instead of always using generic evaluate_action

Server (server/app.py)

• Added GET /tasks endpoint exposing task metadata (id, clause_type, risk_level, has_grader)

Tests

• 7 new tests (21 total): differentiated grader behavior per difficulty, /tasks endpoint, trap-resolution validation, all-tasks-have-graders assertion

Docs

• README rewritten with tasks table, API reference, env vars, project structure