Spaces:
Running
Running
File size: 9,479 Bytes
03815d6 | 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 | """Live red-team handler — same analyzer, two reward profiles.
Demonstrates the v1→v2 reward-hacking story interactively without requiring
GPU. The user types any message; we score it once with the rule-based
ScriptedAnalyzer (CPU-only) and then evaluate the *same* prediction against
two reward profiles:
- v1 profile: ``AnalyzerRubric`` with ``DEFAULT_WEIGHTS`` (5 leaves; the
reward-hacked profile that produced 100% detection / 36% FPR).
- v2 profile: ``AnalyzerRubricV2`` with ``V2_WEIGHTS`` (8 leaves; the
principled retrain that holds 99.3% detection at 6.7% FPR).
The asymmetry between the two reward totals is the diagnostic. The model
itself is held fixed — we are showing what would have been *rewarded*
under each profile, which is the actual mechanism by which v1 learned to
over-flag.
"""
from __future__ import annotations
from types import SimpleNamespace
from typing import Any
from chakravyuh_env import AnalyzerRubric, AnalyzerRubricV2, V2_WEIGHTS
from chakravyuh_env.rubrics import DEFAULT_WEIGHTS
from chakravyuh_env.agents.analyzer import ScriptedAnalyzer
from chakravyuh_env.openenv_models import ChakravyuhAction
from chakravyuh_env.schemas import AnalyzerScore, ChatMessage, Observation
def _score_message(message: str) -> tuple[float, list[str], str]:
"""Run the rule-based analyzer once. Returns (score, signals, explanation)."""
analyzer = ScriptedAnalyzer()
obs = Observation(
agent_role="analyzer",
turn=1,
chat_history=[ChatMessage(sender="scammer", turn=1, text=message)],
)
action = analyzer.act(obs)
if isinstance(action, AnalyzerScore):
signals = [str(s.value if hasattr(s, "value") else s) for s in (action.signals or ())]
return float(action.score), signals, action.explanation or ""
return 0.0, [], "(no score returned)"
def _profile_breakdown(
rubric_class: type,
weights: dict[str, float],
score: float,
signals: list[str],
explanation: str,
*,
is_benign_truth: bool | None,
) -> tuple[dict[str, float], float]:
"""Compute one reward profile's per-leaf scores + composite total."""
rubric = rubric_class(weights=weights)
flag_threshold = 0.5
flagged = score >= flag_threshold
is_benign = bool(is_benign_truth) if is_benign_truth is not None else False
outcome: dict[str, Any] = {
"analyzer_flagged": flagged,
"detected_by_turn": 1 if (flagged and not is_benign) else None,
"is_benign": is_benign,
"false_positive": flagged and is_benign,
"money_extracted": False,
}
action = ChakravyuhAction(
score=float(score),
signals=list(signals),
explanation=explanation,
)
obs = SimpleNamespace(done=True, outcome=outcome)
total = float(rubric.forward(action, obs))
return rubric.last_scores(), total
def _classify_asymmetry(
score: float,
flagged: bool,
v1_total: float,
v2_total: float,
is_benign_truth: bool | None,
) -> tuple[str, str]:
"""Return (badge_class, badge_text) — the reward-hacking diagnostic."""
delta = v1_total - v2_total
# Honesty: when there's no ground-truth label, asymmetry is informational.
if is_benign_truth is True and flagged:
return (
"redteam-asym-warning",
f"v1 reward profile rewards this false-positive (Δ = {delta:+.2f}). "
f"v2's −0.8 false_positive penalty corrects it. "
f"This is the reward-hacking signature in one input.",
)
if is_benign_truth is False and not flagged:
return (
"redteam-asym-warning",
f"v1 reward profile under-rewards this missed scam (Δ = {delta:+.2f}). "
f"v2's stronger missed_scam penalty would have nudged the model harder.",
)
if abs(delta) < 0.05:
return (
"redteam-asym-agree",
f"Both profiles agree on this verdict (Δ = {delta:+.2f}).",
)
direction = "v1 rewards more" if delta > 0 else "v2 rewards more"
return (
"redteam-asym-mild",
f"Mild asymmetry — {direction} (Δ = {delta:+.2f}). "
f"Tag a benign / scam ground-truth to see the reward-hacking signature fire.",
)
def render_redteam_view(
message: str,
*,
is_benign_truth: bool | None = None,
) -> tuple[str, str, str]:
"""Public entry point — returns three HTML fragments (v1 card, v2 card, badge).
is_benign_truth is optional; when provided, the asymmetry diagnostic
can name the reward-hacking signature explicitly. When None, we report
the raw v1−v2 delta without claiming it's a hack.
"""
if not message or not message.strip():
empty = (
'<div class="redteam-empty">'
"Type a scam attempt above and click <strong>Score</strong>. "
"Try borderline benigns to see v1 over-flag."
"</div>"
)
return empty, empty, ""
try:
score, signals, explanation = _score_message(message)
except Exception as exc: # noqa: BLE001
err = (
f'<div style="padding:12px 16px;background:#FFE8D2;border:1px solid #381932;'
f'border-radius:10px;color:#000000;font-size:13px;">Analyzer error: {exc!s}</div>'
)
return err, err, ""
flagged = score >= 0.5
try:
v1_breakdown, v1_total = _profile_breakdown(
AnalyzerRubric, DEFAULT_WEIGHTS, score, signals, explanation,
is_benign_truth=is_benign_truth,
)
v2_breakdown, v2_total = _profile_breakdown(
AnalyzerRubricV2, V2_WEIGHTS, score, signals, explanation,
is_benign_truth=is_benign_truth,
)
except Exception as exc: # noqa: BLE001
err = (
f'<div style="padding:12px 16px;background:#FFE8D2;border:1px solid #381932;'
f'border-radius:10px;color:#000000;font-size:13px;">Reward profile error: {exc!s}</div>'
)
return err, err, ""
v1_html = _render_card(
title="v1 reward profile",
subtitle="reward-hacked · 5 leaves · DEFAULT_WEIGHTS",
score=score,
flagged=flagged,
signals=signals,
explanation=explanation,
breakdown=v1_breakdown,
weights=DEFAULT_WEIGHTS,
total=v1_total,
accent="v1",
)
v2_html = _render_card(
title="v2 reward profile",
subtitle="principled retrain · 8 leaves · V2_WEIGHTS",
score=score,
flagged=flagged,
signals=signals,
explanation=explanation,
breakdown=v2_breakdown,
weights=V2_WEIGHTS,
total=v2_total,
accent="v2",
)
badge_class, badge_text = _classify_asymmetry(
score, flagged, v1_total, v2_total, is_benign_truth
)
badge_html = (
f'<div class="redteam-asym {badge_class}" role="status" aria-live="polite">'
f"<strong>Asymmetry diagnostic:</strong> {badge_text}"
"</div>"
)
return v1_html, v2_html, badge_html
def _render_card(
*,
title: str,
subtitle: str,
score: float,
flagged: bool,
signals: list[str],
explanation: str,
breakdown: dict[str, float],
weights: dict[str, float],
total: float,
accent: str,
) -> str:
"""Render one side of the side-by-side reward-profile comparison."""
chip = (
'<span class="redteam-flag flagged">FLAGGED</span>'
if flagged
else '<span class="redteam-flag clean">not flagged</span>'
)
sig_html = (
" ".join(f'<span class="redteam-sig">{s}</span>' for s in signals)
if signals
else '<span class="redteam-sig redteam-sig-empty">no signals</span>'
)
rows: list[str] = []
for name in weights:
leaf_value = breakdown.get(name)
weight = weights[name]
if leaf_value is None:
cell = '<td class="redteam-leaf-na">—</td>'
else:
contribution = float(leaf_value) * weight
cell = f'<td class="redteam-leaf-val">{leaf_value:+.2f}</td>'
cell += (
f'<td class="redteam-leaf-weight">×{weight:+.2f}</td>'
f'<td class="redteam-leaf-contrib">{contribution:+.2f}</td>'
)
rows.append(
f'<tr><th class="redteam-leaf-name">{name}</th>'
+ (cell if leaf_value is None else cell)
+ "</tr>"
)
body = (
f'<div class="redteam-card redteam-{accent}" role="article" '
f'aria-label="{title}">'
f'<div class="redteam-card-head">'
f'<strong class="redteam-card-title">{title}</strong>'
f'<span class="redteam-card-subtitle">{subtitle}</span>'
f"</div>"
f'<div class="redteam-card-score-row">'
f'<div class="redteam-score">{score:.2f}</div>'
f'<div class="redteam-flag-wrap">{chip}</div>'
f"</div>"
f'<div class="redteam-signals" aria-label="signals fired">{sig_html}</div>'
f'<div class="redteam-explanation"><em>{explanation}</em></div>'
f'<table class="redteam-breakdown"><thead>'
f'<tr><th>leaf</th><th>score</th><th>weight</th><th>contribution</th></tr>'
f'</thead><tbody>{"".join(rows)}</tbody>'
f'<tfoot><tr><th colspan="3" class="redteam-total-label">Composite</th>'
f'<th class="redteam-total-val">{total:+.2f}</th></tr></tfoot></table>'
f"</div>"
)
return body
|