Spaces:
Sleeping
Sleeping
File size: 10,785 Bytes
cf82034 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Any, Optional
from uuid import uuid4
from fastmcp import FastMCP
try:
from openenv.core.env_server.mcp_environment import MCPEnvironment
from openenv.core.env_server.types import Action, Observation, State
except ImportError:
from openenv.core.env_server.mcp_environment import MCPEnvironment
from openenv.core.env_server.types import Action, Observation, State
@dataclass
class Clause:
id: str
section: str
text: str
@dataclass
class Draft:
topic: str
requirements: list[str]
forbidden_phrases: list[str]
numeric_constraints: dict[str, int]
clauses: list[Clause] = field(default_factory=list)
definitions: dict[str, str] = field(default_factory=dict)
def to_markdown(self) -> str:
by_section: dict[str, list[Clause]] = {}
for c in self.clauses:
by_section.setdefault(c.section, []).append(c)
parts: list[str] = [f"# Policy Draft\n\n**Topic**: {self.topic}\n"]
for sec in sorted(by_section.keys()):
parts.append(f"## {sec}\n")
for c in by_section[sec]:
parts.append(f"- ({c.id}) {c.text}\n")
parts.append("\n")
if self.definitions:
parts.append("## Definitions\n")
for k in sorted(self.definitions.keys()):
parts.append(f"- **{k}**: {self.definitions[k]}\n")
parts.append("\n")
return "".join(parts).strip() + "\n"
def _contains_any(text: str, phrases: list[str]) -> list[str]:
t = text.lower()
return [p for p in phrases if p.lower() in t]
def _coverage_score(draft_md: str, requirements: list[str]) -> tuple[int, list[str]]:
missing: list[str] = []
t = draft_md.lower()
for r in requirements:
if r.lower() not in t:
missing.append(r)
covered = len(requirements) - len(missing)
return covered, missing
def _contradiction_checks(draft_md: str) -> list[str]:
# Lightweight, objective checks (expand later):
# - retention days must be consistent with constraint marker if present.
issues: list[str] = []
t = draft_md.lower()
if "retain" in t and "indefinitely" in t:
issues.append("Mentions retention and 'indefinitely' (potentially contradictory).")
if "we will never" in t and "may" in t and "share" in t:
issues.append("Contains 'we will never' and 'may share' (potential contradiction).")
return issues
def _compute_reward(draft: Draft) -> dict[str, Any]:
md = draft.to_markdown()
covered, missing = _coverage_score(md, draft.requirements)
forbidden_hits = _contains_any(md, draft.forbidden_phrases)
contradictions = _contradiction_checks(md)
# Reward components (simple but informative; RLVR-friendly)
r_coverage = covered / max(1, len(draft.requirements)) # 0..1
r_forbidden = -0.25 * len(forbidden_hits) # penalty
r_contra = -0.15 * len(contradictions)
r_defs = 0.05 * min(6, len(draft.definitions)) # small bonus for defining terms
total = (1.5 * r_coverage) + r_forbidden + r_contra + r_defs
return {
"reward_total": float(total),
"reward_components": {
"coverage": float(r_coverage),
"forbidden_penalty": float(r_forbidden),
"contradiction_penalty": float(r_contra),
"definitions_bonus": float(r_defs),
},
"audit": {
"covered": covered,
"missing": missing,
"forbidden_hits": forbidden_hits,
"contradictions": contradictions,
"constraints": draft.numeric_constraints,
},
"draft_markdown": md,
}
class ContractComplianceEnvironment(MCPEnvironment):
"""
Contract-to-Compliance environment (multi-agent roles via tools).
Tools expose stepwise actions (draft clause edits) and return an objective
reward breakdown from multiple verifiers (coverage, forbidden phrases,
contradictions, etc).
"""
def __init__(self):
mcp = FastMCP("contract_to_compliance")
self._state = State(episode_id=str(uuid4()), step_count=0)
self._draft: Draft | None = None
@mcp.tool
def reset_episode(topic: str = "AI should be open source") -> dict:
"""Start a fresh episode with a topic and default compliance rubric."""
self._state = State(episode_id=str(uuid4()), step_count=0)
# Demo rubric (expand to GDPR/CCPA + org constraints later)
requirements = [
"purpose limitation",
"data minimization",
"lawful basis",
"user consent",
"data retention",
"access & deletion request",
"security safeguards",
"third-party sharing disclosure",
]
forbidden_phrases = [
"we sell your data",
"no refunds",
"we are not responsible",
"indefinitely",
"without consent",
]
numeric_constraints = {"retention_days_max": 30}
self._draft = Draft(
topic=topic,
requirements=requirements,
forbidden_phrases=forbidden_phrases,
numeric_constraints=numeric_constraints,
clauses=[],
definitions={},
)
payload = _compute_reward(self._draft)
payload["event"] = "reset"
payload["episode_id"] = self._state.episode_id
payload["step_count"] = self._state.step_count
return payload
@mcp.tool
def get_state() -> dict:
"""Return current draft + audit + reward breakdown."""
if self._draft is None:
return {"error": "Call reset_episode first."}
payload = _compute_reward(self._draft)
payload["event"] = "state"
payload["episode_id"] = self._state.episode_id
payload["step_count"] = self._state.step_count
return payload
@mcp.tool
def add_clause(section: str, text: str) -> dict:
"""Policy Drafter: add a clause in a section."""
if self._draft is None:
return {"error": "Call reset_episode first."}
cid = f"c_{uuid4().hex[:8]}"
self._draft.clauses.append(Clause(id=cid, section=section.strip(), text=text.strip()))
self._state.step_count += 1
payload = _compute_reward(self._draft)
payload["event"] = "add_clause"
payload["clause_id"] = cid
payload["episode_id"] = self._state.episode_id
payload["step_count"] = self._state.step_count
return payload
@mcp.tool
def edit_clause(clause_id: str, new_text: str) -> dict:
"""Policy Drafter: edit an existing clause by id."""
if self._draft is None:
return {"error": "Call reset_episode first."}
for c in self._draft.clauses:
if c.id == clause_id:
c.text = new_text.strip()
self._state.step_count += 1
payload = _compute_reward(self._draft)
payload["event"] = "edit_clause"
payload["clause_id"] = clause_id
payload["episode_id"] = self._state.episode_id
payload["step_count"] = self._state.step_count
return payload
return {"error": f"Unknown clause_id: {clause_id}"}
@mcp.tool
def delete_clause(clause_id: str) -> dict:
"""Policy Drafter: delete a clause by id."""
if self._draft is None:
return {"error": "Call reset_episode first."}
before = len(self._draft.clauses)
self._draft.clauses = [c for c in self._draft.clauses if c.id != clause_id]
if len(self._draft.clauses) == before:
return {"error": f"Unknown clause_id: {clause_id}"}
self._state.step_count += 1
payload = _compute_reward(self._draft)
payload["event"] = "delete_clause"
payload["clause_id"] = clause_id
payload["episode_id"] = self._state.episode_id
payload["step_count"] = self._state.step_count
return payload
@mcp.tool
def add_definition(term: str, definition: str) -> dict:
"""Policy Drafter: add/overwrite a definition."""
if self._draft is None:
return {"error": "Call reset_episode first."}
self._draft.definitions[term.strip()] = definition.strip()
self._state.step_count += 1
payload = _compute_reward(self._draft)
payload["event"] = "add_definition"
payload["episode_id"] = self._state.episode_id
payload["step_count"] = self._state.step_count
return payload
@mcp.tool
def redteam_probe(question: str) -> dict:
"""Red Team: add an adversarial probe (doesn't mutate draft, but scores robustness later)."""
if self._draft is None:
return {"error": "Call reset_episode first."}
# For v1 we simply surface the probe in the audit payload.
payload = _compute_reward(self._draft)
payload["event"] = "redteam_probe"
payload["probe"] = question.strip()
payload["episode_id"] = self._state.episode_id
payload["step_count"] = self._state.step_count
return payload
super().__init__(mcp)
def reset(
self,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
**kwargs: Any,
) -> Observation:
# This env is driven via MCP tools; reset() just returns readiness.
self._state = State(episode_id=episode_id or str(uuid4()), step_count=0)
return Observation(done=False, reward=0.0, metadata={"status": "ready"})
def _step_impl(
self,
action: Action,
timeout_s: Optional[float] = None,
**kwargs: Any,
) -> Observation:
return Observation(
done=False,
reward=0.0,
metadata={
"error": f"Unknown action type: {type(action).__name__}. "
"Use ListToolsAction or CallToolAction for MCP interactions."
},
)
|