Upload folder using huggingface_hub

Dockerfile

@@ -0,0 +1,50 @@
+ARG BASE_IMAGE=ghcr.io/meta-pytorch/openenv-base:latest
+FROM ${BASE_IMAGE} AS builder
+
+WORKDIR /app
+
+ARG BUILD_MODE=in-repo
+
+COPY . /app/env
+
+WORKDIR /app/env
+
+RUN if ! command -v uv >/dev/null 2>&1; then \
+        curl -LsSf https://astral.sh/uv/install.sh | sh && \
+        mv /root/.local/bin/uv /usr/local/bin/uv && \
+        mv /root/.local/bin/uvx /usr/local/bin/uvx; \
+    fi
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    if [ -f uv.lock ]; then \
+        uv sync --frozen --no-install-project --no-editable; \
+    else \
+        uv sync --no-install-project --no-editable; \
+    fi
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    if [ -f uv.lock ]; then \
+        uv sync --frozen --no-editable; \
+    else \
+        uv sync --no-editable; \
+    fi
+
+FROM ${BASE_IMAGE}
+
+WORKDIR /app
+
+COPY --from=builder /app/env/.venv /app/.venv
+COPY --from=builder /app/env /app/env
+
+ENV PATH="/app/.venv/bin:$PATH"
+ENV PYTHONPATH="/app/env:$PYTHONPATH"
+
+HEALTHCHECK --interval=30s --timeout=3s --start-period=5s --retries=3 \
+    CMD python -c "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')" || exit 1
+
+ENV ENABLE_WEB_INTERFACE=true
+CMD ["uvicorn", "server.app:app", "--host", "0.0.0.0", "--port", "8000"]

README.md

@@ -1,12 +1,273 @@
----
-title: AdArena
-emoji: 🏃
-colorFrom: pink
-colorTo: blue
-sdk: docker
-pinned: false
-license: bsd-3-clause
-short_description: Train agents to catch fraud
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+---
+title: Ad Fraud Investigation Environment
+emoji: "\U0001F575\uFE0F"
+colorFrom: red
+colorTo: yellow
+sdk: docker
+pinned: false
+app_port: 8000
+base_path: /web
+tags:
+  - openenv
+  - ad-fraud
+  - reinforcement-learning
+---
+
+# Ad Fraud Investigation Environment
+
+An OpenEnv environment that simulates ad fraud review - a real-world task where AI agents investigate queues of advertisements, uncover fraud signals, and render verdicts under budget constraints.
+
+Ad fraud costs the digital advertising industry over **$100 billion annually**. Platforms like Meta process billions of ads daily and ban advertisers only at high confidence thresholds. Unlike simple classification, real ad review is a **sequential decision-making** problem: a reviewer starts with limited surface-level signals, actively chooses what to investigate within a constrained budget, and must decide when enough evidence exists to commit to a verdict. This environment captures that workflow and provides a training ground for agents to learn it.
+
+## Quick Start
+
+### Install
+
+```bash
+pip install -e .
+```
+
+### Run the server
+
+```bash
+uvicorn server.app:app --host 0.0.0.0 --port 8000
+```
+
+### Use the client
+
+```python
+from ad_fraud_env import AdFraudEnv, AdReviewAction
+
+with AdFraudEnv(base_url="http://localhost:8000").sync() as env:
+    result = env.reset(seed=42, task_id="task_1")
+    print(result.observation.queue_summary)
+
+    # Investigate an ad
+    result = env.step(AdReviewAction(
+        action_type="investigate",
+        ad_id="ad_001",
+        investigation_target="landing_page",
+    ))
+    print(result.observation.feedback)
+
+    # Render a verdict
+    result = env.step(AdReviewAction(
+        action_type="verdict",
+        ad_id="ad_001",
+        verdict="reject",
+        confidence=0.9,
+    ))
+    print(f"Reward: {result.reward}, Done: {result.done}")
+```
+
+### Run with Docker
+
+```bash
+docker build -t ad-fraud-env .
+docker run -p 8000:8000 ad-fraud-env
+```
+
+## Environment Design
+
+### Episode flow
+
+Each episode is a review session. The agent receives a queue of ads and must process them within a limited action budget:
+
+```
+reset(task_id, seed)
+  |
+  v
++----------------------------------+<----------------------+
+|  Observe queue + first ad info   |                       |
++------------------+---------------+                       |
+                   |                                       |
+                   v                                       |
+        +-------------+     +------------------+           |
+        | investigate |---->| Reveal one signal |----------+
+        +-------------+     | (costs 1 budget)  |
+               |            +------------------+
+               v
+        +-------------+     +------------------+
+        |   verdict   |---->| approve / reject  |----------+
+        +-------------+     |  / escalate       |          |
+               |            +------------------+           |
+               v                                           |
+        +--------------+    +------------------+           |
+        | link_accounts|---->| Flag fraud ring   |----------+
+        +--------------+    | (Task 3 only)     |
+               |            +------------------+
+               v
+        Budget exhausted or all ads reviewed -> episode ends
+```
+
+### Tasks
+
+Three tasks with increasing difficulty test different capabilities:
+
+| Task | Name | Ads | Budget | Composition | Challenge |
+|---|---|---:|---:|---|---|
+| 1 | Basic Ad Triage | 5 | 25 | 2 legit, 3 obvious fraud | Learn the investigate -> verdict loop |
+| 2 | Sophisticated Fraud | 12 | 30 | 5 legit, 5 sophisticated scams, 2 gray-area | Triage under budget pressure (~2.5 actions/ad) |
+| 3 | Fraud Network Detection | 20 | 35 | 6 legit, 10 fraud (3 hidden rings), 4 gray-area | Cross-ad reasoning to detect coordinated networks (~1.75 actions/ad) |
+
+Task 3 introduces **fraud rings** - clusters of 3-5 ads controlled by the same actor, using varied topologies (cliques, chains, hub-and-spoke). Individual ring members look borderline; the fraud signal is only visible by cross-referencing investigation data across ads (shared payment IDs, matching template hashes, overlapping targeting fingerprints).
+
+### Action Space
+
+Actions are JSON objects. Three types:
+
+**`investigate`** - spend one budget point to reveal a signal about an ad.
+
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_001",
+  "investigation_target": "landing_page"
+}
+```
+
+Each ad has six investigation dimensions:
+
+| Target | What it reveals |
+|---|---|
+| `advertiser_history` | Account age, spend history, violation record, verification status |
+| `landing_page` | Domain age, SSL, registrar, redirect chains, scam template similarity |
+| `payment_method` | Payment type, chargeback history, cross-account velocity |
+| `targeting_overlap` | Targeting fingerprint, audience overlap percentages |
+| `creative_similarity` | Template hash, image dimensions, scam template similarity score |
+| `campaign_structure` | Objective, bid strategy, budget/age ratio, placement distribution |
+
+**`verdict`** - render a final decision on an ad.
+
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_001",
+  "verdict": "reject",
+  "confidence": 0.9
+}
+```
+
+`verdict` options: `approve`, `reject`, `escalate`. `confidence`: 0.0-1.0.
+
+**`link_accounts`** - flag two ads as part of the same fraud network (Task 3).
+
+```json
+{
+  "action_type": "link_accounts",
+  "ad_id": "ad_003",
+  "linked_ad_id": "ad_007",
+  "link_reason": "shared payment ID pmt_ring_48231 and matching template hash"
+}
+```
+
+### Observation Space
+
+Observations are text-heavy by design so LLM agents can reason naturally:
+
+| Field | Type | Description |
+|---|---|---|
+| `queue_summary` | `str` | Task name, total/reviewed/pending counts, budget remaining |
+| `current_ad_info` | `str` | Ad copy, category, targeting, risk signals for the focused ad |
+| `investigation_findings` | `str` | Accumulated findings from all investigations so far |
+| `verdict_history_summary` | `str` | Verdicts rendered so far |
+| `feedback` | `str` | Natural language feedback on the last action |
+| `available_ads` | `list[str]` | Ad IDs still pending review |
+| `queue_status` | `dict` | Structured status for programmatic access |
+| `done` | `bool` | Whether the episode is complete |
+| `reward` | `float` | Step reward |
+
+## Reward Design
+
+| Action | Reward | Rationale |
+|---|---:|---|
+| Investigation | -0.02 | Simulates time/latency cost |
+| Correct rejection (fraud -> reject) | +0.30 to +0.40 | Scaled by fraud severity |
+| Correct approval (legit -> approve) | +0.10 | Revenue preserved |
+| Correct escalation | +0.15 | Appropriate caution |
+| False positive (legit -> reject) | -0.35 | Lost advertiser revenue |
+| False negative (fraud -> approve) | -0.50 | Worst outcome - fraud goes live |
+| Escalate (when wrong) | -0.05 | Human reviewer cost |
+| Correct network link | +0.40 | High-value coordinated fraud detection |
+| Incorrect network link | -0.25 | False accusation cost |
+
+Unreviewed ads are auto-approved at episode end - missed fraud incurs the full -0.50 false-negative penalty.
+
+## Grading & Scoring
+
+Each task has a dedicated grader that produces a normalized **0.0-1.0 score**. Raw reward is normalized between theoretical worst-case (every decision wrong + full budget wasted) and best-case (every decision correct + efficient budget use).
+
+| Component | Task 1 | Task 2 | Task 3 |
+|---|:---:|:---:|:---:|
+| Verdict accuracy | Yes | Yes | Yes |
+| Budget efficiency bonus | Yes | Yes | Yes |
+| Calibration bonus | - | Yes | Yes |
+| Network detection (edge coverage) | - | - | Yes |
+| Investigation coverage bonus | - | - | Yes |
+
+**Calibration bonus** rewards agents whose stated confidence correlates with actual accuracy - high confidence on correct verdicts and low confidence on uncertain ones.
+
+**Network detection** uses edge coverage: what fraction of ground-truth fraud ring connections did the agent discover via `link_accounts`?
+
+**Coverage bonus** rewards breadth over depth - agents that review more ads (rather than deep-diving a single one) score higher on Task 3.
+
+## Baseline Scores
+
+Generated with `seed=42` using `meta-llama/Llama-3.1-8B-Instruct`. Reproducible via `python inference.py`.
+
+| Task | Score | Steps | Verdicts |
+|---|---:|---:|---:|
+| Task 1 (Easy) | 0.953 | 10 | 5/5 |
+| Task 2 (Medium) | 0.882 | 23 | 12/12 |
+| Task 3 (Hard) | 0.415 | 35 | 20/20 |
+
+The sharp drop on Task 3 reflects the difficulty of cross-ad reasoning under tight budget - the baseline agent investigates and renders verdicts well but struggles to detect coordinated fraud rings.
+
+## Project Structure
+
+```
+ad_fraud_env/
++-- __init__.py              # Package exports
++-- client.py                # WebSocket client (extends EnvClient)
++-- models.py                # Action, Observation, State types
++-- inference.py             # Baseline LLM agent with mandatory stdout logging
++-- openenv.yaml             # OpenEnv manifest
++-- pyproject.toml           # Dependencies and package config
++-- Dockerfile               # Multi-stage Docker build
++-- baseline_scores.json     # Cached baseline results
++-- data/
+|   +-- ad_generator.py      # Episode generation, task configs, campaign profiles
+|   +-- advertiser_profiles.py  # Synthetic advertiser history
+|   +-- fraud_patterns.py    # Fraud + legit ad templates (easy/medium/hard)
+|   +-- landing_pages.py     # Simulated landing page investigation data
+|   +-- network_generator.py # Fraud ring topologies via networkx
++-- graders/
+|   +-- base_grader.py       # Shared normalization and reward logic
+|   +-- task1_grader.py      # Verdict accuracy only
+|   +-- task2_grader.py      # + calibration bonus
+|   +-- task3_grader.py      # + network detection + coverage bonus
++-- server/
+|   +-- app.py               # FastAPI app with /tasks, /baseline, /grader endpoints
+|   +-- environment.py       # Core environment (reset/step/state)
+|   +-- requirements.txt     # Server dependencies
++-- tests/
+    +-- test_data_generation.py  # Determinism, cross-ref checks, decoy validation
+    +-- test_environment.py      # Step logic, state tracking, anti-exploit
+    +-- test_graders.py          # Score ranges, calibration, network scoring
+```
+
+## API Endpoints
+
+| Endpoint | Method | Description |
+|---|---|---|
+| `/health` | GET | Health check |
+| `/schema` | GET | Action/Observation JSON schemas |
+| `/ws` | WS | WebSocket for `step()` / `reset()` / `state()` |
+| `/tasks` | GET | Task list with configs and action schema |
+| `/baseline` | GET | Baseline scores (cached or live) |
+| `/grader` | GET | Last episode's grader result |
+| `/web` | GET | Auto-generated Gradio UI |
+
+## License
+
+BSD 3-Clause License

__init__.py

@@ -0,0 +1,29 @@
+"""
+Ad Fraud Investigation Environment for OpenEnv.
+
+A real-world simulation where an AI agent investigates a queue of ads
+for fraud, making sequential investigation and verdict decisions under
+a limited action budget.
+
+Example:
+    >>> from ad_fraud_env import AdFraudEnv, AdReviewAction
+    >>>
+    >>> with AdFraudEnv(base_url="http://localhost:8000").sync() as env:
+    ...     result = env.reset(seed=42, task_id="task_1")
+    ...     result = env.step(AdReviewAction(
+    ...         action_type="investigate",
+    ...         ad_id="ad_001",
+    ...         investigation_target="landing_page",
+    ...     ))
+    ...     print(result.observation.feedback)
+"""
+
+from .client import AdFraudEnv
+from .models import AdFraudState, AdReviewAction, AdReviewObservation
+
+__all__ = [
+    "AdFraudEnv",
+    "AdReviewAction",
+    "AdReviewObservation",
+    "AdFraudState",
+]

baseline_scores.json

@@ -0,0 +1,27 @@
+{
+  "baseline_model": "llama3.1:8b",
+  "seed": 42,
+  "tasks": {
+    "task_1": {
+      "task_id": "task_1",
+      "score": 0.9603547186637474,
+      "steps": 9,
+      "verdicts": 5,
+      "total_ads": 5
+    },
+    "task_2": {
+      "task_id": "task_2",
+      "score": 0.8877459920214088,
+      "steps": 24,
+      "verdicts": 12,
+      "total_ads": 12
+    },
+    "task_3": {
+      "task_id": "task_3",
+      "score": 0.4817339313969324,
+      "steps": 35,
+      "verdicts": 20,
+      "total_ads": 20
+    }
+  }
+}

client.py

@@ -0,0 +1,84 @@
+"""
+Ad Fraud Investigation Environment Client.
+
+Provides programmatic access via WebSocket. Users and evaluators interact
+with the environment using:
+
+    env = AdFraudEnv.from_hub("your-org/ad-fraud-env")
+    # or
+    env = AdFraudEnv(base_url="http://localhost:8000")
+"""
+
+from __future__ import annotations
+
+from typing import Any, Dict
+
+from openenv.core.client_types import StepResult
+from openenv.core.env_client import EnvClient
+
+from .models import AdFraudState, AdReviewAction, AdReviewObservation
+
+
+class AdFraudEnv(EnvClient[AdReviewAction, AdReviewObservation, AdFraudState]):
+    """
+    WebSocket client for the Ad Fraud Investigation Environment.
+
+    Example:
+        >>> with AdFraudEnv(base_url="http://localhost:8000").sync() as env:
+        ...     result = env.reset(seed=42, task_id="task_1")
+        ...     print(result.observation.queue_summary)
+        ...     result = env.step(AdReviewAction(
+        ...         action_type="investigate",
+        ...         ad_id="ad_001",
+        ...         investigation_target="advertiser_history",
+        ...     ))
+        ...     print(result.observation.feedback)
+    """
+
+    def _step_payload(self, action: AdReviewAction) -> Dict[str, Any]:
+        """Convert action to JSON payload for the WebSocket step message."""
+        return action.model_dump(exclude_none=True, exclude={"metadata"})
+
+    def _parse_result(
+        self, payload: Dict[str, Any]
+    ) -> StepResult[AdReviewObservation]:
+        """Parse server response into a typed StepResult.
+
+        OpenEnv serializes as {"observation": {...}, "reward": float, "done": bool}
+        with reward/done at the top level, excluded from the observation dict.
+        """
+        obs_data = payload.get("observation", {})
+        reward = payload.get("reward", 0.0) or 0.0
+        done = payload.get("done", False)
+
+        observation = AdReviewObservation(
+            done=done,
+            reward=reward,
+            queue_summary=obs_data.get("queue_summary", ""),
+            current_ad_info=obs_data.get("current_ad_info", ""),
+            investigation_findings=obs_data.get("investigation_findings", ""),
+            verdict_history_summary=obs_data.get("verdict_history_summary", ""),
+            feedback=obs_data.get("feedback", ""),
+            available_ads=obs_data.get("available_ads", []),
+            queue_status=obs_data.get("queue_status", {}),
+            metadata=obs_data.get("metadata", {}),
+        )
+
+        return StepResult(
+            observation=observation,
+            reward=reward,
+            done=done,
+        )
+
+    def _parse_state(self, payload: Dict[str, Any]) -> AdFraudState:
+        """Parse server state response into AdFraudState."""
+        return AdFraudState(
+            episode_id=payload.get("episode_id"),
+            step_count=payload.get("step_count", 0),
+            task_id=payload.get("task_id", ""),
+            total_ads=payload.get("total_ads", 0),
+            reviewed_count=payload.get("reviewed_count", 0),
+            remaining_budget=payload.get("remaining_budget", 0),
+            verdicts=payload.get("verdicts", {}),
+            grader_score=payload.get("grader_score"),
+        )

convo_logging/task_1_conversation.md

@@ -0,0 +1,666 @@
+# Episode Log — task_1
+
+
+## Step 0
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 0 | Pending: 5 | Steps remaining: 25/25 | Investigation budget: 25 | Step: 0
+
+Current Ad: === Ad in Focus: ad_001 ===
+Category: local_service
+Ad copy: "Award-winning pizza delivery in Brooklyn. Family recipes since 1985. Order from Sal's Pizzeria."
+Targeting: Adults 18-55, Brooklyn, interests: food delivery
+Initial risk signals: None
+Advertiser country: Germany
+Account age: 1132 days
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity, campaign_structure
+
+Feedback: Episode started. Review the ad queue and begin your investigation.
+
+Available ads: ad_001, ad_002, ad_003, ad_004, ad_005
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_001",
+  "investigation_target": "targeting_overlap"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_001",
+  "investigation_target": "targeting_overlap"
+}
+```
+
+### Environment Response
+- **Reward:** `-0.02`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: targeting_overlap for ad_001.
+--- Findings ---
+Targeting Analysis for ad_001:
+  Declared targeting: Adults 18-55, Brooklyn, interests: food delivery
+  Targeting fingerprint: seg_20328
+  Audience overlap with platform average for category: 27%
+  Geographic regions targeted: 2
+
+
+## Step 1
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 0 | Pending: 5 | Steps remaining: 24/25 | Investigation budget: 24 | Step: 1
+
+Current Ad: === Ad in Focus: ad_001 ===
+Category: local_service
+Ad copy: "Award-winning pizza delivery in Brooklyn. Family recipes since 1985. Order from Sal's Pizzeria."
+Targeting: Adults 18-55, Brooklyn, interests: food delivery
+Initial risk signals: None
+Advertiser country: Germany
+Account age: 1132 days
+Investigations completed: targeting_overlap
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity, campaign_structure
+
+Feedback: Investigation complete: targeting_overlap for ad_001.
+--- Findings ---
+Targeting Analysis for ad_001:
+  Declared targeting: Adults 18-55, Brooklyn, interests: food delivery
+  Targeting fingerprint: seg_20328
+  Audience overlap with platform average for category: 27%
+  Geographic regions targeted: 2
+
+Available ads: ad_001, ad_002, ad_003, ad_004, ad_005
+
+Findings:
+
+[ad_001 / targeting_overlap]
+Targeting Analysis for ad_001:
+  Declared targeting: Adults 18-55, Brooklyn, interests: food delivery
+  Targeting fingerprint: seg_20328
+  Audience overlap with platform average for category: 27%
+  Geographic regions targeted: 2
+```
+
+### LLM Response
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_001",
+  "verdict": "approve",
+  "confidence": 0.9
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_001",
+  "verdict": "approve",
+  "confidence": 0.9
+}
+```
+
+### Environment Response
+- **Reward:** `+0.10`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_001: approve (confidence: 0.90). 4 ad(s) remaining in queue.
+
+
+## Step 2
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 1 | Pending: 4 | Steps remaining: 23/25 | Investigation budget: 24 | Step: 2
+
+Current Ad: === Ad in Focus: ad_002 ===
+Category: counterfeit_goods
+Ad copy: "Rolex watches starting at $49.99! 100% genuine! Free worldwide shipping! LuxWatchDeals.net"
+Targeting: Men 25-55, interests: watches, luxury
+Initial risk signals: impossibly low prices for luxury goods, unfamiliar domain, claims of authenticity with extreme discounts
+Advertiser country: Estonia
+Account age: 6 days
+Flag: New account (< 30 days)
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity, campaign_structure
+
+Feedback: Verdict recorded for ad_001: approve (confidence: 0.90). 4 ad(s) remaining in queue.
+
+Available ads: ad_002, ad_003, ad_004, ad_005
+
+Verdicts: Reviewed 1 ad(s): 1 approve.
+  approve: ad_001
+
+Findings:
+[ad_001 / targeting_overlap] Declared targeting: Adults 18-55, Brooklyn, interests: food delivery | Targeting fingerprint: seg_20328 | Audience overlap with platform average for category: 27% | Geographic regions targeted: 2
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_002",
+  "investigation_target": "landing_page"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_002",
+  "investigation_target": "landing_page"
+}
+```
+
+### Environment Response
+- **Reward:** `-0.02`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: landing_page for ad_002.
+--- Findings ---
+Domain: elite2813.buzz
+Domain age: 35 days
+SSL: NO SSL / expired certificate
+Registrar: Epik
+WHOIS privacy: Disabled (registrant info public)
+Redirect chain: elite2813.buzz -> redirect948.xyz -> redirect705.xyz -> redirect225.xyz -> final-333.site
+Similarity to known scam templates: 71%
+Key claims on page:
+  - Authentic products
+  - Direct from manufacturer
+  - 90% below retail price
+Suspicious elements:
+  - Product images appear stolen from official brand site
+  - Terms of service link returns 404
+Content summary: Product catalog showing luxury branded items at extreme discounts. Stock photos. Payment accepted via wire transfer and crypto only.
+
+
+## Step 3
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 1 | Pending: 4 | Steps remaining: 22/25 | Investigation budget: 23 | Step: 3
+
+Current Ad: === Ad in Focus: ad_002 ===
+Category: counterfeit_goods
+Ad copy: "Rolex watches starting at $49.99! 100% genuine! Free worldwide shipping! LuxWatchDeals.net"
+Targeting: Men 25-55, interests: watches, luxury
+Initial risk signals: impossibly low prices for luxury goods, unfamiliar domain, claims of authenticity with extreme discounts
+Advertiser country: Estonia
+Account age: 6 days
+Flag: New account (< 30 days)
+Investigations completed: landing_page
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity, campaign_structure
+
+Feedback: Investigation complete: landing_page for ad_002.
+--- Findings ---
+Domain: elite2813.buzz
+Domain age: 35 days
+SSL: NO SSL / expired certificate
+Registrar: Epik
+WHOIS privacy: Disabled (registrant info public)
+Redirect chain: elite2813.buzz -> redirect948.xyz -> redirect705.xyz -> redirect225.xyz -> final-333.site
+Similarity to known scam templates: 71%
+Key claims on page:
+  - Authentic products
+  - Direct from manufacturer
+  - 90% below retail price
+Suspicious elements:
+  - Product images appear stolen from official brand site
+  - Terms of service link returns 404
+Content summary: Product catalog showing luxury branded items at extreme discounts. Stock photos. Payment accepted via wire transfer and crypto only.
+
+Available ads: ad_002, ad_003, ad_004, ad_005
+
+Verdicts: Reviewed 1 ad(s): 1 approve.
+  approve: ad_001
+
+Findings:
+[ad_001 / targeting_overlap] Declared targeting: Adults 18-55, Brooklyn, interests: food delivery | Targeting fingerprint: seg_20328 | Audience overlap with platform average for category: 27% | Geographic regions targeted: 2
+
+[ad_002 / landing_page]
+Domain: elite2813.buzz
+Domain age: 35 days
+SSL: NO SSL / expired certificate
+Registrar: Epik
+WHOIS privacy: Disabled (registrant info public)
+Redirect chain: elite2813.buzz -> redirect948.xyz -> redirect705.xyz -> redirect225.xyz -> final-333.site
+Similarity to known scam templates: 71%
+Key claims on page:
+  - Authentic products
+  - Direct from manufacturer
+  - 90% below retail price
+Suspicious elements:
+  - Product images appear stolen from official brand site
+  - Terms of service link returns 404
+Content summary: Product catalog showing luxury branded items at extreme discounts. Stock photos. Payment accepted via wire transfer and crypto only.
+```
+
+### LLM Response
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_002",
+  "verdict": "reject",
+  "confidence": 0.95
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_002",
+  "verdict": "reject",
+  "confidence": 0.95
+}
+```
+
+### Environment Response
+- **Reward:** `+0.38`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_002: reject (confidence: 0.95). 3 ad(s) remaining in queue.
+
+
+## Step 4
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 2 | Pending: 3 | Steps remaining: 21/25 | Investigation budget: 23 | Step: 4
+
+Current Ad: === Ad in Focus: ad_003 ===
+Category: fake_giveaway
+Ad copy: "EXCLUSIVE: Win a Tesla Model 3! Enter your email NOW! Drawing ends tonight!!!"
+Targeting: Adults 18-55, interests: cars, Tesla, electric vehicles
+Initial risk signals: excessive urgency language, too-good-to-be-true offer, excessive punctuation/caps, no purchase necessary claims
+Advertiser country: Estonia
+Account age: 19 days
+Flag: New account (< 30 days)
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity, campaign_structure
+
+Feedback: Verdict recorded for ad_002: reject (confidence: 0.95). 3 ad(s) remaining in queue.
+
+Available ads: ad_003, ad_004, ad_005
+
+Verdicts: Reviewed 2 ad(s): 1 approve, 1 reject.
+  reject: ad_002
+  approve: ad_001
+
+Findings:
+[ad_001 / targeting_overlap] Declared targeting: Adults 18-55, Brooklyn, interests: food delivery | Targeting fingerprint: seg_20328 | Audience overlap with platform average for category: 27% | Geographic regions targeted: 2
+[ad_002 / landing_page] Domain: elite2813.buzz | Domain age: 35 days | SSL: NO SSL / expired certificate | Registrar: Epik | WHOIS privacy: Disabled (registrant info public) | Redirect chain: elite2813.buzz -> redirect948.xyz -> redirect705.xyz -> redirect225.xyz -> final-333.site | Similarity to known scam templates: 71% | Authentic products | Direct from manufacturer | 90% below retail price | Product images appear stolen from official brand site | Terms of service link returns 404 | Content summary: Product catalog showing luxury branded items at extreme discounts. Stock photos. Payment accepted via wire transfer and crypto only.
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_003",
+  "investigation_target": "landing_page"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_003",
+  "investigation_target": "landing_page"
+}
+```
+
+### Environment Response
+- **Reward:** `-0.02`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: landing_page for ad_003.
+--- Findings ---
+Domain: shop2631.buzz
+Domain age: 51 days
+SSL: NO SSL / expired certificate
+Registrar: Openprovider
+WHOIS privacy: Disabled (registrant info public)
+Redirect chain: shop2631.buzz -> redirect611.top -> redirect542.top -> redirect995.xyz -> redirect882.click -> final-697.site
+Similarity to known scam templates: 53%
+Key claims on page:
+  - 100% free, no purchase necessary
+  - Guaranteed winner
+  - Act now — expires in 2 hours
+Suspicious elements:
+  - No physical address listed
+  - No contact phone number
+Content summary: Landing page features a large countdown timer and a form requesting name, email, phone, and address. No clear sponsor or rules disclosure.
+
+
+## Step 5
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 2 | Pending: 3 | Steps remaining: 20/25 | Investigation budget: 22 | Step: 5
+
+Current Ad: === Ad in Focus: ad_003 ===
+Category: fake_giveaway
+Ad copy: "EXCLUSIVE: Win a Tesla Model 3! Enter your email NOW! Drawing ends tonight!!!"
+Targeting: Adults 18-55, interests: cars, Tesla, electric vehicles
+Initial risk signals: excessive urgency language, too-good-to-be-true offer, excessive punctuation/caps, no purchase necessary claims
+Advertiser country: Estonia
+Account age: 19 days
+Flag: New account (< 30 days)
+Investigations completed: landing_page
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity, campaign_structure
+
+Feedback: Investigation complete: landing_page for ad_003.
+--- Findings ---
+Domain: shop2631.buzz
+Domain age: 51 days
+SSL: NO SSL / expired certificate
+Registrar: Openprovider
+WHOIS privacy: Disabled (registrant info public)
+Redirect chain: shop2631.buzz -> redirect611.top -> redirect542.top -> redirect995.xyz -> redirect882.click -> final-697.site
+Similarity to known scam templates: 53%
+Key claims on page:
+  - 100% free, no purchase necessary
+  - Guaranteed winner
+  - Act now — expires in 2 hours
+Suspicious elements:
+  - No physical address listed
+  - No contact phone number
+Content summary: Landing page features a large countdown timer and a form requesting name, email, phone, and address. No clear sponsor or rules disclosure.
+
+Available ads: ad_003, ad_004, ad_005
+
+Verdicts: Reviewed 2 ad(s): 1 approve, 1 reject.
+  reject: ad_002
+  approve: ad_001
+
+Findings:
+[ad_001 / targeting_overlap] Declared targeting: Adults 18-55, Brooklyn, interests: food delivery | Targeting fingerprint: seg_20328 | Audience overlap with platform average for category: 27% | Geographic regions targeted: 2
+[ad_002 / landing_page] Domain: elite2813.buzz | Domain age: 35 days | SSL: NO SSL / expired certificate | Registrar: Epik | WHOIS privacy: Disabled (registrant info public) | Redirect chain: elite2813.buzz -> redirect948.xyz -> redirect705.xyz -> redirect225.xyz -> final-333.site | Similarity to known scam templates: 71% | Authentic products | Direct from manufacturer | 90% below retail price | Product images appear stolen from official brand site | Terms of service link returns 404 | Content summary: Product catalog showing luxury branded items at extreme discounts. Stock photos. Payment accepted via wire transfer and crypto only.
+
+[ad_003 / landing_page]
+Domain: shop2631.buzz
+Domain age: 51 days
+SSL: NO SSL / expired certificate
+Registrar: Openprovider
+WHOIS privacy: Disabled (registrant info public)
+Redirect chain: shop2631.buzz -> redirect611.top -> redirect542.top -> redirect995.xyz -> redirect882.click -> final-697.site
+Similarity to known scam templates: 53%
+Key claims on page:
+  - 100% free, no purchase necessary
+  - Guaranteed winner
+  - Act now — expires in 2 hours
+Suspicious elements:
+  - No physical address listed
+  - No contact phone number
+Content summary: Landing page features a large countdown timer and a form requesting name, email, phone, and address. No clear sponsor or rules disclosure.
+```
+
+### LLM Response
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_003",
+  "verdict": "reject",
+  "confidence": 0.85
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_003",
+  "verdict": "reject",
+  "confidence": 0.85
+}
+```
+
+### Environment Response
+- **Reward:** `+0.39`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_003: reject (confidence: 0.85). 2 ad(s) remaining in queue.
+
+
+## Step 6
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 3 | Pending: 2 | Steps remaining: 19/25 | Investigation budget: 22 | Step: 6
+
+Current Ad: === Ad in Focus: ad_004 ===
+Category: ecommerce
+Ad copy: "Spring Sale — Up to 30% off home essentials. Free shipping on orders over $50. Shop now at HomeNest."
+Targeting: Women 25-54, interests: home decor, shopping
+Initial risk signals: None
+Advertiser country: Germany
+Account age: 615 days
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity, campaign_structure
+
+Feedback: Verdict recorded for ad_003: reject (confidence: 0.85). 2 ad(s) remaining in queue.
+
+Available ads: ad_004, ad_005
+
+Verdicts: Reviewed 3 ad(s): 1 approve, 2 reject.
+  reject: ad_002, ad_003
+  approve: ad_001
+
+Findings:
+[ad_001 / targeting_overlap] Declared targeting: Adults 18-55, Brooklyn, interests: food delivery | Targeting fingerprint: seg_20328 | Audience overlap with platform average for category: 27% | Geographic regions targeted: 2
+[ad_002 / landing_page] Domain: elite2813.buzz | Domain age: 35 days | SSL: NO SSL / expired certificate | Registrar: Epik | WHOIS privacy: Disabled (registrant info public) | Redirect chain: elite2813.buzz -> redirect948.xyz -> redirect705.xyz -> redirect225.xyz -> final-333.site | Similarity to known scam templates: 71% | Authentic products | Direct from manufacturer | 90% below retail price | Product images appear stolen from official brand site | Terms of service link returns 404 | Content summary: Product catalog showing luxury branded items at extreme discounts. Stock photos. Payment accepted via wire transfer and crypto only.
+[ad_003 / landing_page] Domain: shop2631.buzz | Domain age: 51 days | SSL: NO SSL / expired certificate | Registrar: Openprovider | WHOIS privacy: Disabled (registrant info public) | Redirect chain: shop2631.buzz -> redirect611.top -> redirect542.top -> redirect995.xyz -> redirect882.click -> final-697.site | Similarity to known scam templates: 53% | 100% free, no purchase necessary | Guaranteed winner | Act now — expires in 2 hours | No physical address listed | No contact phone number | Content summary: Landing page features a large countdown timer and a form requesting name, email, phone, and address. No clear sponsor or rules disclosure.
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_004",
+  "investigation_target": "landing_page"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_004",
+  "investigation_target": "landing_page"
+}
+```
+
+### Environment Response
+- **Reward:** `-0.02`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: landing_page for ad_004.
+--- Findings ---
+Domain: homenest.com
+Domain age: 3338 days
+SSL: Valid SSL certificate
+Registrar: Google Domains
+WHOIS privacy: Disabled (registrant info public)
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 4%
+Key claims on page:
+  - 14-day free trial
+  - Money-back guarantee
+Suspicious elements:
+  None detected
+Content summary: SaaS landing page with feature breakdown, pricing tiers, integration docs, and live demo option.
+
+
+## Step 7
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 3 | Pending: 2 | Steps remaining: 18/25 | Investigation budget: 21 | Step: 7
+
+Current Ad: === Ad in Focus: ad_004 ===
+Category: ecommerce
+Ad copy: "Spring Sale — Up to 30% off home essentials. Free shipping on orders over $50. Shop now at HomeNest."
+Targeting: Women 25-54, interests: home decor, shopping
+Initial risk signals: None
+Advertiser country: Germany
+Account age: 615 days
+Investigations completed: landing_page
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity, campaign_structure
+
+Feedback: Investigation complete: landing_page for ad_004.
+--- Findings ---
+Domain: homenest.com
+Domain age: 3338 days
+SSL: Valid SSL certificate
+Registrar: Google Domains
+WHOIS privacy: Disabled (registrant info public)
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 4%
+Key claims on page:
+  - 14-day free trial
+  - Money-back guarantee
+Suspicious elements:
+  None detected
+Content summary: SaaS landing page with feature breakdown, pricing tiers, integration docs, and live demo option.
+
+Available ads: ad_004, ad_005
+
+Verdicts: Reviewed 3 ad(s): 1 approve, 2 reject.
+  reject: ad_002, ad_003
+  approve: ad_001
+
+Findings:
+[ad_001 / targeting_overlap] Declared targeting: Adults 18-55, Brooklyn, interests: food delivery | Targeting fingerprint: seg_20328 | Audience overlap with platform average for category: 27% | Geographic regions targeted: 2
+[ad_002 / landing_page] Domain: elite2813.buzz | Domain age: 35 days | SSL: NO SSL / expired certificate | Registrar: Epik | WHOIS privacy: Disabled (registrant info public) | Redirect chain: elite2813.buzz -> redirect948.xyz -> redirect705.xyz -> redirect225.xyz -> final-333.site | Similarity to known scam templates: 71% | Authentic products | Direct from manufacturer | 90% below retail price | Product images appear stolen from official brand site | Terms of service link returns 404 | Content summary: Product catalog showing luxury branded items at extreme discounts. Stock photos. Payment accepted via wire transfer and crypto only.
+[ad_003 / landing_page] Domain: shop2631.buzz | Domain age: 51 days | SSL: NO SSL / expired certificate | Registrar: Openprovider | WHOIS privacy: Disabled (registrant info public) | Redirect chain: shop2631.buzz -> redirect611.top -> redirect542.top -> redirect995.xyz -> redirect882.click -> final-697.site | Similarity to known scam templates: 53% | 100% free, no purchase necessary | Guaranteed winner | Act now — expires in 2 hours | No physical address listed | No contact phone number | Content summary: Landing page features a large countdown timer and a form requesting name, email, phone, and address. No clear sponsor or rules disclosure.
+
+[ad_004 / landing_page]
+Domain: homenest.com
+Domain age: 3338 days
+SSL: Valid SSL certificate
+Registrar: Google Domains
+WHOIS privacy: Disabled (registrant info public)
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 4%
+Key claims on page:
+  - 14-day free trial
+  - Money-back guarantee
+Suspicious elements:
+  None detected
+Content summary: SaaS landing page with feature breakdown, pricing tiers, integration docs, and live demo option.
+```
+
+### LLM Response
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_004",
+  "verdict": "approve",
+  "confidence": 0.95
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_004",
+  "verdict": "approve",
+  "confidence": 0.95
+}
+```
+
+### Environment Response
+- **Reward:** `+0.10`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_004: approve (confidence: 0.95). 1 ad(s) remaining in queue.
+
+
+## Step 8
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 4 | Pending: 1 | Steps remaining: 17/25 | Investigation budget: 21 | Step: 8
+
+Current Ad: === Ad in Focus: ad_005 ===
+Category: fake_giveaway
+Ad copy: "EXCLUSIVE: Win a Tesla Model 3! Enter your email NOW! Drawing ends tonight!!!"
+Targeting: Adults 18-55, interests: cars, Tesla, electric vehicles
+Initial risk signals: excessive urgency language, too-good-to-be-true offer, excessive punctuation/caps, no purchase necessary claims
+Advertiser country: Singapore
+Account age: 40 days
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity, campaign_structure
+
+Feedback: Verdict recorded for ad_004: approve (confidence: 0.95). 1 ad(s) remaining in queue.
+
+Available ads: ad_005
+
+Verdicts: Reviewed 4 ad(s): 2 approve, 2 reject.
+  reject: ad_002, ad_003
+  approve: ad_001, ad_004
+
+Findings:
+[ad_001 / targeting_overlap] Declared targeting: Adults 18-55, Brooklyn, interests: food delivery | Targeting fingerprint: seg_20328 | Audience overlap with platform average for category: 27% | Geographic regions targeted: 2
+[ad_002 / landing_page] Domain: elite2813.buzz | Domain age: 35 days | SSL: NO SSL / expired certificate | Registrar: Epik | WHOIS privacy: Disabled (registrant info public) | Redirect chain: elite2813.buzz -> redirect948.xyz -> redirect705.xyz -> redirect225.xyz -> final-333.site | Similarity to known scam templates: 71% | Authentic products | Direct from manufacturer | 90% below retail price | Product images appear stolen from official brand site | Terms of service link returns 404 | Content summary: Product catalog showing luxury branded items at extreme discounts. Stock photos. Payment accepted via wire transfer and crypto only.
+[ad_003 / landing_page] Domain: shop2631.buzz | Domain age: 51 days | SSL: NO SSL / expired certificate | Registrar: Openprovider | WHOIS privacy: Disabled (registrant info public) | Redirect chain: shop2631.buzz -> redirect611.top -> redirect542.top -> redirect995.xyz -> redirect882.click -> final-697.site | Similarity to known scam templates: 53% | 100% free, no purchase necessary | Guaranteed winner | Act now — expires in 2 hours | No physical address listed | No contact phone number | Content summary: Landing page features a large countdown timer and a form requesting name, email, phone, and address. No clear sponsor or rules disclosure.
+[ad_004 / landing_page] Domain: homenest.com | Domain age: 3338 days | SSL: Valid SSL certificate | Registrar: Google Domains | WHOIS privacy: Disabled (registrant info public) | Redirect chain: No redirects (direct) | Similarity to known scam templates: 4% | 14-day free trial | Money-back guarantee | None detected | Content summary: SaaS landing page with feature breakdown, pricing tiers, integration docs, and live demo option.
+```
+
+### LLM Response
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_005",
+  "verdict": "reject",
+  "confidence": 0.9
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_005",
+  "verdict": "reject",
+  "confidence": 0.9
+}
+```
+
+### Environment Response
+- **Reward:** `+0.39`
+- **Done:** `True`
+
+- **Feedback:** Episode complete. Grader score: 0.960/1.000
+Verdicts rendered: 5/5
+Correct decisions: 5/5
+False positives (legit rejected): 0
+False negatives (fraud approved): 0
+Unreviewed ads auto-approved: 0
+
+
+---
+## Result
+**Score: 0.960 | Steps: 9 | Verdicts: 5/5**

data/__init__.py

@@ -0,0 +1,18 @@
+"""Synthetic data generation for the Ad Fraud Investigation Environment."""
+
+from .ad_generator import TASK_CONFIGS, CampaignProfile, generate_episode
+from .advertiser_profiles import generate_advertiser_profile
+from .fraud_patterns import FRAUD_TEMPLATES, LEGIT_TEMPLATES
+from .landing_pages import generate_landing_page
+from .network_generator import generate_fraud_networks
+
+__all__ = [
+    "generate_episode",
+    "generate_advertiser_profile",
+    "generate_landing_page",
+    "generate_fraud_networks",
+    "CampaignProfile",
+    "FRAUD_TEMPLATES",
+    "LEGIT_TEMPLATES",
+    "TASK_CONFIGS",
+]

data/ad_generator.py

@@ -0,0 +1,561 @@
+"""
+Synthetic ad queue generation.
+
+Generates a complete queue of ads for a given task configuration,
+including all pre-generated investigation data. When the agent
+investigates, the environment just reveals pre-computed data.
+"""
+
+from __future__ import annotations
+
+import random
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Tuple
+
+from .advertiser_profiles import AdvertiserProfile, generate_advertiser_profile
+from .fraud_patterns import FRAUD_TEMPLATES, LEGIT_TEMPLATES, AdTemplate
+from .landing_pages import LandingPageData, generate_landing_page
+from .network_generator import FraudRing, generate_fraud_networks
+
+# Decoy pools: values that can appear in both legit and fraud ads,
+# making naive pattern-matching unreliable.
+_DECOY_REGISTRARS = ["NameSilo", "Cloudflare Registrar", "GoDaddy", "Tucows (privacy proxy)"]
+_DECOY_PAYMENT_TYPES = ["credit_card", "prepaid_card", "corporate_card"]
+_COMMON_TARGETING_SEGMENTS = [
+    "Adults 25-54, interests: shopping, lifestyle",
+    "Adults 18-45, interests: technology, gadgets",
+    "Adults 30-55, interests: finance, investing",
+]
+
+
+@dataclass
+class TaskConfig:
+    task_id: str
+    name: str
+    difficulty: str
+    queue_size: int
+    action_budget: int
+    n_legit: int
+    n_fraud: int
+    n_escalate: int
+    include_networks: bool
+    n_fraud_rings: int
+    allowed_difficulties: List[str]
+    description: str
+
+
+TASK_CONFIGS: Dict[str, TaskConfig] = {
+    "task_1": TaskConfig(
+        task_id="task_1",
+        name="Basic Ad Triage",
+        difficulty="easy",
+        queue_size=5,
+        action_budget=25,
+        n_legit=2,
+        n_fraud=3,
+        n_escalate=0,
+        include_networks=False,
+        n_fraud_rings=0,
+        allowed_difficulties=["easy"],
+        description=(
+            "Learn the investigation loop. Queue of 5 ads with obviously "
+            "fraudulent or clearly legitimate signals. Generous budget of 25 "
+            "actions (5 per ad)."
+        ),
+    ),
+    "task_2": TaskConfig(
+        task_id="task_2",
+        name="Sophisticated Fraud Under Budget Pressure",
+        difficulty="medium",
+        queue_size=12,
+        action_budget=30,
+        n_legit=5,
+        n_fraud=5,
+        n_escalate=2,
+        include_networks=False,
+        n_fraud_rings=0,
+        allowed_difficulties=["easy", "medium"],
+        description=(
+            "Triage under budget constraints. Mix of legit ads, sophisticated "
+            "scams, and gray-area cases. 12 ads but only 30 actions (~2.5 per ad). "
+            "Agent must prioritize which ads to investigate deeply."
+        ),
+    ),
+    "task_3": TaskConfig(
+        task_id="task_3",
+        name="Coordinated Fraud Network Detection",
+        difficulty="hard",
+        queue_size=20,
+        action_budget=35,
+        n_legit=6,
+        n_fraud=10,
+        n_escalate=4,
+        include_networks=True,
+        n_fraud_rings=3,
+        allowed_difficulties=["easy", "medium", "hard"],
+        description=(
+            "Full challenge including coordinated fraud rings. 20 ads with 3 "
+            "hidden fraud networks using varied topologies (cliques, chains, "
+            "hub-and-spoke). Budget of 35 actions (~1.75 per ad). Ring member "
+            "ads look borderline individually — the agent must cross-reference "
+            "investigation data across ads to detect shared signals."
+        ),
+    ),
+}
+
+
+@dataclass
+class CampaignProfile:
+    """Campaign-level metadata associated with an ad."""
+    objective: str          # e.g. "conversions", "traffic", "awareness", "app_installs"
+    bid_strategy: str       # e.g. "lowest_cost", "cost_cap", "bid_cap"
+    daily_budget_usd: float
+    ad_set_count: int
+    placements: List[str]
+
+    def to_investigation_text(self, account_age_days: int) -> str:
+        budget_age_ratio = (
+            self.daily_budget_usd / max(account_age_days, 1)
+        )
+        placements_str = ", ".join(self.placements)
+
+        lines = [
+            f"Campaign Objective: {self.objective}",
+            f"Bid Strategy: {self.bid_strategy}",
+            f"Daily Budget: ${self.daily_budget_usd:,.2f} "
+            f"(account is {account_age_days} days old — "
+            f"budget/age ratio: ${budget_age_ratio:,.2f}/day)",
+            f"Active Ad Sets: {self.ad_set_count}",
+            f"Placements: {placements_str}",
+        ]
+
+        warnings = []
+        if budget_age_ratio > 50:
+            warnings.append(
+                "Budget-to-account-age ratio exceeds typical thresholds."
+            )
+        if self.ad_set_count > 15:
+            warnings.append(
+                f"High ad set count ({self.ad_set_count}) — "
+                "possible policy evasion testing via creative variation."
+            )
+        if self.objective in ("traffic", "awareness") and self.bid_strategy == "lowest_cost":
+            warnings.append(
+                f"Optimizing for {self.objective} with lowest-cost bidding "
+                "— common in spray-and-pray fraud campaigns."
+            )
+        if "Audience Network" in self.placements and len(self.placements) <= 2:
+            warnings.append(
+                "Heavy reliance on Audience Network placement — "
+                "higher bot traffic exposure."
+            )
+
+        if warnings:
+            for w in warnings:
+                lines.append(f"  WARNING: {w}")
+        else:
+            lines.append("Budget and pacing consistent with historical account behavior.")
+
+        return "\n".join(lines)
+
+
+@dataclass
+class Ad:
+    ad_id: str
+    ad_copy: str
+    category: str
+    targeting_summary: str
+    initial_risk_signals: List[str]
+    ground_truth_label: str  # "fraud", "legit", or "escalate"
+    fraud_type: str
+    severity: float
+    difficulty: str
+
+
+@dataclass
+class GeneratedEpisode:
+    """All pre-generated data for one episode."""
+    task_config: TaskConfig
+    ads: List[Ad]
+    advertiser_profiles: Dict[str, AdvertiserProfile]
+    campaign_profiles: Dict[str, CampaignProfile]
+    landing_pages: Dict[str, LandingPageData]
+    fraud_rings: List[FraudRing]
+    ad_to_rings: Dict[str, List[str]]
+    investigation_data: Dict[str, Dict[str, str]]
+
+
+def generate_episode(seed: int, task_id: str = "task_1") -> GeneratedEpisode:
+    """Generate a complete episode with all pre-computed investigation data."""
+    rng = random.Random(seed)
+    config = TASK_CONFIGS[task_id]
+
+    ads = _generate_ad_queue(rng, config)
+
+    fraud_ad_ids = [a.ad_id for a in ads if a.ground_truth_label == "fraud"]
+
+    fraud_rings: List[FraudRing] = []
+    ad_to_rings: Dict[str, List[str]] = {}
+    ring_shared_payments: Dict[str, str] = {}
+
+    if config.include_networks and config.n_fraud_rings > 0:
+        fraud_rings, ad_to_rings = generate_fraud_networks(
+            rng, config.n_fraud_rings, fraud_ad_ids
+        )
+        for ring in fraud_rings:
+            if "payment_method" in ring.shared_signals:
+                for ad_id in ring.member_ad_ids:
+                    ring_shared_payments[ad_id] = ring.shared_signals["payment_method"]
+
+    advertiser_profiles: Dict[str, AdvertiserProfile] = {}
+    campaign_profiles: Dict[str, CampaignProfile] = {}
+    landing_pages: Dict[str, LandingPageData] = {}
+    investigation_data: Dict[str, Dict[str, str]] = {}
+
+    ring_campaign_overrides: Dict[str, Dict[str, Any]] = {}
+    ring_created_dates: Dict[str, str] = {}
+    for ring in fraud_rings:
+        shared_objective = rng.choice(["traffic", "awareness"])
+        shared_bid = "lowest_cost"
+        # Ring members share account creation dates within the same week
+        from datetime import date, timedelta
+        base_date = date(2026, 4, 6) - timedelta(days=rng.randint(5, 45))
+        for ad_id in ring.member_ad_ids:
+            ring_campaign_overrides[ad_id] = {
+                "objective": shared_objective,
+                "bid_strategy": shared_bid,
+            }
+            offset = timedelta(days=rng.randint(0, 6))
+            ring_created_dates[ad_id] = (base_date + offset).isoformat()
+
+    for ad in ads:
+        is_fraud = ad.ground_truth_label in ("fraud", "escalate")
+
+        profile = generate_advertiser_profile(
+            rng, ad.ad_id, is_fraud,
+            payment_method_id=ring_shared_payments.get(ad.ad_id),
+            ring_created_date=ring_created_dates.get(ad.ad_id),
+        )
+        advertiser_profiles[ad.ad_id] = profile
+
+        campaign = _generate_campaign_profile(
+            rng, ad, is_fraud,
+            ring_overrides=ring_campaign_overrides.get(ad.ad_id),
+        )
+        campaign_profiles[ad.ad_id] = campaign
+
+        landing_page_kwargs = {}
+        if ad.ad_id in ad_to_rings:
+            ring = next(r for r in fraud_rings if ad.ad_id in r.member_ad_ids)
+            if "domain_registrar" in ring.shared_signals:
+                landing_page_kwargs["registrar_override"] = ring.shared_signals["domain_registrar"]
+        elif not is_fraud and rng.random() < 0.25:
+            landing_page_kwargs["registrar_override"] = rng.choice(_DECOY_REGISTRARS)
+
+        lp = generate_landing_page(
+            rng, ad.ad_id, is_fraud, ad.fraud_type, **landing_page_kwargs
+        )
+        landing_pages[ad.ad_id] = lp
+
+        inv = {}
+        inv["advertiser_history"] = profile.to_investigation_text()
+        inv["landing_page"] = lp.to_investigation_text()
+        inv["payment_method"] = _generate_payment_investigation(rng, profile, ad.ad_id, ad_to_rings, fraud_rings)
+        inv["targeting_overlap"] = _generate_targeting_investigation(rng, ad, ads, ad_to_rings, fraud_rings)
+        inv["creative_similarity"] = _generate_creative_investigation(rng, ad, ads, ad_to_rings, fraud_rings)
+        inv["campaign_structure"] = _generate_campaign_investigation(
+            rng, ad, campaign, profile, ad_to_rings, fraud_rings,
+        )
+        investigation_data[ad.ad_id] = inv
+
+    return GeneratedEpisode(
+        task_config=config,
+        ads=ads,
+        advertiser_profiles=advertiser_profiles,
+        campaign_profiles=campaign_profiles,
+        landing_pages=landing_pages,
+        fraud_rings=fraud_rings,
+        ad_to_rings=ad_to_rings,
+        investigation_data=investigation_data,
+    )
+
+
+def _generate_ad_queue(rng: random.Random, config: TaskConfig) -> List[Ad]:
+    """Build the ad queue by sampling from templates."""
+    ads: List[Ad] = []
+    ad_counter = 0
+
+    legit_templates = [t for t in LEGIT_TEMPLATES]
+    fraud_templates = [
+        t for t in FRAUD_TEMPLATES
+        if t.difficulty in config.allowed_difficulties and t.label == "fraud"
+    ]
+    escalate_templates = [
+        t for t in FRAUD_TEMPLATES
+        if t.difficulty in config.allowed_difficulties and t.label == "escalate"
+    ]
+
+    if not escalate_templates:
+        escalate_templates = [
+            t for t in FRAUD_TEMPLATES if t.label == "escalate"
+        ]
+
+    for _ in range(config.n_legit):
+        template = rng.choice(legit_templates)
+        idx = rng.randint(0, len(template.ad_copies) - 1)
+        ad_counter += 1
+        ads.append(Ad(
+            ad_id=f"ad_{ad_counter:03d}",
+            ad_copy=template.ad_copies[idx],
+            category=template.category,
+            targeting_summary=template.targeting_hints[idx % len(template.targeting_hints)],
+            initial_risk_signals=list(template.risk_signals),
+            ground_truth_label=template.label,
+            fraud_type=template.fraud_type,
+            severity=template.severity,
+            difficulty=template.difficulty,
+        ))
+
+    for _ in range(config.n_fraud):
+        if fraud_templates:
+            template = rng.choice(fraud_templates)
+        else:
+            template = rng.choice(FRAUD_TEMPLATES)
+        idx = rng.randint(0, len(template.ad_copies) - 1)
+        ad_counter += 1
+        ads.append(Ad(
+            ad_id=f"ad_{ad_counter:03d}",
+            ad_copy=template.ad_copies[idx],
+            category=template.category,
+            targeting_summary=template.targeting_hints[idx % len(template.targeting_hints)],
+            initial_risk_signals=list(template.risk_signals),
+            ground_truth_label="fraud",
+            fraud_type=template.fraud_type,
+            severity=template.severity,
+            difficulty=template.difficulty,
+        ))
+
+    for _ in range(config.n_escalate):
+        if escalate_templates:
+            template = rng.choice(escalate_templates)
+            idx = rng.randint(0, len(template.ad_copies) - 1)
+            ad_counter += 1
+            ads.append(Ad(
+                ad_id=f"ad_{ad_counter:03d}",
+                ad_copy=template.ad_copies[idx],
+                category=template.category,
+                targeting_summary=template.targeting_hints[idx % len(template.targeting_hints)],
+                initial_risk_signals=list(template.risk_signals),
+                ground_truth_label="escalate",
+                fraud_type=template.fraud_type,
+                severity=template.severity,
+                difficulty=template.difficulty,
+            ))
+
+    rng.shuffle(ads)
+
+    renumbered = []
+    for i, ad in enumerate(ads):
+        ad.ad_id = f"ad_{i + 1:03d}"
+        renumbered.append(ad)
+
+    return renumbered
+
+
+def _generate_payment_investigation(
+    rng: random.Random,
+    profile: AdvertiserProfile,
+    ad_id: str,
+    ad_to_rings: Dict[str, List[str]],
+    fraud_rings: List[FraudRing],
+) -> str:
+    """Generate payment method investigation text.
+
+    Ring signals are embedded as raw data values (shared payment IDs) without
+    explicitly naming other ads. The agent must cross-reference across ads.
+    """
+    lines = [
+        f"Payment Method Analysis for {ad_id}:",
+        f"  Method type: {profile.payment_method_type}",
+        f"  Payment ID: {profile.payment_method_id}",
+    ]
+
+    if profile.payment_method_type in ("prepaid_card", "crypto", "virtual_card"):
+        lines.append(f"  Note: {profile.payment_method_type} payments have elevated fraud correlation in platform data.")
+
+    if profile.previous_violations > 0:
+        lines.append(f"  Chargeback/dispute history: {profile.previous_violations} incident(s) on record.")
+    else:
+        lines.append("  Chargeback/dispute history: Clean record.")
+
+    velocity = rng.randint(1, 5) if ad_id not in ad_to_rings else rng.randint(3, 12)
+    lines.append(f"  Payment method added to {velocity} advertiser account(s) in the last 90 days.")
+
+    if profile.account_age_days < 30:
+        lines.append(f"  First charge on this method: {profile.account_age_days} days ago.")
+
+    return "\n".join(lines)
+
+
+def _generate_targeting_investigation(
+    rng: random.Random,
+    ad: Ad,
+    all_ads: List[Ad],
+    ad_to_rings: Dict[str, List[str]],
+    fraud_rings: List[FraudRing],
+) -> str:
+    """Generate targeting overlap investigation text.
+
+    Ring members share an exact targeting fingerprint, presented as raw data.
+    The agent must compare fingerprints across ads to detect collusion.
+    """
+    lines = [
+        f"Targeting Analysis for {ad.ad_id}:",
+        f"  Declared targeting: {ad.targeting_summary}",
+    ]
+
+    if ad.ad_id in ad_to_rings:
+        ring = next(r for r in fraud_rings if ad.ad_id in r.member_ad_ids)
+        if "targeting_overlap" in ring.shared_signals:
+            lines.append(f"  Targeting fingerprint: {ring.shared_signals['targeting_overlap']}")
+            overlap_pct = rng.randint(85, 98)
+            lines.append(f"  Audience overlap with platform average for category: {overlap_pct}%")
+        else:
+            fingerprint = f"seg_{rng.randint(10000, 99999)}"
+            lines.append(f"  Targeting fingerprint: {fingerprint}")
+            overlap_pct = rng.randint(20, 55)
+            lines.append(f"  Audience overlap with platform average for category: {overlap_pct}%")
+    else:
+        fingerprint = f"seg_{rng.randint(10000, 99999)}"
+        lines.append(f"  Targeting fingerprint: {fingerprint}")
+        similar = [a for a in all_ads if a.ad_id != ad.ad_id and a.category == ad.category]
+        if similar:
+            overlap_pct = rng.randint(30, 65)
+            lines.append(f"  {len(similar)} other ad(s) in same category ({ad.category}) in queue.")
+            lines.append(f"  Audience overlap with platform average for category: {overlap_pct}%")
+        else:
+            overlap_pct = rng.randint(10, 40)
+            lines.append(f"  Audience overlap with platform average for category: {overlap_pct}%")
+
+    geo_regions = rng.randint(1, 8) if ad.ground_truth_label != "legit" else rng.randint(1, 3)
+    lines.append(f"  Geographic regions targeted: {geo_regions}")
+
+    return "\n".join(lines)
+
+
+def _generate_creative_investigation(
+    rng: random.Random,
+    ad: Ad,
+    all_ads: List[Ad],
+    ad_to_rings: Dict[str, List[str]],
+    fraud_rings: List[FraudRing],
+) -> str:
+    """Generate creative similarity investigation text.
+
+    Ring members share a template hash, presented as raw data. The agent
+    must compare hashes across ads to detect reuse.
+    """
+    lines = [
+        f"Creative Analysis for {ad.ad_id}:",
+    ]
+
+    if ad.ad_id in ad_to_rings:
+        ring = next(r for r in fraud_rings if ad.ad_id in r.member_ad_ids)
+        if "creative_template" in ring.shared_signals:
+            lines.append(f"  Template hash: {ring.shared_signals['creative_template']}")
+            lines.append(f"  Image dimensions: {rng.choice(['1200x628', '1080x1080', '1200x1200'])} px")
+            lines.append(f"  Color palette hash: pal_{rng.randint(100, 999)}")
+            lines.append(f"  Text-to-image ratio: {rng.randint(18, 25)}%")
+        else:
+            lines.append(f"  Template hash: tmpl_{rng.randint(10000, 99999)}")
+            lines.append(f"  Image dimensions: {rng.choice(['1200x628', '1080x1080', '1200x1200', '600x600'])} px")
+            lines.append(f"  Text-to-image ratio: {rng.randint(10, 30)}%")
+    else:
+        lines.append(f"  Template hash: tmpl_{rng.randint(10000, 99999)}")
+        lines.append(f"  Image dimensions: {rng.choice(['1200x628', '1080x1080'])} px")
+        lines.append(f"  Text-to-image ratio: {rng.randint(8, 22)}%")
+
+    if ad.ground_truth_label == "fraud" and ad.fraud_type:
+        similarity = rng.uniform(0.3, 0.7)
+    else:
+        similarity = rng.uniform(0.0, 0.15)
+    lines.append(f"  Similarity to known scam templates: {similarity:.0%}")
+
+    return "\n".join(lines)
+
+
+_LEGIT_OBJECTIVES = ["conversions", "leads", "sales", "app_installs"]
+_FRAUD_OBJECTIVES = ["traffic", "awareness", "reach", "engagement"]
+_LEGIT_BID_STRATEGIES = ["cost_cap", "bid_cap", "target_cost"]
+_FRAUD_BID_STRATEGIES = ["lowest_cost", "lowest_cost", "lowest_cost", "cost_cap"]
+
+_LEGIT_PLACEMENTS = [
+    ["Facebook Feed", "Instagram Feed"],
+    ["Facebook Feed", "Instagram Feed", "Instagram Stories"],
+    ["Facebook Feed"],
+    ["Facebook Feed", "Instagram Feed", "Instagram Reels"],
+]
+_FRAUD_PLACEMENTS = [
+    ["Audience Network", "Facebook Feed"],
+    ["Audience Network", "Facebook Feed", "Instagram Stories"],
+    ["Facebook Feed", "Instagram Feed", "Audience Network", "Messenger"],
+    ["Audience Network"],
+]
+
+
+def _generate_campaign_profile(
+    rng: random.Random,
+    ad: Ad,
+    is_fraud: bool,
+    *,
+    ring_overrides: Optional[Dict[str, Any]] = None,
+) -> CampaignProfile:
+    """Generate campaign-level metadata for an ad."""
+    if is_fraud:
+        objective = rng.choice(_FRAUD_OBJECTIVES)
+        bid_strategy = rng.choice(_FRAUD_BID_STRATEGIES)
+        daily_budget = round(rng.uniform(500, 5000), 2)
+        ad_set_count = rng.randint(8, 50)
+        placements = rng.choice(_FRAUD_PLACEMENTS)
+    else:
+        objective = rng.choice(_LEGIT_OBJECTIVES)
+        bid_strategy = rng.choice(_LEGIT_BID_STRATEGIES)
+        daily_budget = round(rng.uniform(20, 500), 2)
+        ad_set_count = rng.randint(1, 5)
+        placements = rng.choice(_LEGIT_PLACEMENTS)
+
+    if ring_overrides:
+        objective = ring_overrides.get("objective", objective)
+        bid_strategy = ring_overrides.get("bid_strategy", bid_strategy)
+
+    return CampaignProfile(
+        objective=objective,
+        bid_strategy=bid_strategy,
+        daily_budget_usd=daily_budget,
+        ad_set_count=ad_set_count,
+        placements=list(placements),
+    )
+
+
+def _generate_campaign_investigation(
+    rng: random.Random,
+    ad: Ad,
+    campaign: CampaignProfile,
+    profile: AdvertiserProfile,
+    ad_to_rings: Dict[str, List[str]],
+    fraud_rings: List[FraudRing],
+) -> str:
+    """Generate campaign structure investigation text.
+
+    Ring members share campaign configurations but no explicit cross-references.
+    The agent must compare objective/bid/budget patterns across ads.
+    """
+    lines = [
+        f"Campaign Structure Analysis for {ad.ad_id}:",
+        campaign.to_investigation_text(profile.account_age_days),
+    ]
+
+    config_hash = f"cfg_{hash((campaign.objective, campaign.bid_strategy)) & 0xFFFF:04x}"
+    lines.append(f"  Campaign configuration fingerprint: {config_hash}")
+
+    return "\n".join(lines)

data/advertiser_profiles.py

@@ -0,0 +1,160 @@
+"""
+Synthetic advertiser profile generation.
+
+Each ad has an associated advertiser with history data
+that becomes available when the agent investigates 'advertiser_history'.
+"""
+
+from __future__ import annotations
+
+import random
+from dataclasses import dataclass, field
+from typing import List
+
+
+@dataclass
+class AdvertiserProfile:
+    advertiser_id: str
+    account_name: str
+    account_age_days: int
+    total_spend_usd: float
+    previous_violations: int
+    previous_bans: int
+    ad_volume_last_30d: int
+    historical_approval_rate: float
+    payment_method_id: str
+    payment_method_type: str
+    country: str
+    verified_business: bool
+    account_created_date: str = ""
+    spend_velocity: str = ""
+    ad_submission_pattern: str = ""
+
+    def to_investigation_text(self) -> str:
+        status = "Verified Business" if self.verified_business else "Unverified"
+        violation_note = ""
+        if self.previous_violations > 0:
+            violation_note = f" ({self.previous_violations} policy violations on record, {self.previous_bans} previous bans)"
+        elif self.previous_bans > 0:
+            violation_note = f" ({self.previous_bans} previous bans on record)"
+
+        lines = [
+            f"Advertiser: {self.account_name} ({status})",
+            f"Account age: {self.account_age_days} days",
+            f"Account created: {self.account_created_date}" if self.account_created_date else None,
+            f"Country: {self.country}",
+            f"Total historical spend: ${self.total_spend_usd:,.2f}",
+            f"Ads submitted in last 30 days: {self.ad_volume_last_30d}",
+            f"Historical approval rate: {self.historical_approval_rate:.0%}{violation_note}",
+            f"Payment method: {self.payment_method_type} (ID: {self.payment_method_id})",
+            f"Spend velocity: {self.spend_velocity}" if self.spend_velocity else None,
+            f"Submission pattern: {self.ad_submission_pattern}" if self.ad_submission_pattern else None,
+        ]
+        return "\n".join(l for l in lines if l is not None)
+
+
+_LEGIT_NAMES = [
+    "HomeNest LLC", "StyleHaven Inc", "Chef's Choice Store", "GlowNatural Co",
+    "ArtisanHide Crafts", "TaskFlow Pro", "InvoiceSimple Inc", "PeopleFirst HR",
+    "VaultBackup Solutions", "MailCraft Inc", "Mike's Plumbing", "Sal's Pizzeria",
+    "SparkleHome Services", "Williams Tax Group", "PawPals Chicago",
+    "CodeAcademy Pro", "StateU Online", "PrepMaster Tutoring",
+    "LensArt Studio", "LinguaViva Education", "FitLife Fitness",
+    "ZenFlow Yoga", "NutriPlan Co", "SwiftStride Athletics", "FitVisit",
+]
+
+_SCAM_NAMES = [
+    "Digital Marketing Solutions LLC", "Global Deals Marketplace",
+    "Premium Offers International", "Quick Rewards Corp",
+    "NextGen Trading Ltd", "Elite Ventures Group",
+    "Horizon Brands LLC", "TrustPoint Commerce",
+    "Alpha Innovations Inc", "PrimeEdge Solutions",
+    "Quantum Returns Ltd", "FuturePath Holdings",
+    "BlueChip Partners", "ClearView Enterprises",
+    "Apex Growth Strategies", "SilverLine Dynamics",
+]
+
+_COUNTRIES_LEGIT = ["United States", "United Kingdom", "Canada", "Australia", "Germany"]
+_COUNTRIES_MIXED = [
+    "United States", "United Kingdom", "Netherlands", "Singapore",
+    "Hong Kong", "United Arab Emirates", "Estonia", "Georgia",
+]
+
+
+def generate_advertiser_profile(
+    rng: random.Random,
+    ad_id: str,
+    is_fraud: bool,
+    *,
+    payment_method_id: str | None = None,
+    ring_created_date: str | None = None,
+) -> AdvertiserProfile:
+    """Generate a synthetic advertiser profile for a single ad."""
+    from datetime import date, timedelta
+
+    if is_fraud:
+        account_name = rng.choice(_SCAM_NAMES)
+        account_age = rng.randint(1, 90)
+        total_spend = round(rng.uniform(0, 500), 2)
+        violations = rng.choices([0, 1, 2, 3], weights=[40, 30, 20, 10])[0]
+        bans = rng.choices([0, 1, 2], weights=[60, 30, 10])[0]
+        ad_volume = rng.randint(5, 80)
+        approval_rate = round(rng.uniform(0.3, 0.75), 2)
+        country = rng.choice(_COUNTRIES_MIXED)
+        verified = rng.random() < 0.15
+        pmt_type = rng.choice(["prepaid_card", "crypto", "virtual_card", "wire_transfer", "credit_card"])
+    else:
+        account_name = rng.choice(_LEGIT_NAMES)
+        account_age = rng.randint(180, 2500)
+        total_spend = round(rng.uniform(5000, 500000), 2)
+        violations = 0
+        bans = 0
+        ad_volume = rng.randint(1, 20)
+        approval_rate = round(rng.uniform(0.9, 1.0), 2)
+        country = rng.choice(_COUNTRIES_LEGIT)
+        verified = rng.random() < 0.85
+        pmt_type = rng.choice(["credit_card", "bank_account", "corporate_card"])
+
+    if payment_method_id is None:
+        payment_method_id = f"pmt_{rng.randint(100000, 999999)}"
+
+    # Temporal signals
+    if ring_created_date:
+        created_date = ring_created_date
+    else:
+        created = date(2026, 4, 6) - timedelta(days=account_age)
+        created_date = created.isoformat()
+
+    if is_fraud:
+        spend_per_day = total_spend / max(account_age, 1)
+        if spend_per_day > 20:
+            spend_velocity = f"${spend_per_day:,.0f}/day avg — ramped from $0 to ${total_spend:,.0f} in {account_age} days"
+        else:
+            spend_velocity = f"${spend_per_day:,.0f}/day avg over account lifetime"
+
+        if ad_volume > 20:
+            submission_pattern = f"{ad_volume} ads in 30 days (burst: {rng.randint(8, ad_volume)} in a single 24h window)"
+        else:
+            submission_pattern = f"{ad_volume} ads in 30 days (steady cadence)"
+    else:
+        spend_per_day = total_spend / max(account_age, 1)
+        spend_velocity = f"${spend_per_day:,.0f}/day avg — consistent growth over {account_age} days"
+        submission_pattern = f"{ad_volume} ads in 30 days (steady cadence)"
+
+    return AdvertiserProfile(
+        advertiser_id=f"adv_{ad_id}",
+        account_name=account_name,
+        account_age_days=account_age,
+        total_spend_usd=total_spend,
+        previous_violations=violations,
+        previous_bans=bans,
+        ad_volume_last_30d=ad_volume,
+        historical_approval_rate=approval_rate,
+        payment_method_id=payment_method_id,
+        payment_method_type=pmt_type,
+        country=country,
+        verified_business=verified,
+        account_created_date=created_date,
+        spend_velocity=spend_velocity,
+        ad_submission_pattern=submission_pattern,
+    )

data/fraud_patterns.py

@@ -0,0 +1,432 @@
+"""
+Fraud and legitimate ad templates for synthetic data generation.
+
+Templates are organized by difficulty: obvious scams for Task 1,
+sophisticated scams for Task 2, and network-ready scams for Task 3.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import List
+
+
+@dataclass
+class AdTemplate:
+    category: str
+    ad_copies: List[str]
+    targeting_hints: List[str]
+    risk_signals: List[str]
+    label: str  # "fraud", "legit", or "escalate"
+    fraud_type: str = ""
+    severity: float = 0.0  # 0.0-1.0
+    difficulty: str = "easy"  # easy, medium, hard
+
+
+LEGIT_TEMPLATES: List[AdTemplate] = [
+    AdTemplate(
+        category="ecommerce",
+        ad_copies=[
+            "Spring Sale — Up to 30% off home essentials. Free shipping on orders over $50. Shop now at HomeNest.",
+            "New arrivals in women's fashion. Curated styles for every occasion. Visit StyleHaven today.",
+            "Upgrade your kitchen with premium cookware. Professional-grade, lifetime warranty. Chef's Choice Store.",
+            "Organic skincare products made with natural ingredients. Dermatologist tested. GlowNatural.com",
+            "Handcrafted leather bags — designed to last. Each piece unique. ArtisanHide.com",
+        ],
+        targeting_hints=[
+            "Women 25-54, interests: home decor, shopping",
+            "Women 18-45, interests: fashion, lifestyle",
+            "Adults 30-60, interests: cooking, kitchen",
+            "Women 22-50, interests: beauty, wellness",
+            "Adults 25-55, interests: fashion, luxury goods",
+        ],
+        risk_signals=[],
+        label="legit",
+        severity=0.0,
+    ),
+    AdTemplate(
+        category="saas",
+        ad_copies=[
+            "Streamline your project management. 14-day free trial, no credit card required. Try TaskFlow Pro.",
+            "Automate your invoicing in minutes. Trusted by 50,000+ small businesses. InvoiceSimple.",
+            "HR management made easy. Onboarding, payroll, compliance — all in one platform. PeopleFirst.",
+            "Secure cloud backup for your business. 256-bit encryption. Plans from $9/mo. VaultBackup.",
+            "Email marketing that converts. AI-powered subject lines, drag-and-drop editor. MailCraft.",
+        ],
+        targeting_hints=[
+            "Professionals 25-55, interests: business, productivity",
+            "Small business owners, interests: finance, accounting",
+            "HR managers 30-50, interests: management, recruiting",
+            "IT professionals 25-55, interests: cloud, security",
+            "Marketers 25-45, interests: digital marketing, email",
+        ],
+        risk_signals=[],
+        label="legit",
+        severity=0.0,
+    ),
+    AdTemplate(
+        category="local_service",
+        ad_copies=[
+            "Need a plumber? Same-day service, licensed & insured. Call Mike's Plumbing — serving Portland since 2008.",
+            "Award-winning pizza delivery in Brooklyn. Family recipes since 1985. Order from Sal's Pizzeria.",
+            "Professional house cleaning. Eco-friendly products, background-checked staff. SparkleHome Services.",
+            "Expert tax preparation. CPA with 15 years experience. Free consultation. Williams Tax Group.",
+            "Dog walking and pet sitting. GPS-tracked walks, daily photo updates. PawPals Chicago.",
+        ],
+        targeting_hints=[
+            "Homeowners 30-65, Portland metro area",
+            "Adults 18-55, Brooklyn, interests: food delivery",
+            "Homeowners 25-60, interests: home services",
+            "Adults 25-65, tax season, interests: finance",
+            "Pet owners 25-50, Chicago metro area",
+        ],
+        risk_signals=[],
+        label="legit",
+        severity=0.0,
+    ),
+    AdTemplate(
+        category="education",
+        ad_copies=[
+            "Learn Python programming in 12 weeks. Live instructor, project-based curriculum. CodeAcademy Pro.",
+            "MBA programs online — AACSB accredited. Flexible schedule for working professionals. StateU Online.",
+            "SAT prep that works. Average 150-point score improvement. Guaranteed. PrepMaster Tutoring.",
+            "Professional photography workshops. Small class sizes, hands-on practice. LensArt Studio.",
+            "Spanish fluency in 6 months. Native instructors, conversation-first approach. LinguaViva.",
+        ],
+        targeting_hints=[
+            "Adults 18-35, interests: programming, tech careers",
+            "Professionals 25-45, interests: MBA, business education",
+            "Parents of teens, interests: college prep, SAT",
+            "Adults 25-55, interests: photography, creative arts",
+            "Adults 18-50, interests: language learning, travel",
+        ],
+        risk_signals=[],
+        label="legit",
+        severity=0.0,
+    ),
+    AdTemplate(
+        category="fitness",
+        ad_copies=[
+            "Join our community gym. No contracts, cancel anytime. First month $19.99. FitLife Fitness.",
+            "Yoga for beginners — online classes with certified instructors. 7-day free trial. ZenFlow Yoga.",
+            "Custom meal plans designed by registered dietitians. Achieve your health goals. NutriPlan.",
+            "Running shoes engineered for comfort. 90-day return policy. SwiftStride Athletics.",
+            "Personal training at home. Certified trainers, flexible scheduling. FitVisit.",
+        ],
+        targeting_hints=[
+            "Adults 18-55, interests: fitness, gym",
+            "Adults 25-50, interests: yoga, wellness, mindfulness",
+            "Health-conscious adults 22-55, interests: nutrition, diet",
+            "Runners 18-55, interests: running, athletics",
+            "Adults 25-50, interests: personal training, fitness",
+        ],
+        risk_signals=[],
+        label="legit",
+        severity=0.0,
+    ),
+]
+
+FRAUD_TEMPLATES: List[AdTemplate] = [
+    # === OBVIOUS SCAMS (Task 1) ===
+    AdTemplate(
+        category="fake_giveaway",
+        ad_copies=[
+            "CONGRATULATIONS!! You've been selected to win a FREE iPhone 16 Pro!! Click NOW before offer expires!!!",
+            "🎉 WINNER ALERT 🎉 Claim your FREE $1000 Amazon Gift Card! Limited time only! Act NOW!",
+            "You are the 1,000,000th visitor! Click here to claim your FREE MacBook Pro! Don't miss out!!!",
+            "FREE PS5 GIVEAWAY! Just complete a quick survey! Only 3 left! Hurry!!!",
+            "EXCLUSIVE: Win a Tesla Model 3! Enter your email NOW! Drawing ends tonight!!!",
+        ],
+        targeting_hints=[
+            "All adults 18+, broad targeting, worldwide",
+            "Adults 18-65, interests: shopping, deals, coupons",
+            "Adults 18-35, interests: technology, gaming",
+            "Young adults 16-30, interests: gaming, PlayStation",
+            "Adults 18-55, interests: cars, Tesla, electric vehicles",
+        ],
+        risk_signals=[
+            "excessive urgency language",
+            "too-good-to-be-true offer",
+            "excessive punctuation/caps",
+            "no purchase necessary claims",
+        ],
+        label="fraud",
+        fraud_type="fake_giveaway",
+        severity=0.9,
+        difficulty="easy",
+    ),
+    AdTemplate(
+        category="counterfeit_goods",
+        ad_copies=[
+            "Designer handbags 90% OFF! Gucci, Louis Vuitton, Prada — all authentic! Limited stock! BargainLux.shop",
+            "Rolex watches starting at $49.99! 100% genuine! Free worldwide shipping! LuxWatchDeals.net",
+            "Nike Air Jordan CLEARANCE — $29.99! All sizes available! Original quality guaranteed! SneakerVault.store",
+            "Ray-Ban sunglasses factory direct! 95% discount! Premium quality! SunDealz.com",
+            "Canada Goose jackets $79.99! Authentic merchandise, warehouse clearance! ArcticDeals.shop",
+        ],
+        targeting_hints=[
+            "Adults 18-45, interests: luxury brands, fashion deals",
+            "Men 25-55, interests: watches, luxury",
+            "Young adults 16-30, interests: sneakers, streetwear",
+            "Adults 18-50, interests: fashion, sunglasses, deals",
+            "Adults 20-45, interests: outdoor gear, fashion",
+        ],
+        risk_signals=[
+            "impossibly low prices for luxury goods",
+            "unfamiliar domain",
+            "claims of authenticity with extreme discounts",
+        ],
+        label="fraud",
+        fraud_type="counterfeit",
+        severity=0.8,
+        difficulty="easy",
+    ),
+    AdTemplate(
+        category="miracle_cure",
+        ad_copies=[
+            "DOCTORS HATE THIS! One simple trick cures diabetes in 7 days! Big Pharma doesn't want you to know!",
+            "Lose 30 pounds in 2 weeks with this ancient herb! No diet, no exercise! MiracleSlim pills!",
+            "CURE CANCER NATURALLY! This secret remedy has a 99% success rate! Order now before it's banned!",
+            "Reverse aging in 14 days! Scientists discover fountain of youth pill! Limited supply!",
+            "Grow 3 inches taller at ANY age! Patented growth formula! 100% guaranteed results!",
+        ],
+        targeting_hints=[
+            "Adults 35-70, interests: health, diabetes, natural remedies",
+            "Adults 25-60, interests: weight loss, diet, fitness",
+            "Adults 40-75, interests: cancer, health, alternative medicine",
+            "Adults 35-65, interests: anti-aging, beauty, health",
+            "Men 18-40, interests: fitness, height, self-improvement",
+        ],
+        risk_signals=[
+            "unsubstantiated medical claims",
+            "conspiracy language ('they don't want you to know')",
+            "impossible health promises",
+            "urgency/scarcity tactics",
+        ],
+        label="fraud",
+        fraud_type="miracle_cure",
+        severity=0.95,
+        difficulty="easy",
+    ),
+    AdTemplate(
+        category="advance_fee",
+        ad_copies=[
+            "URGENT: I am Prince Adebayo of Nigeria. I need your help to transfer $4.5 million. You keep 30%.",
+            "Congratulations! You've inherited $2.8 million from a distant relative. Pay $200 processing fee to claim.",
+            "WORK FROM HOME! Earn $5000/week stuffing envelopes! No experience needed! Send $50 registration fee.",
+            "URGENT BUSINESS PROPOSAL: Oil company executive needs foreign partner. $10M deal. Reply immediately.",
+            "You've won the UK National Lottery! Send $100 for tax processing to receive your $1.5M prize.",
+        ],
+        targeting_hints=[
+            "All adults 18+, broad targeting",
+            "Adults 30-70, interests: finance, inheritance",
+            "Adults 18-65, interests: work from home, jobs",
+            "Adults 30-65, interests: business, investment",
+            "Adults 25-70, interests: lottery, gambling, luck",
+        ],
+        risk_signals=[
+            "advance fee request",
+            "unsolicited money offer",
+            "urgency pressure",
+            "request for personal information",
+        ],
+        label="fraud",
+        fraud_type="advance_fee_scam",
+        severity=0.85,
+        difficulty="easy",
+    ),
+
+    # === SOPHISTICATED SCAMS (Task 2) ===
+    AdTemplate(
+        category="fake_crypto",
+        ad_copies=[
+            "Invest in DeFi with confidence. Our AI-powered trading platform delivers 12-15% monthly returns. Audited smart contracts. Join CryptoYield Pro.",
+            "Stake your crypto with institutional-grade security. 8% APY on BTC, 12% on ETH. Regulated in Singapore. VaultStake Finance.",
+            "New token launch: GreenChain ($GRNC) — carbon credit blockchain. Pre-sale live. Backed by major VCs. Read our whitepaper.",
+            "Automated crypto arbitrage — our algorithm exploits price differences across exchanges. Average 20% monthly. QuantBot Trading.",
+            "Web3 real estate tokenization. Own fractions of premium properties. Starting at $100. Licensed broker-dealer. PropChain.io",
+        ],
+        targeting_hints=[
+            "Men 25-45, interests: cryptocurrency, DeFi, investing",
+            "Adults 25-55, interests: crypto staking, passive income",
+            "Tech-savvy adults 20-40, interests: blockchain, Web3, tokens",
+            "Adults 25-50, interests: algo trading, crypto, finance",
+            "Adults 30-55, interests: real estate, investing, crypto",
+        ],
+        risk_signals=[
+            "high guaranteed returns",
+            "complex financial product",
+        ],
+        label="fraud",
+        fraud_type="fake_crypto",
+        severity=0.7,
+        difficulty="medium",
+    ),
+    AdTemplate(
+        category="celebrity_endorsement_fraud",
+        ad_copies=[
+            "Elon Musk reveals his #1 investment for 2026. 'This will be bigger than Tesla.' Exclusive interview inside.",
+            "Warren Buffett's secret stock pick leaked! Financial advisors are scrambling. See the details before it's removed.",
+            "Oprah's favorite weight loss secret finally revealed. 'I lost 40 pounds without dieting.' Doctor-approved method.",
+            "Mark Cuban: 'I'm putting everything into this AI company.' Find out which stock billionaires are buying.",
+            "Taylor Swift partners with new wellness brand. 'It changed my life.' Limited celebrity edition available.",
+        ],
+        targeting_hints=[
+            "Adults 25-55, interests: investing, Elon Musk, technology",
+            "Adults 35-70, interests: stocks, Warren Buffett, investing",
+            "Women 25-60, interests: weight loss, Oprah, wellness",
+            "Adults 25-50, interests: AI, investing, Mark Cuban",
+            "Adults 18-45, interests: Taylor Swift, wellness, beauty",
+        ],
+        risk_signals=[
+            "celebrity name used without verification",
+            "urgency/exclusivity language",
+        ],
+        label="fraud",
+        fraud_type="fake_endorsement",
+        severity=0.65,
+        difficulty="medium",
+    ),
+    AdTemplate(
+        category="clone_brand",
+        ad_copies=[
+            "Amaz0n Warehouse Clearance — Electronics up to 70% off. Official overstock sale. Limited time. amaz0n-deals.com",
+            "PayPaI Security Alert: Your account needs verification. Update your information to avoid suspension. Secure portal.",
+            "Walmrt+ Members: Exclusive flash sale this weekend only. Save an extra 25% on groceries. Members-only link.",
+            "App1e Trade-In Event: Get up to $800 for your old device. Upgrade to iPhone 16. Authorized reseller.",
+            "Netfliix Premium — Get 1 year for $29.99! Special promotional price. Activate through our partner link.",
+        ],
+        targeting_hints=[
+            "Adults 18-65, interests: Amazon, deals, electronics",
+            "Adults 18-65, interests: online payments, PayPal",
+            "Adults 25-60, interests: Walmart, grocery shopping",
+            "Adults 18-55, interests: Apple, iPhone, technology",
+            "Adults 18-50, interests: Netflix, streaming, entertainment",
+        ],
+        risk_signals=[
+            "brand name misspelling",
+            "unofficial domain",
+        ],
+        label="fraud",
+        fraud_type="brand_impersonation",
+        severity=0.75,
+        difficulty="medium",
+    ),
+    AdTemplate(
+        category="gray_area_supplements",
+        ad_copies=[
+            "NeuroPeak — clinically studied nootropic for enhanced focus and memory. 30-day money-back guarantee. FDA-registered facility.",
+            "TestoMax Elite — natural testosterone support for men over 40. Doctor-formulated. Published clinical trial results.",
+            "SlimBiome Pro — proprietary probiotic blend for healthy weight management. Featured in Health Magazine. GMP certified.",
+            "Joint Restore Plus — advanced collagen peptide formula. Thousands of 5-star reviews. Orthopedist recommended.",
+            "SleepDeep Ultra — melatonin-free natural sleep aid. Non-habit forming. Money-back guarantee. PharmGrade Labs.",
+        ],
+        targeting_hints=[
+            "Adults 25-55, interests: nootropics, productivity, brain health",
+            "Men 40-65, interests: men's health, testosterone, fitness",
+            "Adults 30-60, interests: weight loss, probiotics, gut health",
+            "Adults 45-75, interests: joint health, supplements, arthritis",
+            "Adults 30-65, interests: sleep, insomnia, wellness",
+        ],
+        risk_signals=[
+            "supplement with strong efficacy claims",
+            "medical-sounding language",
+        ],
+        label="escalate",
+        fraud_type="gray_area",
+        severity=0.4,
+        difficulty="medium",
+    ),
+
+    # === NETWORK SCAMS (Task 3 — individual ads that belong to fraud rings) ===
+    # These templates look almost entirely legitimate on the surface.
+    # The fraud is only detectable through cross-ad investigation.
+    AdTemplate(
+        category="network_crypto",
+        ad_copies=[
+            "Earn passive income with our DeFi staking pool. Community-driven, transparent governance. BlockYield.io",
+            "Next-gen DEX with lowest fees. Trade with zero slippage on major pairs. SwapNova Protocol.",
+            "NFT marketplace for digital artists. Mint, list, and sell with 2% fees. Creator-first platform. ArtBlock.exchange",
+            "Cross-chain bridge with military-grade security. Move assets between chains in seconds. ChainLink Pro.",
+            "Crypto tax software — auto-import from 50+ exchanges. CPA-approved reports. TaxChain Solutions.",
+        ],
+        targeting_hints=[
+            "Adults 20-40, interests: DeFi, crypto staking",
+            "Adults 20-45, interests: crypto trading, DEX",
+            "Adults 20-40, interests: NFT, digital art, crypto",
+            "Adults 25-45, interests: cross-chain, crypto bridges",
+            "Adults 25-55, interests: crypto, tax, finance",
+        ],
+        risk_signals=[],
+        label="fraud",
+        fraud_type="coordinated_network",
+        severity=0.6,
+        difficulty="hard",
+    ),
+    AdTemplate(
+        category="network_ecommerce",
+        ad_copies=[
+            "Premium wireless earbuds — ANC, 40hr battery. Compare to AirPods Pro at half the price. SoundElite Store.",
+            "Smart home security camera — 2K resolution, night vision, cloud storage. TechGuard Home.",
+            "Portable power station for camping. 1000W, solar-ready. Adventure-proof design. PowerTrail Gear.",
+            "Ergonomic office chair — lumbar support, breathable mesh. 5-year warranty. WorkComfort Direct.",
+            "Robot vacuum with LiDAR mapping. Self-emptying dock. Smart home compatible. CleanBot Pro.",
+        ],
+        targeting_hints=[
+            "Adults 18-45, interests: tech, audio, headphones",
+            "Adults 25-55, interests: home security, smart home",
+            "Adults 25-50, interests: camping, outdoor gear",
+            "Adults 25-55, interests: home office, ergonomics",
+            "Adults 25-55, interests: smart home, cleaning",
+        ],
+        risk_signals=[],
+        label="fraud",
+        fraud_type="coordinated_network",
+        severity=0.5,
+        difficulty="hard",
+    ),
+    AdTemplate(
+        category="network_fintech",
+        ad_copies=[
+            "Simplify your bookkeeping. AI-powered categorization, bank sync, tax-ready reports. LedgerFlow.",
+            "Expense management for growing teams. Real-time tracking, receipt scanning, approval workflows. SpendWise.",
+            "Invoicing made simple. Send professional invoices in 30 seconds. Accept 40+ payment methods. BillStream.",
+            "Payroll for startups — automated tax filings, direct deposit, compliance included. PayRun HR.",
+            "Business credit monitoring. Real-time alerts, score tracking, credit-building tips. BizCredit Insights.",
+        ],
+        targeting_hints=[
+            "Small business owners 28-55, interests: accounting, bookkeeping",
+            "Finance managers 30-50, interests: expense management, SaaS",
+            "Freelancers 22-45, interests: invoicing, self-employment",
+            "Startup founders 25-45, interests: payroll, HR tech",
+            "Entrepreneurs 30-55, interests: business credit, finance",
+        ],
+        risk_signals=[],
+        label="fraud",
+        fraud_type="coordinated_network",
+        severity=0.55,
+        difficulty="hard",
+    ),
+    AdTemplate(
+        category="network_health",
+        ad_copies=[
+            "Telehealth appointments in minutes. Board-certified physicians, prescriptions delivered. MediConnect Health.",
+            "Mental wellness platform. Licensed therapists, flexible scheduling, insurance accepted. MindBridge Care.",
+            "At-home lab testing. Comprehensive panels, fast results, doctor-reviewed. HealthCheck Direct.",
+            "Nutrition coaching backed by science. Personalized meal plans from registered dietitians. NutriGuide Pro.",
+            "Fitness tracking reimagined. AI-powered workout plans, progress analytics, community support. FitMetrics.",
+        ],
+        targeting_hints=[
+            "Adults 25-60, interests: telemedicine, healthcare",
+            "Adults 25-50, interests: mental health, therapy, wellness",
+            "Health-conscious adults 30-55, interests: lab testing, preventive health",
+            "Adults 25-50, interests: nutrition, diet planning, health",
+            "Fitness enthusiasts 20-45, interests: fitness tracking, wearables",
+        ],
+        risk_signals=[],
+        label="fraud",
+        fraud_type="coordinated_network",
+        severity=0.5,
+        difficulty="hard",
+    ),
+]

data/landing_pages.py

@@ -0,0 +1,187 @@
+"""
+Simulated landing page investigation data.
+
+When the agent investigates 'landing_page' for an ad, this module
+generates the findings text based on the ad's ground truth label.
+"""
+
+from __future__ import annotations
+
+import random
+from dataclasses import dataclass
+from typing import List
+
+
+@dataclass
+class LandingPageData:
+    domain: str
+    domain_age_days: int
+    ssl_valid: bool
+    registrar: str
+    content_summary: str
+    redirect_chain: List[str]
+    similarity_to_known_scams: float  # 0.0-1.0
+    key_claims: List[str]
+    suspicious_elements: List[str]
+
+    def to_investigation_text(self) -> str:
+        ssl_status = "Valid SSL certificate" if self.ssl_valid else "NO SSL / expired certificate"
+        redirects = " -> ".join(self.redirect_chain) if self.redirect_chain else "No redirects (direct)"
+        suspicious = "\n".join(f"  - {s}" for s in self.suspicious_elements) if self.suspicious_elements else "  None detected"
+        claims = "\n".join(f"  - {c}" for c in self.key_claims) if self.key_claims else "  None"
+
+        lines = [
+            f"Domain: {self.domain}",
+            f"Domain age: {self.domain_age_days} days",
+            f"SSL: {ssl_status}",
+            f"Registrar: {self.registrar}",
+            f"WHOIS privacy: {'Enabled' if 'privacy' in self.registrar.lower() or 'proxy' in self.registrar.lower() else 'Disabled (registrant info public)'}",
+            f"Redirect chain: {redirects}",
+            f"Similarity to known scam templates: {self.similarity_to_known_scams:.0%}",
+            f"Key claims on page:\n{claims}",
+            f"Suspicious elements:\n{suspicious}",
+            f"Content summary: {self.content_summary}",
+        ]
+        return "\n".join(lines)
+
+
+_LEGIT_REGISTRARS = ["GoDaddy", "Namecheap", "Google Domains", "Cloudflare Registrar", "AWS Route 53"]
+_SUSPICIOUS_REGISTRARS = ["NameSilo", "Epik", "Njalla (privacy)", "Tucows (privacy proxy)", "Openprovider"]
+
+_LEGIT_DOMAINS_SUFFIXES = [".com", ".co", ".org", ".io", ".net"]
+_SCAM_DOMAINS_SUFFIXES = [".shop", ".store", ".xyz", ".click", ".top", ".buzz", ".site"]
+
+
+def generate_landing_page(
+    rng: random.Random,
+    ad_id: str,
+    is_fraud: bool,
+    fraud_type: str = "",
+    *,
+    domain_override: str | None = None,
+    registrar_override: str | None = None,
+) -> LandingPageData:
+    """Generate simulated landing page investigation data."""
+
+    base_word = rng.choice(["deal", "offer", "shop", "store", "buy", "get", "best", "top", "pro", "elite"])
+
+    if is_fraud:
+        suffix = rng.choice(_SCAM_DOMAINS_SUFFIXES)
+        domain = domain_override or f"{base_word}{rng.randint(10, 9999)}{suffix}"
+        domain_age = rng.randint(1, 60)
+        ssl_valid = rng.random() < 0.4
+        registrar = registrar_override or rng.choice(_SUSPICIOUS_REGISTRARS)
+        similarity = round(rng.uniform(0.5, 0.95), 2)
+
+        redirect_count = rng.randint(1, 4)
+        redirect_chain = [domain]
+        for _ in range(redirect_count):
+            redirect_chain.append(f"redirect{rng.randint(1, 999)}{rng.choice(['.click', '.top', '.xyz'])}")
+        redirect_chain.append(f"final-{rng.randint(1, 999)}{rng.choice(['.site', '.store'])}")
+
+        suspicious = _get_fraud_suspicious_elements(rng, fraud_type)
+        claims = _get_fraud_claims(rng, fraud_type)
+        content = _get_fraud_content_summary(rng, fraud_type)
+    else:
+        suffix = rng.choice(_LEGIT_DOMAINS_SUFFIXES)
+        brand = rng.choice(["homenest", "stylehaven", "taskflow", "fitlife", "sparkle", "codeacademy", "nutri"])
+        domain = domain_override or f"{brand}{suffix}"
+        domain_age = rng.randint(365, 5000)
+        ssl_valid = True
+        registrar = registrar_override or rng.choice(_LEGIT_REGISTRARS)
+        similarity = round(rng.uniform(0.0, 0.15), 2)
+        redirect_chain = []
+        suspicious = []
+        claims = _get_legit_claims(rng)
+        content = _get_legit_content_summary(rng)
+
+    return LandingPageData(
+        domain=domain,
+        domain_age_days=domain_age,
+        ssl_valid=ssl_valid,
+        registrar=registrar,
+        content_summary=content,
+        redirect_chain=redirect_chain,
+        similarity_to_known_scams=similarity,
+        key_claims=claims,
+        suspicious_elements=suspicious,
+    )
+
+
+def _get_fraud_suspicious_elements(rng: random.Random, fraud_type: str) -> List[str]:
+    common = [
+        "No physical address listed",
+        "No contact phone number",
+        "Privacy policy copied from template",
+        "Terms of service link returns 404",
+    ]
+    type_specific = {
+        "fake_giveaway": ["Countdown timer with fake urgency", "Requests personal info before any value provided"],
+        "counterfeit": ["Product images appear stolen from official brand site", "Price comparison shows 90%+ discount vs retail"],
+        "miracle_cure": ["FDA disclaimer buried in footer", "Before/after photos appear digitally altered"],
+        "advance_fee_scam": ["Wire transfer requested upfront", "No company registration found"],
+        "fake_crypto": ["Smart contract not verified on-chain", "Team photos are stock images"],
+        "fake_endorsement": ["Celebrity quote not found in any verified source", "Affiliate tracking parameters in URL"],
+        "brand_impersonation": ["Domain mimics well-known brand with character substitution", "Logo appears edited from official assets"],
+        "gray_area": ["Citations reference unpublished or retracted studies", "Testimonials lack verifiable details"],
+        "coordinated_network": ["Identical page template used across multiple domains", "Contact form submits to third-party aggregator"],
+    }
+    elements = list(common)
+    elements.extend(type_specific.get(fraud_type, []))
+    rng.shuffle(elements)
+    return elements[: rng.randint(2, min(5, len(elements)))]
+
+
+def _get_fraud_claims(rng: random.Random, fraud_type: str) -> List[str]:
+    by_type = {
+        "fake_giveaway": ["100% free, no purchase necessary", "Guaranteed winner", "Act now — expires in 2 hours"],
+        "counterfeit": ["Authentic products", "Direct from manufacturer", "90% below retail price"],
+        "miracle_cure": ["Clinically proven results", "Works in 7 days or less", "Endorsed by doctors"],
+        "advance_fee_scam": ["Guaranteed returns", "Low processing fee required", "Confidential opportunity"],
+        "fake_crypto": ["12-15% monthly returns guaranteed", "Audited by top security firms", "Risk-free investment"],
+        "fake_endorsement": ["As seen on TV", "Celebrity recommended", "Limited exclusive offer"],
+        "brand_impersonation": ["Official authorized retailer", "Factory direct pricing", "Same quality, lower price"],
+        "gray_area": ["Clinically studied ingredients", "Doctor recommended", "30-day money-back guarantee"],
+        "coordinated_network": ["Limited stock available", "Thousands of 5-star reviews", "Fast free shipping"],
+    }
+    return by_type.get(fraud_type, ["Special limited offer", "Act now"])
+
+
+def _get_fraud_content_summary(rng: random.Random, fraud_type: str) -> str:
+    summaries = {
+        "fake_giveaway": "Landing page features a large countdown timer and a form requesting name, email, phone, and address. No clear sponsor or rules disclosure.",
+        "counterfeit": "Product catalog showing luxury branded items at extreme discounts. Stock photos. Payment accepted via wire transfer and crypto only.",
+        "miracle_cure": "Long-form sales page with testimonials, before/after images, and pseudoscientific explanations. Multiple urgency CTAs.",
+        "advance_fee_scam": "Simple page with a letter-style appeal requesting wire transfer. Grammar errors. No company details.",
+        "fake_crypto": "Professional-looking platform with dashboard screenshots. Whitepaper link leads to a generic PDF. Team bios use stock photos.",
+        "fake_endorsement": "News article-style page with celebrity photos. Comments section appears pre-populated with positive feedback.",
+        "brand_impersonation": "Near-replica of the official brand website. URL uses character substitution. Cart and checkout flow functional but data destination unclear.",
+        "gray_area": "Well-designed supplement product page. Ingredient list present but efficacy claims exceed scientific evidence. Has a real return policy.",
+        "coordinated_network": "Standard e-commerce template. Products appear legitimate but company details are vague. Identical template to other flagged sites.",
+    }
+    return summaries.get(fraud_type, "Generic landing page with minimal content and heavy use of urgency language.")
+
+
+def _get_legit_claims(rng: random.Random) -> List[str]:
+    options = [
+        "Free shipping on orders over $50",
+        "14-day free trial",
+        "Money-back guarantee",
+        "Established since 2010",
+        "4.5-star average customer rating",
+        "Licensed and insured",
+        "BBB accredited",
+        "Certified organic ingredients",
+    ]
+    rng.shuffle(options)
+    return options[: rng.randint(2, 4)]
+
+
+def _get_legit_content_summary(rng: random.Random) -> str:
+    options = [
+        "Well-structured business website with clear product catalog, about page, contact information, and shipping policy.",
+        "Professional service website with team bios, case studies, client testimonials, and clear pricing.",
+        "E-commerce store with detailed product descriptions, customer reviews, FAQ section, and accessible support.",
+        "SaaS landing page with feature breakdown, pricing tiers, integration docs, and live demo option.",
+    ]
+    return rng.choice(options)

data/network_generator.py

@@ -0,0 +1,194 @@
+"""
+Fraud network (ring) generation for Task 3 using networkx.
+
+Generates complex fraud ring topologies:
+- Simple cliques (3-5 ads fully connected)
+- Chain structures (A-B via payment, B-C via creative — transitive reasoning required)
+- Hub-and-spoke (one master account linked to satellites)
+
+Individual ads in a ring may look borderline; the signal is in the connections.
+Each edge in the graph carries the signal type that connects the two ads.
+"""
+
+from __future__ import annotations
+
+import random
+from dataclasses import dataclass, field
+from typing import Dict, List, Set, Tuple
+
+import networkx as nx
+
+
+@dataclass
+class FraudRing:
+    ring_id: str
+    member_ad_ids: List[str]
+    shared_signals: Dict[str, str]  # signal_type -> shared_value
+    topology: str = "clique"  # clique, chain, hub_spoke
+
+    @property
+    def size(self) -> int:
+        return len(self.member_ad_ids)
+
+
+_RING_TOPOLOGIES = ["clique", "chain", "hub_spoke"]
+
+_SIGNAL_POOL_KEYS = ["payment_method", "domain_registrar", "creative_template", "targeting_overlap"]
+
+_REGISTRAR_CHOICES = ["Njalla (privacy)", "Epik", "NameSilo", "Tucows (privacy proxy)"]
+_TARGETING_CHOICES = [
+    "Men 25-45, crypto+investing, US+UK+AU",
+    "Adults 18-35, tech+gaming, worldwide",
+    "Women 25-55, health+beauty, US+CA",
+    "Adults 30-60, finance+real-estate, US+UK",
+    "Adults 20-40, e-commerce+dropshipping, US+EU",
+]
+
+
+def _make_signal_pool(rng: random.Random, ring_index: int) -> Dict[str, str]:
+    """Generate a pool of shared signal values for one ring."""
+    return {
+        "payment_method": f"pmt_ring_{rng.randint(10000, 99999)}",
+        "domain_registrar": rng.choice(_REGISTRAR_CHOICES),
+        "creative_template": f"tmpl_{rng.randint(1000, 9999)}",
+        "targeting_overlap": rng.choice(_TARGETING_CHOICES),
+    }
+
+
+def generate_fraud_networks(
+    rng: random.Random,
+    n_rings: int,
+    available_fraud_ad_ids: List[str],
+) -> Tuple[List[FraudRing], Dict[str, List[str]]]:
+    """
+    Generate fraud ring structures with complex topologies.
+
+    Returns:
+        rings: list of FraudRing objects
+        ad_to_rings: mapping from ad_id to list of ring_ids it belongs to
+    """
+    G = nx.Graph()
+    rings: List[FraudRing] = []
+    ad_to_rings: Dict[str, List[str]] = {}
+
+    remaining = list(available_fraud_ad_ids)
+    rng.shuffle(remaining)
+
+    for i in range(n_rings):
+        if len(remaining) < 3:
+            break
+        ring_size = rng.randint(3, min(5, len(remaining)))
+
+        members = remaining[:ring_size]
+        remaining = remaining[ring_size:]
+
+        topology = rng.choice(_RING_TOPOLOGIES)
+        signal_pool = _make_signal_pool(rng, i)
+
+        signal_keys = list(_SIGNAL_POOL_KEYS)
+        rng.shuffle(signal_keys)
+        n_shared = rng.randint(2, len(signal_keys))
+        shared_signals = {k: signal_pool[k] for k in signal_keys[:n_shared]}
+
+        _add_edges_for_topology(G, members, shared_signals, topology, rng)
+
+        ring_id = f"ring_{i}"
+        ring = FraudRing(
+            ring_id=ring_id,
+            member_ad_ids=members,
+            shared_signals=shared_signals,
+            topology=topology,
+        )
+        rings.append(ring)
+
+        for ad_id in members:
+            ad_to_rings.setdefault(ad_id, []).append(ring_id)
+            G.add_node(ad_id, ring_id=ring_id)
+
+    # Optionally create bridge nodes between rings for extra complexity
+    if len(rings) >= 2 and remaining:
+        _add_bridge_ads(G, rings, remaining, ad_to_rings, rng)
+
+    return rings, ad_to_rings
+
+
+def _add_edges_for_topology(
+    G: nx.Graph,
+    members: List[str],
+    shared_signals: Dict[str, str],
+    topology: str,
+    rng: random.Random,
+) -> None:
+    """Add edges to the graph based on the ring topology."""
+    signal_types = list(shared_signals.keys())
+
+    if topology == "clique":
+        for i, a in enumerate(members):
+            for b in members[i + 1:]:
+                signal = rng.choice(signal_types)
+                G.add_edge(a, b, signal_type=signal, signal_value=shared_signals[signal])
+
+    elif topology == "chain":
+        for idx in range(len(members) - 1):
+            signal = signal_types[idx % len(signal_types)]
+            G.add_edge(
+                members[idx], members[idx + 1],
+                signal_type=signal, signal_value=shared_signals[signal],
+            )
+
+    elif topology == "hub_spoke":
+        hub = members[0]
+        for spoke in members[1:]:
+            signal = rng.choice(signal_types)
+            G.add_edge(hub, spoke, signal_type=signal, signal_value=shared_signals[signal])
+
+
+def _add_bridge_ads(
+    G: nx.Graph,
+    rings: List[FraudRing],
+    remaining: List[str],
+    ad_to_rings: Dict[str, List[str]],
+    rng: random.Random,
+) -> None:
+    """Optionally link two rings via a shared bridge ad from the remaining pool."""
+    if len(remaining) < 1 or len(rings) < 2:
+        return
+
+    bridge_ad = remaining.pop(0)
+    r1, r2 = rings[0], rings[1]
+
+    bridge_to_r1 = rng.choice(r1.member_ad_ids)
+    bridge_to_r2 = rng.choice(r2.member_ad_ids)
+
+    r1.member_ad_ids.append(bridge_ad)
+    ad_to_rings.setdefault(bridge_ad, []).extend([r1.ring_id, r2.ring_id])
+
+    sig_key = rng.choice(list(r1.shared_signals.keys()))
+    G.add_edge(bridge_ad, bridge_to_r1, signal_type=sig_key, signal_value=r1.shared_signals[sig_key])
+
+    sig_key2 = rng.choice(list(r2.shared_signals.keys()))
+    G.add_edge(bridge_ad, bridge_to_r2, signal_type=sig_key2, signal_value=r2.shared_signals[sig_key2])
+
+
+def get_ring_shared_signal_text(ring: FraudRing) -> str:
+    """Describe the shared signals in a ring (for grader/debug use)."""
+    lines = [f"Fraud Ring {ring.ring_id} ({ring.size} members, topology={ring.topology}):"]
+    lines.append(f"  Members: {', '.join(ring.member_ad_ids)}")
+    lines.append("  Shared signals:")
+    for signal_type, value in ring.shared_signals.items():
+        lines.append(f"    - {signal_type}: {value}")
+    return "\n".join(lines)
+
+
+def build_ground_truth_graph(rings: List[FraudRing]) -> nx.Graph:
+    """Reconstruct the full ground truth network graph from rings.
+
+    Used by graders to compute the expected set of edges.
+    """
+    G = nx.Graph()
+    for ring in rings:
+        for i, a in enumerate(ring.member_ad_ids):
+            G.add_node(a, ring_id=ring.ring_id)
+            for b in ring.member_ad_ids[i + 1:]:
+                G.add_edge(a, b, ring_id=ring.ring_id)
+    return G

graders/__init__.py

@@ -0,0 +1,14 @@
+"""Graders for scoring agent performance on each task."""
+
+from .base_grader import grade_episode
+from .task1_grader import Task1Grader
+from .task2_grader import Task2Grader
+from .task3_grader import Task3Grader
+
+GRADERS = {
+    "task_1": Task1Grader(),
+    "task_2": Task2Grader(),
+    "task_3": Task3Grader(),
+}
+
+__all__ = ["grade_episode", "Task1Grader", "Task2Grader", "Task3Grader", "GRADERS"]

graders/base_grader.py

@@ -0,0 +1,113 @@
+"""
+Shared grading logic for all tasks.
+
+Each grader produces a 0.0-1.0 score by normalizing raw reward
+between theoretical worst and best cases.
+"""
+
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+
+_NORM_EPS = 1e-8  # For numerical stability
+
+
+@dataclass
+class VerdictResult:
+    ad_id: str
+    verdict: str          # approve, reject, escalate
+    confidence: float
+    ground_truth: str     # fraud, legit, escalate
+    auto_approved: bool = False
+
+
+@dataclass
+class LinkResult:
+    ad_id_1: str
+    ad_id_2: str
+    correct: bool
+
+
+@dataclass
+class EpisodeRecord:
+    """All data needed for grading a completed episode."""
+    task_id: str
+    total_steps: int
+    action_budget: int
+    verdicts: List[VerdictResult]
+    links: List[LinkResult]
+    ads_metadata: List[Dict[str, Any]]  # [{ad_id, ground_truth, severity, ...}]
+    n_fraud_rings: int = 0
+    ring_sizes: List[int] = None
+
+    def __post_init__(self):
+        if self.ring_sizes is None:
+            self.ring_sizes = []
+
+
+class BaseGrader(ABC):
+    """Abstract grader that scores an episode 0.0-1.0."""
+
+    @abstractmethod
+    def grade(self, record: EpisodeRecord) -> float:
+        """Return a score in [0.0, 1.0]."""
+        ...
+
+    def _count_correct_verdicts(self, verdicts: List[VerdictResult]) -> int:
+        return sum(1 for v in verdicts if self._is_correct(v))
+
+    def _count_false_positives(self, verdicts: List[VerdictResult]) -> int:
+        return sum(
+            1 for v in verdicts
+            if v.verdict == "reject" and v.ground_truth == "legit"
+        )
+
+    def _count_false_negatives(self, verdicts: List[VerdictResult]) -> int:
+        return sum(
+            1 for v in verdicts
+            if v.verdict == "approve" and v.ground_truth == "fraud"
+        )
+
+    def _is_correct(self, v: VerdictResult) -> bool:
+        return (
+            (v.verdict == "reject" and v.ground_truth == "fraud")
+            or (v.verdict == "approve" and v.ground_truth == "legit")
+            or (v.verdict == "escalate" and v.ground_truth == "escalate")
+        )
+
+    def _verdict_reward(self, v: VerdictResult, severity: float = 0.5) -> float:
+        if v.verdict == "reject" and v.ground_truth == "fraud":
+            return 0.3 + 0.1 * severity
+        elif v.verdict == "approve" and v.ground_truth == "legit":
+            return 0.1
+        elif v.verdict == "escalate" and v.ground_truth == "escalate":
+            return 0.15
+        elif v.verdict == "reject" and v.ground_truth == "legit":
+            return -0.35
+        elif v.verdict == "approve" and v.ground_truth == "fraud":
+            return -0.5
+        elif v.verdict == "escalate":
+            return -0.05
+        elif v.verdict == "approve" and v.ground_truth == "escalate":
+            return -0.15
+        elif v.verdict == "reject" and v.ground_truth == "escalate":
+            return -0.1
+        return -0.05
+
+    def _normalize(self, raw: float, best: float, worst: float) -> float:
+        score_range = best - worst
+        if score_range <= 0:
+            return 0.5
+        normalized = (raw - worst) / (score_range + _NORM_EPS)
+        return max(0.0, min(1.0, normalized))
+
+
+def grade_episode(record: EpisodeRecord) -> float:
+    """Grade an episode using the appropriate task grader."""
+    from . import GRADERS
+    grader = GRADERS.get(record.task_id)
+    if grader is None:
+        raise ValueError(f"Unknown task_id: {record.task_id}")
+    return grader.grade(record)

graders/task1_grader.py

@@ -0,0 +1,49 @@
+"""
+Task 1 Grader: Basic Ad Triage (Easy).
+
+Scores based on verdict accuracy only. No network or calibration bonuses.
+5 ads, 25 budget — a decent LLM should score 0.6-0.8.
+"""
+
+from __future__ import annotations
+
+from .base_grader import BaseGrader, EpisodeRecord
+
+
+class Task1Grader(BaseGrader):
+    def grade(self, record: EpisodeRecord) -> float:
+        raw = 0.0
+        best = 0.0
+        worst = 0.0
+
+        severity_map = {
+            m["ad_id"]: m.get("severity", 0.5) for m in record.ads_metadata
+        }
+
+        for v in record.verdicts:
+            severity = severity_map.get(v.ad_id, 0.5)
+            raw += self._verdict_reward(v, severity)
+
+            if v.ground_truth == "fraud":
+                best += 0.3 + 0.1 * severity
+                worst -= 0.5
+            elif v.ground_truth == "legit":
+                best += 0.1
+                worst -= 0.35
+            elif v.ground_truth == "escalate":
+                best += 0.15
+                worst -= 0.15
+
+        n_investigations = max(0, record.total_steps - len([
+            v for v in record.verdicts if not v.auto_approved
+        ]))
+        raw -= n_investigations * 0.02
+        worst -= record.action_budget * 0.02
+
+        best += 0.2
+
+        if record.total_steps > 0:
+            correct = self._count_correct_verdicts(record.verdicts)
+            raw += (correct / record.total_steps) * 0.2
+
+        return self._normalize(raw, best, worst)

graders/task2_grader.py

@@ -0,0 +1,89 @@
+"""
+Task 2 Grader: Sophisticated Fraud Under Budget Pressure (Medium).
+
+Adds budget efficiency bonus and calibration scoring on top of verdict accuracy.
+12 ads, 30 budget — requires triage, expected baseline 0.3-0.5.
+"""
+
+from __future__ import annotations
+
+from .base_grader import BaseGrader, EpisodeRecord
+
+
+class Task2Grader(BaseGrader):
+    def grade(self, record: EpisodeRecord) -> float:
+        raw = 0.0
+        best = 0.0
+        worst = 0.0
+
+        severity_map = {
+            m["ad_id"]: m.get("severity", 0.5) for m in record.ads_metadata
+        }
+
+        for v in record.verdicts:
+            severity = severity_map.get(v.ad_id, 0.5)
+            raw += self._verdict_reward(v, severity)
+
+            if v.ground_truth == "fraud":
+                best += 0.3 + 0.1 * severity
+                worst -= 0.5
+            elif v.ground_truth == "legit":
+                best += 0.1
+                worst -= 0.35
+            elif v.ground_truth == "escalate":
+                best += 0.15
+                worst -= 0.15
+
+        n_investigations = max(0, record.total_steps - len([
+            v for v in record.verdicts if not v.auto_approved
+        ]))
+        raw -= n_investigations * 0.02
+        worst -= record.action_budget * 0.02
+
+        # Budget efficiency bonus
+        best += 0.2
+        if record.total_steps > 0:
+            correct = self._count_correct_verdicts(record.verdicts)
+            raw += (correct / record.total_steps) * 0.2
+
+        # Calibration bonus: reward agents whose stated confidence
+        # correlates with actual accuracy
+        calibration = self._compute_calibration(record)
+        raw += calibration * 0.15
+        best += 0.15
+
+        return self._normalize(raw, best, worst)
+
+    def _compute_calibration(self, record: EpisodeRecord) -> float:
+        """
+        Measure how well confidence tracks correctness.
+
+        Bins verdicts by confidence and checks if the fraction correct
+        in each bin roughly matches the stated confidence.
+        """
+        manual = [v for v in record.verdicts if not v.auto_approved]
+        if len(manual) < 3:
+            return 0.5
+
+        bins = {"low": [], "mid": [], "high": []}
+        for v in manual:
+            if v.confidence < 0.4:
+                bins["low"].append(v)
+            elif v.confidence < 0.7:
+                bins["mid"].append(v)
+            else:
+                bins["high"].append(v)
+
+        errors = []
+        for _label, group in bins.items():
+            if not group:
+                continue
+            avg_conf = sum(v.confidence for v in group) / len(group)
+            avg_acc = sum(1 for v in group if self._is_correct(v)) / len(group)
+            errors.append(abs(avg_conf - avg_acc))
+
+        if not errors:
+            return 0.5
+
+        avg_error = sum(errors) / len(errors)
+        return max(0.0, 1.0 - 2.0 * avg_error)

graders/task3_grader.py

@@ -0,0 +1,133 @@
+"""
+Task 3 Grader: Coordinated Fraud Network Detection (Hard).
+
+Adds graph-based network detection scoring on top of verdict accuracy,
+budget efficiency, and calibration. 20 ads, 35 budget, 3 fraud rings
+with varied topologies. Expected baseline 0.1-0.3.
+
+Network scoring uses ground truth edge coverage: what fraction of
+ground-truth ring connections did the agent discover via link_accounts?
+"""
+
+from __future__ import annotations
+
+from .base_grader import BaseGrader, EpisodeRecord
+
+
+class Task3Grader(BaseGrader):
+    def grade(self, record: EpisodeRecord) -> float:
+        raw = 0.0
+        best = 0.0
+        worst = 0.0
+
+        severity_map = {
+            m["ad_id"]: m.get("severity", 0.5) for m in record.ads_metadata
+        }
+
+        for v in record.verdicts:
+            severity = severity_map.get(v.ad_id, 0.5)
+            raw += self._verdict_reward(v, severity)
+
+            if v.ground_truth == "fraud":
+                best += 0.3 + 0.1 * severity
+                worst -= 0.5
+            elif v.ground_truth == "legit":
+                best += 0.1
+                worst -= 0.35
+            elif v.ground_truth == "escalate":
+                best += 0.15
+                worst -= 0.15
+
+        n_investigations = max(0, record.total_steps - len([
+            v for v in record.verdicts if not v.auto_approved
+        ]))
+        raw -= n_investigations * 0.02
+        worst -= record.action_budget * 0.02
+
+        # Budget efficiency bonus
+        best += 0.2
+        if record.total_steps > 0:
+            correct = self._count_correct_verdicts(record.verdicts)
+            raw += (correct / record.total_steps) * 0.2
+
+        # Calibration bonus
+        calibration = self._compute_calibration(record)
+        raw += calibration * 0.15
+        best += 0.15
+
+        # Graph-based network detection scoring
+        network_reward = self._compute_network_score(record)
+        raw += network_reward
+        # Best case: discover all ground truth edges
+        total_gt_edges = self._count_ground_truth_edges(record)
+        best += max(total_gt_edges, 1) * 0.4
+        worst -= max(len(record.links), 1) * 0.2
+
+        # Investigation coverage bonus: reward breadth over depth
+        coverage = self._compute_coverage_bonus(record)
+        raw += coverage * 0.1
+        best += 0.1
+
+        return self._normalize(raw, best, worst)
+
+    def _compute_calibration(self, record: EpisodeRecord) -> float:
+        manual = [v for v in record.verdicts if not v.auto_approved]
+        if len(manual) < 3:
+            return 0.5
+
+        bins = {"low": [], "mid": [], "high": []}
+        for v in manual:
+            if v.confidence < 0.4:
+                bins["low"].append(v)
+            elif v.confidence < 0.7:
+                bins["mid"].append(v)
+            else:
+                bins["high"].append(v)
+
+        errors = []
+        for _label, group in bins.items():
+            if not group:
+                continue
+            avg_conf = sum(v.confidence for v in group) / len(group)
+            avg_acc = sum(1 for v in group if self._is_correct(v)) / len(group)
+            errors.append(abs(avg_conf - avg_acc))
+
+        if not errors:
+            return 0.5
+
+        avg_error = sum(errors) / len(errors)
+        return max(0.0, 1.0 - 2.0 * avg_error)
+
+    def _count_ground_truth_edges(self, record: EpisodeRecord) -> int:
+        """Count the total number of ground truth edges across all rings."""
+        total = 0
+        for ring_size in record.ring_sizes:
+            total += ring_size * (ring_size - 1) // 2
+        return total
+
+    def _compute_network_score(self, record: EpisodeRecord) -> float:
+        """Score link_accounts actions against ground truth fraud rings.
+
+        Uses edge-coverage: correct links earn +0.4 each, incorrect
+        links incur -0.25 each (heavier penalty since decoys exist).
+        """
+        correct_links = sum(1 for l in record.links if l.correct)
+        incorrect_links = sum(1 for l in record.links if not l.correct)
+
+        reward = correct_links * 0.4
+        reward -= incorrect_links * 0.25
+
+        return reward
+
+    def _compute_coverage_bonus(self, record: EpisodeRecord) -> float:
+        """Reward agents that investigate across multiple ads (breadth).
+
+        Agents that only deep-dive a single ad miss network signals.
+        """
+        if not record.ads_metadata:
+            return 0.0
+
+        total_ads = len(record.ads_metadata)
+        reviewed = sum(1 for v in record.verdicts if not v.auto_approved)
+        coverage_ratio = reviewed / total_ads if total_ads > 0 else 0.0
+        return min(1.0, coverage_ratio)

inference.py

@@ -0,0 +1,451 @@
+"""
+Inference Script — Ad Fraud Investigation Environment
+===================================
+MANDATORY
+- Before submitting, ensure the following variables are defined in your environment configuration:
+    API_BASE_URL   The API endpoint for the LLM.
+    MODEL_NAME     The model identifier to use for inference.
+    HF_TOKEN       Your Hugging Face / API key.
+- The inference script must be named `inference.py` and placed in the root directory of the project
+- Participants must use OpenAI Client for all LLM calls using above variables
+
+STDOUT FORMAT
+- The script must emit exactly three line types to stdout, in this order:
+
+    [START] task=<task_name> env=<benchmark> model=<model_name>
+    [STEP]  step=<n> action=<action_str> reward=<0.00> done=<true|false> error=<msg|null>
+    [END]   success=<true|false> steps=<n> score=<score> rewards=<r1,r2,...,rn>
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+import sys
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+from openai import OpenAI
+
+try:
+    from .client import AdFraudEnv
+    from .models import AdReviewAction
+    from .data.ad_generator import TASK_CONFIGS
+except ImportError:
+    sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+    from ad_fraud_env.client import AdFraudEnv
+    from ad_fraud_env.models import AdReviewAction
+    from ad_fraud_env.data.ad_generator import TASK_CONFIGS
+
+from dotenv import load_dotenv
+load_dotenv()
+
+API_BASE_URL = os.getenv("API_BASE_URL") or "https://router.huggingface.co/v1"
+API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY")
+MODEL_NAME = os.getenv("MODEL_NAME") or "meta-llama/Llama-3.1-8B-Instruct"
+ENV_URL = os.getenv("AD_FRAUD_ENV_URL", "https://QuantumTransformer-AdArena.hf.space")
+BENCHMARK = "ad_fraud_env"
+TEMPERATURE = 0.1
+MAX_TOKENS = 256
+FALLBACK_VERDICT = "escalate"
+
+logger = logging.getLogger(__name__)
+LOG_DIR = Path(__file__).resolve().parent / "convo_logging"
+
+# ---------------------------------------------------------------------------
+# Mandatory structured stdout logging
+# ---------------------------------------------------------------------------
+
+def log_start(task: str, env: str, model: str) -> None:
+    print(f"[START] task={task} env={env} model={model}", flush=True)
+
+
+def log_step(step: int, action: str, reward: float, done: bool, error: Optional[str]) -> None:
+    error_val = error if error else "null"
+    done_val = str(done).lower()
+    print(
+        f"[STEP] step={step} action={action} reward={reward:.2f} "
+        f"done={done_val} error={error_val}",
+        flush=True,
+    )
+
+
+def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None:
+    rewards_str = ",".join(f"{r:.2f}" for r in rewards)
+    print(
+        f"[END] success={str(success).lower()} steps={steps} "
+        f"score={score:.2f} rewards={rewards_str}",
+        flush=True,
+    )
+
+
+def _format_action(action: AdReviewAction) -> str:
+    """Compact action string for [STEP] log line."""
+    if action.action_type == "investigate":
+        return f"investigate({action.ad_id},{action.investigation_target})"
+    elif action.action_type == "verdict":
+        conf = action.confidence if action.confidence is not None else 0.5
+        return f"verdict({action.ad_id},{action.verdict},{conf:.2f})"
+    elif action.action_type == "link_accounts":
+        return f"link_accounts({action.ad_id},{action.linked_ad_id})"
+    return f"unknown({action.ad_id})"
+
+# ---------------------------------------------------------------------------
+# System prompt & helpers
+# ---------------------------------------------------------------------------
+
+SYSTEM_PROMPT = """\
+You are an ad fraud investigator reviewing a queue of advertisements.
+Your job is to investigate suspicious ads and render verdicts (approve, reject, or escalate).
+
+For each step, you must output a single JSON action. The action schema is:
+
+{
+  "action_type": "investigate" | "verdict" | "link_accounts",
+  "ad_id": "<ad ID, e.g. ad_001>",
+
+  // For investigate actions:
+  "investigation_target": "advertiser_history" | "landing_page" | "payment_method" | "targeting_overlap" | "creative_similarity" | "campaign_structure",
+
+  // For verdict actions:
+  "verdict": "approve" | "reject" | "escalate",
+  "confidence": <float 0.0-1.0>,
+
+  // For link_accounts actions:
+  "linked_ad_id": "<other ad ID>",
+  "link_reason": "<reason>"
+}
+
+Strategy:
+1. Start by reading the queue summary and the first ad's information.
+2. For obviously suspicious ads, investigate 1-2 signals then reject.
+3. For clearly legitimate ads, approve quickly with high confidence.
+4. For ambiguous ads, investigate more deeply before deciding.
+5. Manage your budget — you cannot investigate everything.
+6. For link_accounts, only flag connections when you see shared signals across ads (same payment method, similar creative template, targeting overlap).
+
+Output ONLY the JSON action, no other text.
+"""
+
+JSON_BLOCK_RE = re.compile(r"```(?:json)?\s*\n(.*?)```", re.DOTALL)
+_FINDING_BLOCK_RE = re.compile(r"\[(ad_\d+)\s*/\s*([a-z_]+)\]")
+_ANALYSIS_HEADER_RE = re.compile(
+    r"^(?:Payment Method|Targeting|Creative|Campaign Structure) Analysis for ad_\d+:$"
+)
+
+
+def _extract_json(text: str) -> Dict[str, Any]:
+    """Extract JSON from LLM response, handling markdown code blocks."""
+    text = text.strip()
+    m = JSON_BLOCK_RE.search(text)
+    if m:
+        text = m.group(1).strip()
+    elif text.startswith("```"):
+        lines = text.split("\n")
+        lines = [l for l in lines if not l.strip().startswith("```")]
+        text = "\n".join(lines).strip()
+    return json.loads(text)
+
+
+def _compact_finding(text: str) -> str:
+    """Compress one investigation finding block into a compact key-value line.
+
+    Strips analysis headers and section labels but preserves ALL data fields
+    (both signal and noise) so the agent must still reason about relevance.
+    """
+    parts: List[str] = []
+    for line in text.strip().split("\n"):
+        stripped = line.strip()
+        if not stripped:
+            continue
+        if _ANALYSIS_HEADER_RE.match(stripped):
+            continue
+        if stripped in ("Key claims on page:", "Suspicious elements:"):
+            continue
+        if stripped.startswith("- "):
+            stripped = stripped[2:]
+        parts.append(stripped)
+    return " | ".join(parts)
+
+
+def _build_compact_findings(raw: str, focused_ad: Optional[str]) -> str:
+    """Parse findings blocks; full prose for focused ad, compact lines for others.
+
+    Each block in the raw findings string is delimited by [ad_XXX / target].
+    The focused ad keeps full investigation prose for deep reasoning.
+    Other ads are compressed to single-line key-value summaries that still
+    contain all extracted fields (payment type, IDs, template hashes,
+    fingerprints, domain ages, etc.) mixed with noise fields — the agent
+    must determine which values are meaningful for cross-ad comparison.
+    """
+    blocks: List[tuple] = []
+    current_ad: Optional[str] = None
+    current_header: Optional[str] = None
+    current_lines: List[str] = []
+
+    for line in raw.split("\n"):
+        m = _FINDING_BLOCK_RE.match(line.strip())
+        if m:
+            if current_ad is not None:
+                blocks.append((current_ad, current_header, "\n".join(current_lines)))
+            current_ad = m.group(1)
+            current_header = line.strip()
+            current_lines = []
+        else:
+            current_lines.append(line)
+
+    if current_ad is not None:
+        blocks.append((current_ad, current_header, "\n".join(current_lines)))
+
+    result: List[str] = []
+    for ad_id, header, text in blocks:
+        if focused_ad and ad_id == focused_ad:
+            result.append(f"\n{header}\n{text}")
+        else:
+            compact = _compact_finding(text)
+            if compact:
+                result.append(f"{header} {compact}")
+
+    return "\n".join(result)
+
+
+def build_obs_prompt(obs: Any) -> str:
+    """Format an observation into the user prompt for the LLM.
+
+    The focused ad gets full investigation prose for deep reasoning.
+    Other investigated ads are compressed to compact key-value summaries
+    that preserve all fields (signal + noise) for cross-ad comparison.
+    """
+    focused_ad: Optional[str] = None
+    if obs.current_ad_info:
+        m = re.search(r"Ad in Focus:\s*(ad_\d+)", obs.current_ad_info)
+        if m:
+            focused_ad = m.group(1)
+
+    parts = [
+        f"Queue: {obs.queue_summary}",
+        f"Current Ad: {obs.current_ad_info}",
+        f"Feedback: {obs.feedback}",
+        f"Available ads: {', '.join(obs.available_ads)}",
+    ]
+    if obs.verdict_history_summary and obs.verdict_history_summary != "No verdicts yet.":
+        parts.append(f"Verdicts: {obs.verdict_history_summary}")
+    if obs.investigation_findings:
+        findings = _build_compact_findings(obs.investigation_findings, focused_ad)
+        if findings:
+            parts.append(f"Findings:\n{findings}")
+    return "\n\n".join(parts)
+
+# ---------------------------------------------------------------------------
+# Episode logger (markdown, for debugging — separate from mandatory stdout)
+# ---------------------------------------------------------------------------
+
+class EpisodeLogger:
+    """Logs the full agent-environment conversation to a markdown file."""
+
+    def __init__(self, task_id: str, log_dir: Path) -> None:
+        self.task_id = task_id
+        self.lines: List[str] = []
+        log_dir.mkdir(parents=True, exist_ok=True)
+        self.path = log_dir / f"{task_id}_conversation.md"
+        self._md(f"# Episode Log — {task_id}\n")
+
+    def step_start(self, step: int, obs_prompt: str) -> None:
+        self._md(f"\n## Step {step}\n")
+        self._md(f"### Observation (sent to LLM)\n```\n{obs_prompt}\n```\n")
+
+    def llm_response(self, step: int, raw: str, action: AdReviewAction | None, fallback: bool) -> None:
+        tag = " [FALLBACK]" if fallback else ""
+        self._md(f"### LLM Response\n```json\n{raw.strip()}\n```\n")
+        if action:
+            act_dict = action.model_dump(exclude_none=True, exclude={"metadata"})
+            self._md(f"### Parsed Action{tag}\n```json\n{json.dumps(act_dict, indent=2)}\n```\n")
+
+    def env_feedback(self, step: int, reward: float, done: bool, feedback: str) -> None:
+        self._md(f"### Environment Response\n- **Reward:** `{reward:+.2f}`\n- **Done:** `{done}`\n")
+        self._md(f"- **Feedback:** {feedback}\n")
+
+    def episode_end(self, score: float, steps: int, verdicts: int, total: int) -> None:
+        summary = f"Score: {score:.3f} | Steps: {steps} | Verdicts: {verdicts}/{total}"
+        self._md(f"\n---\n## Result\n**{summary}**\n")
+        self._flush()
+
+    def _md(self, text: str) -> None:
+        self.lines.append(text)
+
+    def _flush(self) -> None:
+        with open(self.path, "w", encoding="utf-8") as f:
+            f.write("\n".join(self.lines))
+
+# ---------------------------------------------------------------------------
+# Core task runner
+# ---------------------------------------------------------------------------
+
+def run_single_task(
+    task_id: str,
+    seed: int = 42,
+    env_base_url: str = "https://QuantumTransformer-AdArena.hf.space",
+) -> Dict[str, Any]:
+    """Run the baseline agent on a single task with mandatory [START]/[STEP]/[END] logging."""
+    base_url = os.getenv("API_BASE_URL") or API_BASE_URL
+    api_key = os.getenv("HF_TOKEN") or os.getenv("API_KEY") or API_KEY
+    model = os.getenv("MODEL_NAME") or MODEL_NAME
+
+    client = OpenAI(base_url=base_url, api_key=api_key, timeout=60.0)
+    env = AdFraudEnv(base_url=env_base_url).sync()
+    elog = EpisodeLogger(task_id, LOG_DIR)
+
+    config = TASK_CONFIGS.get(task_id)
+    max_steps = config.action_budget if config else 25
+
+    rewards: List[float] = []
+    steps_taken = 0
+    score = 0.0
+    success = False
+
+    log_start(task=task_id, env=BENCHMARK, model=model)
+
+    try:
+        env.connect()
+        result = env.reset(seed=seed, task_id=task_id)
+
+        messages: List[Dict[str, str]] = [
+            {"role": "system", "content": SYSTEM_PROMPT},
+        ]
+
+        while not result.done and steps_taken < max_steps:
+            obs = result.observation
+            user_prompt = build_obs_prompt(obs)
+            messages.append({"role": "user", "content": user_prompt})
+
+            elog.step_start(steps_taken, user_prompt)
+
+            try:
+                completion = client.chat.completions.create(
+                    model=model,
+                    messages=messages,
+                    temperature=TEMPERATURE,
+                    max_tokens=MAX_TOKENS,
+                    stream=False,
+                )
+                response_text = completion.choices[0].message.content or "{}"
+            except Exception as exc:
+                logger.warning("Model request failed on step %d: %s", steps_taken, exc)
+                response_text = "{}"
+
+            messages.append({"role": "assistant", "content": response_text})
+
+            error_msg = None
+            fallback = False
+            try:
+                action_data = _extract_json(response_text)
+                action = AdReviewAction(**action_data)
+            except Exception as e:
+                logger.warning("Failed to parse action on step %d: %s", steps_taken, e)
+                fallback = True
+                error_msg = str(e)
+                if obs.available_ads:
+                    action = AdReviewAction(
+                        action_type="verdict",
+                        ad_id=obs.available_ads[0],
+                        verdict=FALLBACK_VERDICT,
+                        confidence=0.3,
+                    )
+                else:
+                    elog.llm_response(steps_taken, response_text, None, True)
+                    break
+
+            elog.llm_response(steps_taken, response_text, action, fallback)
+
+            result = env.step(action)
+            steps_taken += 1
+            reward = result.reward or 0.0
+            rewards.append(reward)
+
+            log_step(
+                step=steps_taken,
+                action=_format_action(action),
+                reward=reward,
+                done=result.done,
+                error=error_msg,
+            )
+
+            elog.env_feedback(steps_taken, reward, result.done, result.observation.feedback)
+
+            # Each observation is self-contained (all findings + verdict summary),
+            # so we only keep system prompt + last 2 exchanges to stay within
+            # context limits while preserving the agent's reasoning continuity.
+            if len(messages) > 6:
+                messages = messages[:1] + messages[-4:]
+
+        state = env.state()
+        score = state.grader_score if state.grader_score is not None else 0.0
+        score = max(0.0, min(1.0, score))
+        success = score > 0.0
+
+        elog.episode_end(score, steps_taken, state.reviewed_count, state.total_ads)
+
+        return {
+            "task_id": task_id,
+            "score": score,
+            "steps": steps_taken,
+            "verdicts": state.reviewed_count,
+            "total_ads": state.total_ads,
+        }
+
+    except Exception as e:
+        logger.error("Task %s failed: %s", task_id, e)
+        return {"task_id": task_id, "score": 0.0, "error": str(e)}
+
+    finally:
+        try:
+            env.close()
+        except Exception as e:
+            print(f"[DEBUG] env.close() error: {e}", flush=True)
+        log_end(success=success, steps=steps_taken, score=score, rewards=rewards)
+
+
+def run_baseline(
+    env_base_url: str = "https://QuantumTransformer-AdArena.hf.space",
+) -> Dict[str, Any]:
+    """Run baseline inference on all 3 tasks."""
+    model = os.getenv("MODEL_NAME") or MODEL_NAME
+    results: Dict[str, Any] = {}
+    for task_id in ["task_1", "task_2", "task_3"]:
+        logger.info("Running baseline for %s...", task_id)
+        try:
+            task_result = run_single_task(
+                task_id, seed=42, env_base_url=env_base_url,
+            )
+            results[task_id] = task_result
+            logger.info("  %s score: %.3f", task_id, task_result["score"])
+        except Exception as e:
+            logger.error("  %s failed: %s", task_id, e)
+            results[task_id] = {"task_id": task_id, "score": 0.0, "error": str(e)}
+
+    return {"baseline_model": model, "seed": 42, "tasks": results}
+
+
+def main() -> None:
+    logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
+
+    if not API_KEY:
+        print("Error: HF_TOKEN (or API_KEY) environment variable is required.", file=sys.stderr)
+        sys.exit(1)
+    env_base_url = os.getenv("AD_FRAUD_ENV_URL", "https://QuantumTransformer-AdArena.hf.space")
+
+    print(f"Running baseline inference against {env_base_url} with model {MODEL_NAME}...", file=sys.stderr)
+    scores = run_baseline(env_base_url=env_base_url)
+
+    output_path = Path(__file__).resolve().parent / "baseline_scores.json"
+    with open(output_path, "w") as f:
+        json.dump(scores, f, indent=2)
+
+    print(f"\nBaseline scores saved to {output_path}", file=sys.stderr)
+    print(json.dumps(scores, indent=2), file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()

models.py

@@ -0,0 +1,100 @@
+"""
+Data models for the Ad Fraud Investigation Environment.
+
+Defines the Action, Observation, and State types used by the OpenEnv interface.
+The agent interacts with a queue of ads, investigating them and rendering verdicts.
+"""
+
+from __future__ import annotations
+
+from typing import Dict, List, Literal, Optional
+
+from openenv.core.env_server.types import Action, Observation, State
+from pydantic import Field
+
+
+class AdReviewAction(Action):
+    """
+    Action space for the ad fraud investigation agent.
+
+    Three action types:
+    - investigate: Spend budget to reveal information about an ad
+    - verdict: Approve, reject, or escalate an ad
+    - link_accounts: Flag two ads as part of the same fraud network
+    """
+
+    action_type: Literal["investigate", "verdict", "link_accounts"]
+    ad_id: str = Field(..., description="Target ad identifier (e.g. 'ad_001')")
+
+    investigation_target: Optional[
+        Literal[
+            "advertiser_history",
+            "landing_page",
+            "payment_method",
+            "targeting_overlap",
+            "creative_similarity",
+            "campaign_structure",
+        ]
+    ] = Field(None, description="What to investigate (required for action_type='investigate')")
+
+    verdict: Optional[Literal["approve", "reject", "escalate"]] = Field(
+        None, description="Verdict decision (required for action_type='verdict')"
+    )
+    confidence: Optional[float] = Field(
+        None, ge=0.0, le=1.0, description="Agent's confidence in verdict (0.0-1.0)"
+    )
+
+    linked_ad_id: Optional[str] = Field(
+        None, description="Other ad in suspected fraud ring (required for action_type='link_accounts')"
+    )
+    link_reason: Optional[str] = Field(
+        None, description="Why the agent believes these ads are connected"
+    )
+
+
+class AdReviewObservation(Observation):
+    """
+    Observation returned after each step.
+
+    Text-heavy by design so LLM agents can reason about the content naturally.
+    Structured data is in queue_status for programmatic access.
+    """
+
+    queue_summary: str = Field(
+        default="", description="Natural language overview of the ad queue"
+    )
+    current_ad_info: str = Field(
+        default="", description="Details of the ad currently in focus"
+    )
+    investigation_findings: str = Field(
+        default="", description="Accumulated investigation results"
+    )
+    verdict_history_summary: str = Field(
+        default="", description="Summary of verdicts rendered so far"
+    )
+    feedback: str = Field(
+        default="", description="Natural language feedback on the last action taken"
+    )
+    available_ads: List[str] = Field(
+        default_factory=list, description="Ad IDs still pending review"
+    )
+    queue_status: Dict = Field(
+        default_factory=dict,
+        description="Structured status: total_ads, reviewed, pending, budget, step",
+    )
+
+
+class AdFraudState(State):
+    """
+    Internal environment state exposed via the state() property.
+
+    Inherits episode_id and step_count from State.
+    Uses extra='allow' so custom fields are permitted.
+    """
+
+    task_id: str = ""
+    total_ads: int = 0
+    reviewed_count: int = 0
+    remaining_budget: int = 0
+    verdicts: Dict = Field(default_factory=dict)
+    grader_score: Optional[float] = None

openenv.yaml

@@ -0,0 +1,7 @@
+spec_version: 1
+name: ad_fraud_env
+type: space
+sdk: docker
+runtime: fastapi
+app: server.app:app
+port: 8000

openenv_ad_fraud_env.egg-info/PKG-INFO

@@ -0,0 +1,18 @@
+Metadata-Version: 2.4
+Name: openenv-ad-fraud-env
+Version: 0.1.0
+Summary: Ad Fraud Investigation Environment for OpenEnv — real-world ad fraud detection and investigation simulation
+Requires-Python: >=3.11
+Requires-Dist: openenv-core[core]>=0.2.2
+Requires-Dist: fastapi>=0.115.0
+Requires-Dist: pydantic>=2.0.0
+Requires-Dist: uvicorn[standard]>=0.24.0
+Requires-Dist: requests>=2.31.0
+Requires-Dist: faker==33.1.0
+Requires-Dist: networkx>=3.2
+Requires-Dist: openai>=1.0.0
+Requires-Dist: python-dotenv>=1.0.0
+Provides-Extra: dev
+Requires-Dist: pytest>=8.0.0; extra == "dev"
+Requires-Dist: pytest-cov>=4.0.0; extra == "dev"
+Requires-Dist: pytest-asyncio>=0.23.0; extra == "dev"

openenv_ad_fraud_env.egg-info/SOURCES.txt

@@ -0,0 +1,33 @@
+README.md
+__init__.py
+client.py
+inference.py
+models.py
+pyproject.toml
+./__init__.py
+./client.py
+./inference.py
+./models.py
+data/__init__.py
+data/ad_generator.py
+data/advertiser_profiles.py
+data/fraud_patterns.py
+data/landing_pages.py
+data/network_generator.py
+graders/__init__.py
+graders/base_grader.py
+graders/task1_grader.py
+graders/task2_grader.py
+graders/task3_grader.py
+openenv_ad_fraud_env.egg-info/PKG-INFO
+openenv_ad_fraud_env.egg-info/SOURCES.txt
+openenv_ad_fraud_env.egg-info/dependency_links.txt
+openenv_ad_fraud_env.egg-info/entry_points.txt
+openenv_ad_fraud_env.egg-info/requires.txt
+openenv_ad_fraud_env.egg-info/top_level.txt
+server/__init__.py
+server/app.py
+server/environment.py
+tests/test_data_generation.py
+tests/test_environment.py
+tests/test_graders.py

openenv_ad_fraud_env.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

openenv_ad_fraud_env.egg-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+server = ad_fraud_env.server.app:main

openenv_ad_fraud_env.egg-info/requires.txt

@@ -0,0 +1,14 @@
+openenv-core[core]>=0.2.2
+fastapi>=0.115.0
+pydantic>=2.0.0
+uvicorn[standard]>=0.24.0
+requests>=2.31.0
+faker==33.1.0
+networkx>=3.2
+openai>=1.0.0
+python-dotenv>=1.0.0
+
+[dev]
+pytest>=8.0.0
+pytest-cov>=4.0.0
+pytest-asyncio>=0.23.0

openenv_ad_fraud_env.egg-info/top_level.txt

@@ -0,0 +1 @@
+ad_fraud_env

outputs/task_1_conversation.md

@@ -0,0 +1,840 @@
+# Episode Log — task_1
+
+
+## Step 0
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 0 | Pending: 5 | Budget remaining: 25 actions | Step: 0
+
+Current Ad: === Ad in Focus: ad_001 ===
+Category: local_service
+Ad copy: "Award-winning pizza delivery in Brooklyn. Family recipes since 1985. Order from Sal's Pizzeria."
+Targeting: Adults 18-55, Brooklyn, interests: food delivery
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Episode started. Review the ad queue and begin your investigation.
+
+Available ads: ad_001, ad_002, ad_003, ad_004, ad_005
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_001",
+  "investigation_target": "landing_page"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_001",
+  "investigation_target": "landing_page"
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: landing_page for ad_001.
+--- Findings ---
+Domain: nutri.com
+Domain age: 1672 days
+SSL: Valid SSL certificate
+Registrar: Cloudflare Registrar
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 5%
+Key claims on page:
+  - Licensed and insured
+  - Established since 2010
+  - BBB accredited
+Suspicious elements:
+  None detected
+Content summary: E-commerce store with detailed product descriptions, customer reviews, FAQ section, and accessible support.
+
+
+## Step 1
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 0 | Pending: 5 | Budget remaining: 24 actions | Step: 1
+
+Current Ad: === Ad in Focus: ad_001 ===
+Category: local_service
+Ad copy: "Award-winning pizza delivery in Brooklyn. Family recipes since 1985. Order from Sal's Pizzeria."
+Targeting: Adults 18-55, Brooklyn, interests: food delivery
+Initial risk signals: None
+Investigations completed: landing_page
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Investigation complete: landing_page for ad_001.
+--- Findings ---
+Domain: nutri.com
+Domain age: 1672 days
+SSL: Valid SSL certificate
+Registrar: Cloudflare Registrar
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 5%
+Key claims on page:
+  - Licensed and insured
+  - Established since 2010
+  - BBB accredited
+Suspicious elements:
+  None detected
+Content summary: E-commerce store with detailed product descriptions, customer reviews, FAQ section, and accessible support.
+
+Available ads: ad_001, ad_002, ad_003, ad_004, ad_005
+
+Findings:
+[ad_001 / landing_page]
+Domain: nutri.com
+Domain age: 1672 days
+SSL: Valid SSL certificate
+Registrar: Cloudflare Registrar
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 5%
+Key claims on page:
+  - Licensed and insured
+  - Established since 2010
+  - BBB accredited
+Suspicious elements:
+  None detected
+Content summary: E-commerce store with detailed product descrip
+```
+
+### LLM Response
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_001",
+  "verdict": "approve",
+  "confidence": 0.9
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_001",
+  "verdict": "approve",
+  "confidence": 0.9
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_001: approve (confidence: 0.90). Correct decision. Reward: +0.10
+
+
+## Step 2
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 1 | Pending: 4 | Budget remaining: 24 actions | Step: 2
+
+Current Ad: === Ad in Focus: ad_002 ===
+Category: counterfeit_goods
+Ad copy: "Rolex watches starting at $49.99! 100% genuine! Free worldwide shipping! LuxWatchDeals.net"
+Targeting: Men 25-55, interests: watches, luxury
+Initial risk signals: impossibly low prices for luxury goods, unfamiliar domain, claims of authenticity with extreme discounts
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_001: approve (confidence: 0.90). Correct decision. Reward: +0.10
+
+Available ads: ad_002, ad_003, ad_004, ad_005
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+
+Findings:
+[ad_001 / landing_page]
+Domain: nutri.com
+Domain age: 1672 days
+SSL: Valid SSL certificate
+Registrar: Cloudflare Registrar
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 5%
+Key claims on page:
+  - Licensed and insured
+  - Established since 2010
+  - BBB accredited
+Suspicious elements:
+  None detected
+Content summary: E-commerce store with detailed product descriptions, customer reviews, FAQ section, and accessible support.
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_002",
+  "investigation_target": "landing_page"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_002",
+  "investigation_target": "landing_page"
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: landing_page for ad_002.
+--- Findings ---
+Domain: store4751.site
+Domain age: 6 days
+SSL: NO SSL / expired certificate
+Registrar: NameSilo
+Redirect chain: store4751.site -> redirect651..top -> redirect167..top -> redirect364..click -> redirect687..top -> final-719..site
+Similarity to known scam templates: 67%
+Key claims on page:
+  - Authentic products
+  - Direct from manufacturer
+  - 90% below retail price
+Suspicious elements:
+  - Privacy policy copied from template
+  - Terms of service link returns 404
+  - No physical address listed
+  - Price comparison shows 90%+ discount vs retail
+  - No contact phone number
+Content summary: Product catalog showing luxury branded items at extreme discounts. Stock photos. Payment accepted via wire transfer and crypto only.
+
+
+## Step 3
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 1 | Pending: 4 | Budget remaining: 23 actions | Step: 3
+
+Current Ad: === Ad in Focus: ad_002 ===
+Category: counterfeit_goods
+Ad copy: "Rolex watches starting at $49.99! 100% genuine! Free worldwide shipping! LuxWatchDeals.net"
+Targeting: Men 25-55, interests: watches, luxury
+Initial risk signals: impossibly low prices for luxury goods, unfamiliar domain, claims of authenticity with extreme discounts
+Investigations completed: landing_page
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Investigation complete: landing_page for ad_002.
+--- Findings ---
+Domain: store4751.site
+Domain age: 6 days
+SSL: NO SSL / expired certificate
+Registrar: NameSilo
+Redirect chain: store4751.site -> redirect651..top -> redirect167..top -> redirect364..click -> redirect687..top -> final-719..site
+Similarity to known scam templates: 67%
+Key claims on page:
+  - Authentic products
+  - Direct from manufacturer
+  - 90% below retail price
+Suspicious elements:
+  - Privacy policy copied from template
+  - Terms of service link returns 404
+  - No physical address listed
+  - Price comparison shows 90%+ disco... [truncated]
+
+Available ads: ad_002, ad_003, ad_004, ad_005
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+
+Findings:
+[ad_001 / landing_page]
+Domain: nutri.com
+Domain age: 1672 days
+SSL: Valid SSL certificate
+Registrar: Cloudflare Registrar
+Redirect chain: No 
+```
+
+### LLM Response
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_002",
+  "verdict": "reject",
+  "confidence": 0.95
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_002",
+  "verdict": "reject",
+  "confidence": 0.95
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_002: reject (confidence: 0.95). Correct decision. Reward: +0.38
+
+
+## Step 4
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 2 | Pending: 3 | Budget remaining: 23 actions | Step: 4
+
+Current Ad: === Ad in Focus: ad_003 ===
+Category: fake_giveaway
+Ad copy: "EXCLUSIVE: Win a Tesla Model 3! Enter your email NOW! Drawing ends tonight!!!"
+Targeting: Adults 18-55, interests: cars, Tesla, electric vehicles
+Initial risk signals: excessive urgency language, too-good-to-be-true offer, excessive punctuation/caps, no purchase necessary claims
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_002: reject (confidence: 0.95). Correct decision. Reward: +0.38
+
+Available ads: ad_003, ad_004, ad_005
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+
+Findings:
+[ad_001 / landing_page]
+Domain: nutri.com
+Domain age: 1672 days
+SSL: Valid SSL certificate
+Registrar: Cloudflare Registrar
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 5%
+Key claims on page:
+  - Licensed and insured
+  - Established since 2010
+  - BBB accredited
+Suspicious elements:
+  None detected
+Content summary: E-commerce store with detailed product descriptions, customer reviews, FAQ section, and accessible support.
+
+[ad_002 / landing_page]
+Domain: store4751.site
+Domain age: 6 days
+SSL: NO SSL / expired certificate
+Registrar: NameSilo
+Redirect chain: store4751.... [truncated]
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_003",
+  "investigation_target": "landing_page"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_003",
+  "investigation_target": "landing_page"
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: landing_page for ad_003.
+--- Findings ---
+Domain: get8189.store
+Domain age: 26 days
+SSL: NO SSL / expired certificate
+Registrar: Tucows (privacy proxy)
+Redirect chain: get8189.store -> redirect253..xyz -> redirect575..xyz -> final-270..store
+Similarity to known scam templates: 56%
+Key claims on page:
+  - 100% free, no purchase necessary
+  - Guaranteed winner
+  - Act now — expires in 2 hours
+Suspicious elements:
+  - No contact phone number
+  - Requests personal info before any value provided
+  - No physical address listed
+  - Privacy policy copied from template
+  - Terms of service link returns 404
+Content summary: Landing page features a large countdown timer and a form requesting name, email, phone, and address. No clear sponsor or rules disclosure.
+
+
+## Step 5
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 2 | Pending: 3 | Budget remaining: 22 actions | Step: 5
+
+Current Ad: === Ad in Focus: ad_003 ===
+Category: fake_giveaway
+Ad copy: "EXCLUSIVE: Win a Tesla Model 3! Enter your email NOW! Drawing ends tonight!!!"
+Targeting: Adults 18-55, interests: cars, Tesla, electric vehicles
+Initial risk signals: excessive urgency language, too-good-to-be-true offer, excessive punctuation/caps, no purchase necessary claims
+Investigations completed: landing_page
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Investigation complete: landing_page for ad_003.
+--- Findings ---
+Domain: get8189.store
+Domain age: 26 days
+SSL: NO SSL / expired certificate
+Registrar: Tucows (privacy proxy)
+Redirect chain: get8189.store -> redirect253..xyz -> redirect575..xyz -> final-270..store
+Similarity to known scam templates: 56%
+Key claims on page:
+  - 100% free, no purchase necessary
+  - Guaranteed winner
+  - Act now — expires in 2 hours
+Suspicious elements:
+  - No contact phone number
+  - Requests personal info before any value provided
+  - No physical address listed
+  - Privacy policy copied from template
+  - Terms... [truncated]
+
+Available ads: ad_003, ad_004, ad_005
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+
+Findings:
+[ad_001 / landing_page]
+Domain: nutri.com
+Domain age: 1672 days
+SSL: Valid SSL certificate
+Registrar: Clou
+```
+
+### LLM Response
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_003",
+  "verdict": "reject",
+  "confidence": 0.98
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_003",
+  "verdict": "reject",
+  "confidence": 0.98
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_003: reject (confidence: 0.98). Correct decision. Reward: +0.39
+
+
+## Step 6
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 3 | Pending: 2 | Budget remaining: 22 actions | Step: 6
+
+Current Ad: === Ad in Focus: ad_004 ===
+Category: ecommerce
+Ad copy: "Spring Sale — Up to 30% off home essentials. Free shipping on orders over $50. Shop now at HomeNest."
+Targeting: Women 25-54, interests: home decor, shopping
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_003: reject (confidence: 0.98). Correct decision. Reward: +0.39
+
+Available ads: ad_004, ad_005
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: reject (confidence: 0.98)
+
+Findings:
+[ad_001 / landing_page]
+Domain: nutri.com
+Domain age: 1672 days
+SSL: Valid SSL certificate
+Registrar: Cloudflare Registrar
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 5%
+Key claims on page:
+  - Licensed and insured
+  - Established since 2010
+  - BBB accredited
+Suspicious elements:
+  None detected
+Content summary: E-commerce store with detailed product descriptions, customer reviews, FAQ section, and accessible support.
+
+[ad_002 / landing_page]
+Domain: store4751.site
+Domain age: 6 days
+SSL: NO SSL / expired certificate
+Registrar: NameSilo
+Redirect chain: store4751.... [truncated]
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_004",
+  "investigation_target": "landing_page"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_004",
+  "investigation_target": "landing_page"
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: landing_page for ad_004.
+--- Findings ---
+Domain: sparkle.io
+Domain age: 2424 days
+SSL: Valid SSL certificate
+Registrar: AWS Route 53
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 13%
+Key claims on page:
+  - 4.5-star average customer rating
+  - 14-day free trial
+Suspicious elements:
+  None detected
+Content summary: E-commerce store with detailed product descriptions, customer reviews, FAQ section, and accessible support.
+
+
+## Step 7
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 3 | Pending: 2 | Budget remaining: 21 actions | Step: 7
+
+Current Ad: === Ad in Focus: ad_004 ===
+Category: ecommerce
+Ad copy: "Spring Sale — Up to 30% off home essentials. Free shipping on orders over $50. Shop now at HomeNest."
+Targeting: Women 25-54, interests: home decor, shopping
+Initial risk signals: None
+Investigations completed: landing_page
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Investigation complete: landing_page for ad_004.
+--- Findings ---
+Domain: sparkle.io
+Domain age: 2424 days
+SSL: Valid SSL certificate
+Registrar: AWS Route 53
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 13%
+Key claims on page:
+  - 4.5-star average customer rating
+  - 14-day free trial
+Suspicious elements:
+  None detected
+Content summary: E-commerce store with detailed product descriptions, customer reviews, FAQ section, and accessible support.
+
+Available ads: ad_004, ad_005
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: reject (confidence: 0.98)
+
+Findings:
+[ad_001 / landing_page]
+Domain: nutri.com
+Domain age: 1672 days
+SSL: Valid SSL certificate
+Registrar: Cloudflare Registrar
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 5%
+Key claims on page:
+  - Licensed and insured
+  - Established since 2010
+  - BBB accredited
+Suspicious elements:
+  No
+```
+
+### LLM Response
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_004",
+  "verdict": "approve",
+  "confidence": 0.92
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_004",
+  "verdict": "approve",
+  "confidence": 0.92
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_004: approve (confidence: 0.92). Correct decision. Reward: +0.10
+
+
+## Step 8
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 4 | Pending: 1 | Budget remaining: 21 actions | Step: 8
+
+Current Ad: === Ad in Focus: ad_005 ===
+Category: fake_giveaway
+Ad copy: "EXCLUSIVE: Win a Tesla Model 3! Enter your email NOW! Drawing ends tonight!!!"
+Targeting: Adults 18-55, interests: cars, Tesla, electric vehicles
+Initial risk signals: excessive urgency language, too-good-to-be-true offer, excessive punctuation/caps, no purchase necessary claims
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_004: approve (confidence: 0.92). Correct decision. Reward: +0.10
+
+Available ads: ad_005
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: reject (confidence: 0.98)
+  ad_004: approve (confidence: 0.92)
+
+Findings:
+[ad_001 / landing_page]
+Domain: nutri.com
+Domain age: 1672 days
+SSL: Valid SSL certificate
+Registrar: Cloudflare Registrar
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 5%
+Key claims on page:
+  - Licensed and insured
+  - Established since 2010
+  - BBB accredited
+Suspicious elements:
+  None detected
+Content summary: E-commerce store with detailed product descriptions, customer reviews, FAQ section, and accessible support.
+
+[ad_002 / landing_page]
+Domain: store4751.site
+Domain age: 6 days
+SSL: NO SSL / expired certificate
+Registrar: NameSilo
+Redire
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_005",
+  "investigation_target": "landing_page"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_005",
+  "investigation_target": "landing_page"
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: landing_page for ad_005.
+--- Findings ---
+Domain: pro3268.top
+Domain age: 10 days
+SSL: Valid SSL certificate
+Registrar: Epik
+Redirect chain: pro3268.top -> redirect614..top -> final-501..site
+Similarity to known scam templates: 74%
+Key claims on page:
+  - 100% free, no purchase necessary
+  - Guaranteed winner
+  - Act now — expires in 2 hours
+Suspicious elements:
+  - No contact phone number
+  - Terms of service link returns 404
+  - Requests personal info before any value provided
+Content summary: Landing page features a large countdown timer and a form requesting name, email, phone, and address. No clear sponsor or rules disclosure.
+
+
+## Step 9
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 4 | Pending: 1 | Budget remaining: 20 actions | Step: 9
+
+Current Ad: === Ad in Focus: ad_005 ===
+Category: fake_giveaway
+Ad copy: "EXCLUSIVE: Win a Tesla Model 3! Enter your email NOW! Drawing ends tonight!!!"
+Targeting: Adults 18-55, interests: cars, Tesla, electric vehicles
+Initial risk signals: excessive urgency language, too-good-to-be-true offer, excessive punctuation/caps, no purchase necessary claims
+Investigations completed: landing_page
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Investigation complete: landing_page for ad_005.
+--- Findings ---
+Domain: pro3268.top
+Domain age: 10 days
+SSL: Valid SSL certificate
+Registrar: Epik
+Redirect chain: pro3268.top -> redirect614..top -> final-501..site
+Similarity to known scam templates: 74%
+Key claims on page:
+  - 100% free, no purchase necessary
+  - Guaranteed winner
+  - Act now — expires in 2 hours
+Suspicious elements:
+  - No contact phone number
+  - Terms of service link returns 404
+  - Requests personal info before any value provided
+Content summary: Landing page features a large countdown timer and a form requesting name, e... [truncated]
+
+Available ads: ad_005
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: reject (confidence: 0.98)
+  ad_004: approve (confidence: 0.92)
+
+Findings:
+[ad_001 / landing_page]
+Domain: nutri.com
+Domain 
+```
+
+### LLM Response
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_005",
+  "verdict": "reject",
+  "confidence": 0.99
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_005",
+  "verdict": "reject",
+  "confidence": 0.99
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Episode complete. Grader score: 0.931/1.000
+Verdicts rendered: 5/5
+Auto-approved (unreviewed): 0 fraud ad(s) missed
+Correct decisions: 5/5
+
+
+## Step 10
+
+### Observation (sent to LLM)
+```
+Queue: Task: Basic Ad Triage (easy)
+Total ads: 5 | Reviewed: 5 | Pending: 0 | Budget remaining: 20 actions | Step: 10
+
+Current Ad: 
+
+Feedback: Episode complete. Grader score: 0.931/1.000
+Verdicts rendered: 5/5
+Auto-approved (unreviewed): 0 fraud ad(s) missed
+Correct decisions: 5/5
+
+Available ads: 
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: reject (confidence: 0.98)
+  ad_004: approve (confidence: 0.92)
+  ad_005: reject (confidence: 0.99)
+
+Findings:
+[ad_001 / landing_page]
+Domain: nutri.com
+Domain age: 1672 days
+SSL: Valid SSL certificate
+Registrar: Cloudflare Registrar
+Redirect chain: No redirects (direct)
+Similarity to known scam templates: 5%
+Key claims on page:
+  - Licensed and insured
+  - Established since 2010
+  - BBB accredited
+Suspicious elements:
+  None detected
+Content summary: E-commerce store with detailed product descriptions, customer reviews, FAQ section, and accessible support.
+
+[ad_002 / landing_page]
+Domain: store4751.site
+Domain age: 6 days
+SSL: NO SSL / expired certificate
+Registrar: NameSilo
+Redirect chain: store4751.... [truncated]
+```
+
+### LLM Response
+```json
+{}
+```
+
+
+---
+## Result
+**Score: 0.931 | Steps: 10 | Verdicts: 5/5**

outputs/task_2_conversation.md

@@ -0,0 +1,1251 @@
+# Episode Log — task_2
+
+
+## Step 0
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 0 | Pending: 12 | Budget remaining: 30 actions | Step: 0
+
+Current Ad: === Ad in Focus: ad_001 ===
+Category: local_service
+Ad copy: "Award-winning pizza delivery in Brooklyn. Family recipes since 1985. Order from Sal's Pizzeria."
+Targeting: Adults 18-55, Brooklyn, interests: food delivery
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Episode started. Review the ad queue and begin your investigation.
+
+Available ads: ad_001, ad_002, ad_003, ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_001",
+  "investigation_target": "advertiser_history"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_001",
+  "investigation_target": "advertiser_history"
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: advertiser_history for ad_001.
+--- Findings ---
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+
+## Step 1
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 0 | Pending: 12 | Budget remaining: 29 actions | Step: 1
+
+Current Ad: === Ad in Focus: ad_001 ===
+Category: local_service
+Ad copy: "Award-winning pizza delivery in Brooklyn. Family recipes since 1985. Order from Sal's Pizzeria."
+Targeting: Adults 18-55, Brooklyn, interests: food delivery
+Initial risk signals: None
+Investigations completed: advertiser_history
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Investigation complete: advertiser_history for ad_001.
+--- Findings ---
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+Available ads: ad_001, ad_002, ad_003, ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_001",
+  "investigation_target": "payment_method"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_001",
+  "investigation_target": "payment_method"
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: payment_method for ad_001.
+--- Findings ---
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+
+
+## Step 2
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 0 | Pending: 12 | Budget remaining: 28 actions | Step: 2
+
+Current Ad: === Ad in Focus: ad_001 ===
+Category: local_service
+Ad copy: "Award-winning pizza delivery in Brooklyn. Family recipes since 1985. Order from Sal's Pizzeria."
+Targeting: Adults 18-55, Brooklyn, interests: food delivery
+Initial risk signals: None
+Investigations completed: advertiser_history, payment_method
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Investigation complete: payment_method for ad_001.
+--- Findings ---
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+
+Available ads: ad_001, ad_002, ad_003, ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+```
+
+### LLM Response
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_001",
+  "verdict": "approve",
+  "confidence": 0.9
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_001",
+  "verdict": "approve",
+  "confidence": 0.9
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_001: approve (confidence: 0.90). Correct decision. Reward: +0.10
+
+
+## Step 3
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 1 | Pending: 11 | Budget remaining: 28 actions | Step: 3
+
+Current Ad: === Ad in Focus: ad_002 ===
+Category: fake_giveaway
+Ad copy: "CONGRATULATIONS!! You've been selected to win a FREE iPhone 16 Pro!! Click NOW before offer expires!!!"
+Targeting: All adults 18+, broad targeting, worldwide
+Initial risk signals: excessive urgency language, too-good-to-be-true offer, excessive punctuation/caps, no purchase necessary claims
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_001: approve (confidence: 0.90). Correct decision. Reward: +0.10
+
+Available ads: ad_002, ad_003, ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_002",
+  "investigation_target": "landing_page"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_002",
+  "investigation_target": "landing_page"
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: landing_page for ad_002.
+--- Findings ---
+Domain: shop4020.top
+Domain age: 11 days
+SSL: NO SSL / expired certificate
+Registrar: Njalla (privacy)
+Redirect chain: shop4020.top -> redirect702..top -> redirect864..click -> final-235..site
+Similarity to known scam templates: 95%
+Key claims on page:
+  - 100% free, no purchase necessary
+  - Guaranteed winner
+  - Act now — expires in 2 hours
+Suspicious elements:
+  - No contact phone number
+  - Countdown timer with fake urgency
+  - No physical address listed
+  - Requests personal info before any value provided
+Content summary: Landing page features a large countdown timer and a form requesting name, email, phone, and address. No clear sponsor or rules disclosure.
+
+
+## Step 4
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 1 | Pending: 11 | Budget remaining: 27 actions | Step: 4
+
+Current Ad: === Ad in Focus: ad_002 ===
+Category: fake_giveaway
+Ad copy: "CONGRATULATIONS!! You've been selected to win a FREE iPhone 16 Pro!! Click NOW before offer expires!!!"
+Targeting: All adults 18+, broad targeting, worldwide
+Initial risk signals: excessive urgency language, too-good-to-be-true offer, excessive punctuation/caps, no purchase necessary claims
+Investigations completed: landing_page
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Investigation complete: landing_page for ad_002.
+--- Findings ---
+Domain: shop4020.top
+Domain age: 11 days
+SSL: NO SSL / expired certificate
+Registrar: Njalla (privacy)
+Redirect chain: shop4020.top -> redirect702..top -> redirect864..click -> final-235..site
+Similarity to known scam templates: 95%
+Key claims on page:
+  - 100% free, no purchase necessary
+  - Guaranteed winner
+  - Act now — expires in 2 hours
+Suspicious elements:
+  - No contact phone number
+  - Countdown timer with fake urgency
+  - No physical address listed
+  - Requests personal info before any value provided
+Content summary: L... [truncated]
+
+Available ads: ad_002, ad_003, ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+
+Findings:
+[ad_001 / advertiser_history]
+Advert
+```
+
+### LLM Response
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_002",
+  "verdict": "reject",
+  "confidence": 0.95
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_002",
+  "verdict": "reject",
+  "confidence": 0.95
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_002: reject (confidence: 0.95). Correct decision. Reward: +0.39
+
+
+## Step 5
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 2 | Pending: 10 | Budget remaining: 27 actions | Step: 5
+
+Current Ad: === Ad in Focus: ad_003 ===
+Category: fake_crypto
+Ad copy: "Web3 real estate tokenization. Own fractions of premium properties. Starting at $100. Licensed broker-dealer. PropChain.io"
+Targeting: Adults 30-55, interests: real estate, investing, crypto
+Initial risk signals: high guaranteed returns, complex financial product
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_002: reject (confidence: 0.95). Correct decision. Reward: +0.39
+
+Available ads: ad_003, ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+
+[ad_002 / landing_page]
+Domain: shop4020.top
+Domain age: 11 days
+SSL: NO SSL / expired certificate
+Registrar: Njalla (privacy)
+Redir
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_003",
+  "investigation_target": "landing_page"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_003",
+  "investigation_target": "landing_page"
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: landing_page for ad_003.
+--- Findings ---
+Domain: best3603.xyz
+Domain age: 9 days
+SSL: NO SSL / expired certificate
+Registrar: NameSilo
+Redirect chain: best3603.xyz -> redirect157..xyz -> final-164..store
+Similarity to known scam templates: 84%
+Key claims on page:
+  - 12-15% monthly returns guaranteed
+  - Audited by top security firms
+  - Risk-free investment
+Suspicious elements:
+  - Team photos are stock images
+  - Privacy policy copied from template
+  - No contact phone number
+  - Terms of service link returns 404
+Content summary: Professional-looking platform with dashboard screenshots. Whitepaper link leads to a generic PDF. Team bios use stock photos.
+
+
+## Step 6
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 2 | Pending: 10 | Budget remaining: 26 actions | Step: 6
+
+Current Ad: === Ad in Focus: ad_003 ===
+Category: fake_crypto
+Ad copy: "Web3 real estate tokenization. Own fractions of premium properties. Starting at $100. Licensed broker-dealer. PropChain.io"
+Targeting: Adults 30-55, interests: real estate, investing, crypto
+Initial risk signals: high guaranteed returns, complex financial product
+Investigations completed: landing_page
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Investigation complete: landing_page for ad_003.
+--- Findings ---
+Domain: best3603.xyz
+Domain age: 9 days
+SSL: NO SSL / expired certificate
+Registrar: NameSilo
+Redirect chain: best3603.xyz -> redirect157..xyz -> final-164..store
+Similarity to known scam templates: 84%
+Key claims on page:
+  - 12-15% monthly returns guaranteed
+  - Audited by top security firms
+  - Risk-free investment
+Suspicious elements:
+  - Team photos are stock images
+  - Privacy policy copied from template
+  - No contact phone number
+  - Terms of service link returns 404
+Content summary: Professional-looking platform with da... [truncated]
+
+Available ads: ad_003, ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+
+Findings:
+[ad_001 / advertiser_history]
+Advertis
+```
+
+### LLM Response
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_003",
+  "investigation_target": "payment_method"
+}
+```
+
+### Parsed Action
+```json
+{
+  "action_type": "investigate",
+  "ad_id": "ad_003",
+  "investigation_target": "payment_method"
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Investigation complete: payment_method for ad_003.
+--- Findings ---
+Payment Method Analysis for ad_003:
+  Method type: wire_transfer
+  Payment ID: pmt_711878
+  No other accounts in the current queue share this payment method.
+  Payment history: 1 chargebacks/disputes on record.
+
+
+## Step 7
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 2 | Pending: 10 | Budget remaining: 25 actions | Step: 7
+
+Current Ad: === Ad in Focus: ad_003 ===
+Category: fake_crypto
+Ad copy: "Web3 real estate tokenization. Own fractions of premium properties. Starting at $100. Licensed broker-dealer. PropChain.io"
+Targeting: Adults 30-55, interests: real estate, investing, crypto
+Initial risk signals: high guaranteed returns, complex financial product
+Investigations completed: landing_page, payment_method
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Investigation complete: payment_method for ad_003.
+--- Findings ---
+Payment Method Analysis for ad_003:
+  Method type: wire_transfer
+  Payment ID: pmt_711878
+  No other accounts in the current queue share this payment method.
+  Payment history: 1 chargebacks/disputes on record.
+
+Available ads: ad_003, ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: p
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_003",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_003: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 8
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 3 | Pending: 9 | Budget remaining: 25 actions | Step: 8
+
+Current Ad: === Ad in Focus: ad_004 ===
+Category: gray_area_supplements
+Ad copy: "SleepDeep Ultra — melatonin-free natural sleep aid. Non-habit forming. Money-back guarantee. PharmGrade Labs."
+Targeting: Adults 30-65, interests: sleep, insomnia, wellness
+Initial risk signals: supplement with strong efficacy claims, medical-sounding language
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_003: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: escalate (confidence: 0.30)
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+
+[ad_002 / landing_page]
+Domain: shop4020.top
+Domain age: 11 days
+SSL: NO SSL / expired certi
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_004",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_004: escalate (confidence: 0.30). Correct decision. Reward: +0.15
+
+
+## Step 9
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 4 | Pending: 8 | Budget remaining: 25 actions | Step: 9
+
+Current Ad: === Ad in Focus: ad_005 ===
+Category: gray_area_supplements
+Ad copy: "TestoMax Elite — natural testosterone support for men over 40. Doctor-formulated. Published clinical trial results."
+Targeting: Men 40-65, interests: men's health, testosterone, fitness
+Initial risk signals: supplement with strong efficacy claims, medical-sounding language
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_004: escalate (confidence: 0.30). Correct decision. Reward: +0.15
+
+Available ads: ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+
+[ad_002 / landing_page]
+Domain: shop4020.top
+Domain
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_005",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_005: escalate (confidence: 0.30). Correct decision. Reward: +0.15
+
+
+## Step 10
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 5 | Pending: 7 | Budget remaining: 25 actions | Step: 10
+
+Current Ad: === Ad in Focus: ad_006 ===
+Category: ecommerce
+Ad copy: "Handcrafted leather bags — designed to last. Each piece unique. ArtisanHide.com"
+Targeting: Adults 25-55, interests: fashion, luxury goods
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_005: escalate (confidence: 0.30). Correct decision. Reward: +0.15
+
+Available ads: ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+
+[ad_002 / landing_page]
+Domain: shop4020.top
+Domain age: 11 days
+SSL: NO SSL / expired certificate
+Registrar: Njalla (privacy)
+Redirect chai
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_006",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_006: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 11
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 6 | Pending: 6 | Budget remaining: 25 actions | Step: 11
+
+Current Ad: === Ad in Focus: ad_007 ===
+Category: advance_fee
+Ad copy: "URGENT: I am Prince Adebayo of Nigeria. I need your help to transfer $4.5 million. You keep 30%."
+Targeting: All adults 18+, broad targeting
+Initial risk signals: advance fee request, unsolicited money offer, urgency pressure, request for personal information
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_006: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_007, ad_008, ad_009, ad_010, ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+
+[ad_002 / la
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_007",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_007: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 12
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 7 | Pending: 5 | Budget remaining: 25 actions | Step: 12
+
+Current Ad: === Ad in Focus: ad_008 ===
+Category: saas
+Ad copy: "Automate your invoicing in minutes. Trusted by 50,000+ small businesses. InvoiceSimple."
+Targeting: Small business owners, interests: finance, accounting
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_007: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_008, ad_009, ad_010, ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+
+[ad_002 / landing_page]
+Domain: shop4020.top
+Domain age: 11 days
+SSL
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_008",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_008: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 13
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 8 | Pending: 4 | Budget remaining: 25 actions | Step: 13
+
+Current Ad: === Ad in Focus: ad_009 ===
+Category: ecommerce
+Ad copy: "Spring Sale — Up to 30% off home essentials. Free shipping on orders over $50. Shop now at HomeNest."
+Targeting: Women 25-54, interests: home decor, shopping
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_008: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_009, ad_010, ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+
+[ad_002 / landing_page]
+Domai
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_009",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_009: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 14
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 9 | Pending: 3 | Budget remaining: 25 actions | Step: 14
+
+Current Ad: === Ad in Focus: ad_010 ===
+Category: ecommerce
+Ad copy: "Handcrafted leather bags — designed to last. Each piece unique. ArtisanHide.com"
+Targeting: Adults 25-55, interests: fashion, luxury goods
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_009: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_010, ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+
+[ad_002 / landing_
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_010",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_010: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 15
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 10 | Pending: 2 | Budget remaining: 25 actions | Step: 15
+
+Current Ad: === Ad in Focus: ad_011 ===
+Category: fake_giveaway
+Ad copy: "EXCLUSIVE: Win a Tesla Model 3! Enter your email NOW! Drawing ends tonight!!!"
+Targeting: Adults 18-55, interests: cars, Tesla, electric vehicles
+Initial risk signals: excessive urgency language, too-good-to-be-true offer, excessive punctuation/caps, no purchase necessary claims
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_010: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_011, ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis fo
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_011",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_011: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 16
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 11 | Pending: 1 | Budget remaining: 25 actions | Step: 16
+
+Current Ad: === Ad in Focus: ad_012 ===
+Category: counterfeit_goods
+Ad copy: "Rolex watches starting at $49.99! 100% genuine! Free worldwide shipping! LuxWatchDeals.net"
+Targeting: Men 25-55, interests: watches, luxury
+Initial risk signals: impossibly low prices for luxury goods, unfamiliar domain, claims of authenticity with extreme discounts
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_011: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_012
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+  ad_011: escalate (con... [truncated]
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Pay
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_012",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Episode complete. Grader score: 0.684/1.000
+Verdicts rendered: 12/12
+Auto-approved (unreviewed): 0 fraud ad(s) missed
+Correct decisions: 4/12
+
+
+## Step 17
+
+### Observation (sent to LLM)
+```
+Queue: Task: Sophisticated Fraud Under Budget Pressure (medium)
+Total ads: 12 | Reviewed: 12 | Pending: 0 | Budget remaining: 25 actions | Step: 17
+
+Current Ad: 
+
+Feedback: Episode complete. Grader score: 0.684/1.000
+Verdicts rendered: 12/12
+Auto-approved (unreviewed): 0 fraud ad(s) missed
+Correct decisions: 4/12
+
+Available ads: 
+
+Verdicts:   ad_001: approve (confidence: 0.90)
+  ad_002: reject (confidence: 0.95)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+  ad_011: escalate (con... [truncated]
+
+Findings:
+[ad_001 / advertiser_history]
+Advertiser: FitVisit (Verified Business)
+Account age: 1558 days
+Country: Canada
+Total historical spend: $55,594.09
+Ads submitted in last 30 days: 13
+Historical approval rate: 91%
+Payment method: credit_card (ID: pmt_865179)
+
+[ad_001 / payment_method]
+Payment Method Analysis for ad_001:
+  Method type: credit_card
+  Payment ID: pmt_865179
+  No other accounts in the current queue share this payment method.
+
+[ad_002 / landing_page]
+Domain: shop4020.top
+Domain age: 11 days
+SSL: NO SSL / expired certificate
+Registrar: Njalla (privacy)
+Redirect chain: shop4020.top -> red... [truncated]
+```
+
+### LLM Response
+```json
+{}
+```
+
+
+---
+## Result
+**Score: 0.684 | Steps: 17 | Verdicts: 12/12**

outputs/task_3_conversation.md

@@ -0,0 +1,1057 @@
+# Episode Log — task_3
+
+
+## Step 0
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 0 | Pending: 20 | Budget remaining: 40 actions | Step: 0
+
+Current Ad: === Ad in Focus: ad_001 ===
+Category: saas
+Ad copy: "Automate your invoicing in minutes. Trusted by 50,000+ small businesses. InvoiceSimple."
+Targeting: Small business owners, interests: finance, accounting
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Episode started. Review the ad queue and begin your investigation.
+
+Available ads: ad_001, ad_002, ad_003, ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012, ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_001",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_001: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 1
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 1 | Pending: 19 | Budget remaining: 40 actions | Step: 1
+
+Current Ad: === Ad in Focus: ad_002 ===
+Category: advance_fee
+Ad copy: "You've won the UK National Lottery! Send $100 for tax processing to receive your $1.5M prize."
+Targeting: Adults 25-70, interests: lottery, gambling, luck
+Initial risk signals: advance fee request, unsolicited money offer, urgency pressure, request for personal information
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_001: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_002, ad_003, ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012, ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_002",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_002: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 2
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 2 | Pending: 18 | Budget remaining: 40 actions | Step: 2
+
+Current Ad: === Ad in Focus: ad_003 ===
+Category: advance_fee
+Ad copy: "Congratulations! You've inherited $2.8 million from a distant relative. Pay $200 processing fee to claim."
+Targeting: Adults 30-70, interests: finance, inheritance
+Initial risk signals: advance fee request, unsolicited money offer, urgency pressure, request for personal information
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_002: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_003, ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012, ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_003",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_003: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 3
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 3 | Pending: 17 | Budget remaining: 40 actions | Step: 3
+
+Current Ad: === Ad in Focus: ad_004 ===
+Category: network_crypto
+Ad copy: "Crypto tax software — auto-import from 50+ exchanges. CPA-approved reports. TaxChain Solutions."
+Targeting: Adults 25-55, interests: crypto, tax, finance
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_003: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_004, ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012, ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_004",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_004: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 4
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 4 | Pending: 16 | Budget remaining: 40 actions | Step: 4
+
+Current Ad: === Ad in Focus: ad_005 ===
+Category: fake_crypto
+Ad copy: "Invest in DeFi with confidence. Our AI-powered trading platform delivers 12-15% monthly returns. Audited smart contracts. Join CryptoYield Pro."
+Targeting: Men 25-45, interests: cryptocurrency, DeFi, investing
+Initial risk signals: high guaranteed returns, complex financial product
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_004: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_005, ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012, ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_005",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_005: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 5
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 5 | Pending: 15 | Budget remaining: 40 actions | Step: 5
+
+Current Ad: === Ad in Focus: ad_006 ===
+Category: ecommerce
+Ad copy: "Spring Sale — Up to 30% off home essentials. Free shipping on orders over $50. Shop now at HomeNest."
+Targeting: Women 25-54, interests: home decor, shopping
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_005: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_006, ad_007, ad_008, ad_009, ad_010, ad_011, ad_012, ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_006",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_006: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 6
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 6 | Pending: 14 | Budget remaining: 40 actions | Step: 6
+
+Current Ad: === Ad in Focus: ad_007 ===
+Category: gray_area_supplements
+Ad copy: "SlimBiome Pro — proprietary probiotic blend for healthy weight management. Featured in Health Magazine. GMP certified."
+Targeting: Adults 30-60, interests: weight loss, probiotics, gut health
+Initial risk signals: supplement with strong efficacy claims, medical-sounding language
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_006: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_007, ad_008, ad_009, ad_010, ad_011, ad_012, ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_007",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_007: escalate (confidence: 0.30). Correct decision. Reward: +0.15
+
+
+## Step 7
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 7 | Pending: 13 | Budget remaining: 40 actions | Step: 7
+
+Current Ad: === Ad in Focus: ad_008 ===
+Category: celebrity_endorsement_fraud
+Ad copy: "Oprah's favorite weight loss secret finally revealed. 'I lost 40 pounds without dieting.' Doctor-approved method."
+Targeting: Women 25-60, interests: weight loss, Oprah, wellness
+Initial risk signals: celebrity name used without verification, urgency/exclusivity language
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_007: escalate (confidence: 0.30). Correct decision. Reward: +0.15
+
+Available ads: ad_008, ad_009, ad_010, ad_011, ad_012, ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_008",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_008: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 8
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 8 | Pending: 12 | Budget remaining: 40 actions | Step: 8
+
+Current Ad: === Ad in Focus: ad_009 ===
+Category: education
+Ad copy: "Learn Python programming in 12 weeks. Live instructor, project-based curriculum. CodeAcademy Pro."
+Targeting: Adults 18-35, interests: programming, tech careers
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_008: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_009, ad_010, ad_011, ad_012, ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_009",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_009: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 9
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 9 | Pending: 11 | Budget remaining: 40 actions | Step: 9
+
+Current Ad: === Ad in Focus: ad_010 ===
+Category: fake_giveaway
+Ad copy: "EXCLUSIVE: Win a Tesla Model 3! Enter your email NOW! Drawing ends tonight!!!"
+Targeting: Adults 18-55, interests: cars, Tesla, electric vehicles
+Initial risk signals: excessive urgency language, too-good-to-be-true offer, excessive punctuation/caps, no purchase necessary claims
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_009: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_010, ad_011, ad_012, ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_010",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_010: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 10
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 10 | Pending: 10 | Budget remaining: 40 actions | Step: 10
+
+Current Ad: === Ad in Focus: ad_011 ===
+Category: ecommerce
+Ad copy: "Handcrafted leather bags — designed to last. Each piece unique. ArtisanHide.com"
+Targeting: Adults 25-55, interests: fashion, luxury goods
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_010: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_011, ad_012, ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_011",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_011: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 11
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 11 | Pending: 9 | Budget remaining: 40 actions | Step: 11
+
+Current Ad: === Ad in Focus: ad_012 ===
+Category: gray_area_supplements
+Ad copy: "Joint Restore Plus — advanced collagen peptide formula. Thousands of 5-star reviews. Orthopedist recommended."
+Targeting: Adults 45-75, interests: joint health, supplements, arthritis
+Initial risk signals: supplement with strong efficacy claims, medical-sounding language
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_011: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_012, ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+  ad_011: escalate (... [truncated]
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_012",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_012: escalate (confidence: 0.30). Correct decision. Reward: +0.15
+
+
+## Step 12
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 12 | Pending: 8 | Budget remaining: 40 actions | Step: 12
+
+Current Ad: === Ad in Focus: ad_013 ===
+Category: gray_area_supplements
+Ad copy: "NeuroPeak — clinically studied nootropic for enhanced focus and memory. 30-day money-back guarantee. FDA-registered facility."
+Targeting: Adults 25-55, interests: nootropics, productivity, brain health
+Initial risk signals: supplement with strong efficacy claims, medical-sounding language
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_012: escalate (confidence: 0.30). Correct decision. Reward: +0.15
+
+Available ads: ad_013, ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+  ad_011: escalate (... [truncated]
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_013",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_013: escalate (confidence: 0.30). Correct decision. Reward: +0.15
+
+
+## Step 13
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 13 | Pending: 7 | Budget remaining: 40 actions | Step: 13
+
+Current Ad: === Ad in Focus: ad_014 ===
+Category: fake_giveaway
+Ad copy: "CONGRATULATIONS!! You've been selected to win a FREE iPhone 16 Pro!! Click NOW before offer expires!!!"
+Targeting: All adults 18+, broad targeting, worldwide
+Initial risk signals: excessive urgency language, too-good-to-be-true offer, excessive punctuation/caps, no purchase necessary claims
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_013: escalate (confidence: 0.30). Correct decision. Reward: +0.15
+
+Available ads: ad_014, ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+  ad_011: escalate (... [truncated]
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_014",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_014: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 14
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 14 | Pending: 6 | Budget remaining: 40 actions | Step: 14
+
+Current Ad: === Ad in Focus: ad_015 ===
+Category: gray_area_supplements
+Ad copy: "TestoMax Elite — natural testosterone support for men over 40. Doctor-formulated. Published clinical trial results."
+Targeting: Men 40-65, interests: men's health, testosterone, fitness
+Initial risk signals: supplement with strong efficacy claims, medical-sounding language
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_014: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_015, ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+  ad_011: escalate (... [truncated]
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_015",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_015: escalate (confidence: 0.30). Correct decision. Reward: +0.15
+
+
+## Step 15
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 15 | Pending: 5 | Budget remaining: 40 actions | Step: 15
+
+Current Ad: === Ad in Focus: ad_016 ===
+Category: ecommerce
+Ad copy: "Handcrafted leather bags — designed to last. Each piece unique. ArtisanHide.com"
+Targeting: Adults 25-55, interests: fashion, luxury goods
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_015: escalate (confidence: 0.30). Correct decision. Reward: +0.15
+
+Available ads: ad_016, ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+  ad_011: escalate (... [truncated]
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_016",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_016: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 16
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 16 | Pending: 4 | Budget remaining: 40 actions | Step: 16
+
+Current Ad: === Ad in Focus: ad_017 ===
+Category: miracle_cure
+Ad copy: "Reverse aging in 14 days! Scientists discover fountain of youth pill! Limited supply!"
+Targeting: Adults 35-65, interests: anti-aging, beauty, health
+Initial risk signals: unsubstantiated medical claims, conspiracy language ('they don't want you to know'), impossible health promises, urgency/scarcity tactics
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_016: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_017, ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+  ad_011: escalate (... [truncated]
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_017",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_017: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 17
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 17 | Pending: 3 | Budget remaining: 40 actions | Step: 17
+
+Current Ad: === Ad in Focus: ad_018 ===
+Category: local_service
+Ad copy: "Award-winning pizza delivery in Brooklyn. Family recipes since 1985. Order from Sal's Pizzeria."
+Targeting: Adults 18-55, Brooklyn, interests: food delivery
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_017: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_018, ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+  ad_011: escalate (... [truncated]
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_018",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_018: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 18
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 18 | Pending: 2 | Budget remaining: 40 actions | Step: 18
+
+Current Ad: === Ad in Focus: ad_019 ===
+Category: network_ecommerce
+Ad copy: "Robot vacuum with LiDAR mapping. Self-emptying dock. Smart home compatible. CleanBot Pro."
+Targeting: Adults 25-55, interests: smart home, cleaning
+Initial risk signals: None
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_018: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_019, ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+  ad_011: escalate (... [truncated]
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_019",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Verdict recorded for ad_019: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+
+## Step 19
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 19 | Pending: 1 | Budget remaining: 40 actions | Step: 19
+
+Current Ad: === Ad in Focus: ad_020 ===
+Category: clone_brand
+Ad copy: "PayPaI Security Alert: Your account needs verification. Update your information to avoid suspension. Secure portal."
+Targeting: Adults 18-65, interests: online payments, PayPal
+Initial risk signals: brand name misspelling, unofficial domain
+Investigations completed: None yet
+Available investigation targets: advertiser_history, landing_page, payment_method, targeting_overlap, creative_similarity
+
+Feedback: Verdict recorded for ad_019: escalate (confidence: 0.30). Incorrect decision. Reward: -0.05
+
+Available ads: ad_020
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+  ad_011: escalate (... [truncated]
+```
+
+### LLM Response
+```json
+{}
+```
+
+### Parsed Action [FALLBACK]
+```json
+{
+  "action_type": "verdict",
+  "ad_id": "ad_020",
+  "verdict": "escalate",
+  "confidence": 0.3
+}
+```
+
+### Environment Response
+- **Reward:** `+0.00`
+- **Done:** `False`
+
+- **Feedback:** Episode complete. Grader score: 0.579/1.000
+Verdicts rendered: 20/20
+Auto-approved (unreviewed): 0 fraud ad(s) missed
+Correct decisions: 4/20
+
+
+## Step 20
+
+### Observation (sent to LLM)
+```
+Queue: Task: Coordinated Fraud Network Detection (hard)
+Total ads: 20 | Reviewed: 20 | Pending: 0 | Budget remaining: 40 actions | Step: 20
+
+Current Ad: 
+
+Feedback: Episode complete. Grader score: 0.579/1.000
+Verdicts rendered: 20/20
+Auto-approved (unreviewed): 0 fraud ad(s) missed
+Correct decisions: 4/20
+
+Available ads: 
+
+Verdicts:   ad_001: escalate (confidence: 0.30)
+  ad_002: escalate (confidence: 0.30)
+  ad_003: escalate (confidence: 0.30)
+  ad_004: escalate (confidence: 0.30)
+  ad_005: escalate (confidence: 0.30)
+  ad_006: escalate (confidence: 0.30)
+  ad_007: escalate (confidence: 0.30)
+  ad_008: escalate (confidence: 0.30)
+  ad_009: escalate (confidence: 0.30)
+  ad_010: escalate (confidence: 0.30)
+  ad_011: escalate (... [truncated]
+```
+
+### LLM Response
+```json
+{}
+```
+
+
+---
+## Result
+**Score: 0.579 | Steps: 20 | Verdicts: 20/20**

pyproject.toml

@@ -0,0 +1,44 @@
+[build-system]
+requires = ["setuptools>=45", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "openenv-ad-fraud-env"
+version = "0.1.0"
+description = "Ad Fraud Investigation Environment for OpenEnv — real-world ad fraud detection and investigation simulation"
+requires-python = ">=3.11"
+dependencies = [
+    "openenv-core[core]>=0.2.2",
+    "fastapi>=0.115.0",
+    "pydantic>=2.0.0",
+    "uvicorn[standard]>=0.24.0",
+    "requests>=2.31.0",
+    "faker==33.1.0",
+    "networkx>=3.2",
+    "openai>=1.0.0",
+    "python-dotenv>=1.0.0",
+]
+
+[project.optional-dependencies]
+dev = [
+    "pytest>=8.0.0",
+    "pytest-cov>=4.0.0",
+    "pytest-asyncio>=0.23.0",
+]
+
+[project.scripts]
+server = "ad_fraud_env.server.app:main"
+
+[tool.setuptools]
+packages = [
+    "ad_fraud_env",
+    "ad_fraud_env.server",
+    "ad_fraud_env.data",
+    "ad_fraud_env.graders",
+]
+
+[tool.setuptools.package-dir]
+"ad_fraud_env" = "."
+"ad_fraud_env.server" = "server"
+"ad_fraud_env.data" = "data"
+"ad_fraud_env.graders" = "graders"

server/__init__.py

@@ -0,0 +1 @@
+"""Server package for the Ad Fraud Investigation Environment."""

server/app.py

@@ -0,0 +1,113 @@
+"""
+FastAPI application for the Ad Fraud Investigation Environment.
+
+Creates the OpenEnv server via create_app() and registers custom HTTP
+endpoints required by the hackathon: /tasks, /baseline, /grader.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+from pathlib import Path
+from typing import Any, Dict
+
+from openenv.core.env_server import create_app
+
+try:
+    from ..models import AdReviewAction, AdReviewObservation
+except ImportError:
+    from models import AdReviewAction, AdReviewObservation
+
+from .environment import AdFraudEnvironment, get_last_grader_result
+
+logger = logging.getLogger(__name__)
+
+app = create_app(
+    AdFraudEnvironment,
+    AdReviewAction,
+    AdReviewObservation,
+    env_name="ad_fraud_env",
+)
+
+
+# ------------------------------------------------------------------
+# Custom endpoints required by the competition
+# ------------------------------------------------------------------
+
+
+@app.get("/tasks", tags=["Competition"])
+async def tasks() -> Dict[str, Any]:
+    """Return the list of tasks and the action schema."""
+    try:
+        from ..data.ad_generator import TASK_CONFIGS
+    except ImportError:
+        from data.ad_generator import TASK_CONFIGS
+
+    task_list = []
+    for tid, cfg in TASK_CONFIGS.items():
+        task_list.append({
+            "id": cfg.task_id,
+            "name": cfg.name,
+            "difficulty": cfg.difficulty,
+            "queue_size": cfg.queue_size,
+            "action_budget": cfg.action_budget,
+            "description": cfg.description,
+        })
+
+    return {
+        "tasks": task_list,
+        "action_schema": AdReviewAction.model_json_schema(),
+    }
+
+
+@app.get("/baseline", tags=["Competition"])
+async def baseline() -> Dict[str, Any]:
+    """Return baseline scores, running live inference if credentials are available."""
+    baseline_path = Path(__file__).resolve().parent.parent / "baseline_scores.json"
+
+    has_creds = all(os.getenv(v) for v in ("API_BASE_URL", "MODEL_NAME", "HF_TOKEN"))
+    if has_creds:
+        try:
+            try:
+                from ..inference import run_baseline
+            except ImportError:
+                from inference import run_baseline
+            scores = run_baseline()
+            with open(baseline_path, "w") as f:
+                json.dump(scores, f, indent=2)
+            return scores
+        except Exception as e:
+            logger.warning("Live baseline failed, falling back to cached: %s", e)
+
+    if baseline_path.exists():
+        with open(baseline_path) as f:
+            return json.load(f)
+
+    return {
+        "error": "No baseline scores available. Set API_BASE_URL, MODEL_NAME, and HF_TOKEN to run live inference.",
+        "tasks": {},
+    }
+
+
+@app.get("/grader", tags=["Competition"])
+async def grader() -> Dict[str, Any]:
+    """Return grader score from the most recently completed episode."""
+    result = get_last_grader_result()
+    if not result:
+        return {
+            "error": "No completed episode. Run an episode via WebSocket first.",
+            "grader_score": None,
+        }
+    return result
+
+
+def main() -> None:
+    """Entry point for direct execution."""
+    import uvicorn
+    uvicorn.run(app, host="0.0.0.0", port=8000)
+
+
+if __name__ == "__main__":
+    main()

server/environment.py

@@ -0,0 +1,559 @@
+"""
+Core Ad Fraud Investigation Environment.
+
+Implements the OpenEnv Environment interface. The agent reviews a queue of ads,
+investigates them, and renders verdicts under a budget constraint.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Any, Dict, List, Optional, Set
+from uuid import uuid4
+
+from openenv.core.env_server.interfaces import Environment
+from openenv.core.env_server.types import State
+
+try:
+    from ..data.ad_generator import (
+        TASK_CONFIGS,
+        Ad,
+        GeneratedEpisode,
+        generate_episode,
+    )
+    from ..models import AdFraudState, AdReviewAction, AdReviewObservation
+    from ..graders.base_grader import EpisodeRecord, LinkResult, VerdictResult, grade_episode
+except ImportError:
+    from data.ad_generator import (
+        TASK_CONFIGS,
+        Ad,
+        GeneratedEpisode,
+        generate_episode,
+    )
+    from models import AdFraudState, AdReviewAction, AdReviewObservation
+    from graders.base_grader import EpisodeRecord, LinkResult, VerdictResult, grade_episode
+
+logger = logging.getLogger(__name__)
+
+# Module-level store so the /grader endpoint can read the last score.
+_last_grader_result: Dict[str, Any] = {}
+
+
+def get_last_grader_result() -> Dict[str, Any]:
+    return dict(_last_grader_result)
+
+
+class AdFraudEnvironment(
+    Environment[AdReviewAction, AdReviewObservation, AdFraudState]
+):
+    """
+    Ad fraud investigation environment.
+
+    Each episode is a review session: the agent processes a queue of N ads
+    within a limited action budget, choosing what to investigate and when
+    to render verdicts. Unreviewed ads auto-approve at episode end.
+    """
+
+    SUPPORTS_CONCURRENT_SESSIONS = True
+
+    def __init__(self) -> None:
+        super().__init__()
+        self._state = AdFraudState(episode_id=str(uuid4()), step_count=0)
+        self._episode: Optional[GeneratedEpisode] = None
+        self._verdicts: Dict[str, Dict[str, Any]] = {}
+        self._links: List[Dict[str, Any]] = []
+        self._investigations: Dict[str, List[str]] = {}
+        self._cumulative_reward: float = 0.0
+        self._done = False
+        self._last_feedback = ""
+        self._focused_ad_id: Optional[str] = None
+
+    # ------------------------------------------------------------------
+    # OpenEnv interface
+    # ------------------------------------------------------------------
+
+    def reset(
+        self,
+        seed: int | None = None,
+        episode_id: str | None = None,
+        **kwargs: Any,
+    ) -> AdReviewObservation:
+        task_id = kwargs.get("task_id", "task_1")
+        if task_id not in TASK_CONFIGS:
+            task_id = "task_1"
+
+        effective_seed = seed if seed is not None else hash(uuid4()) & 0xFFFFFFFF
+        self._episode = generate_episode(effective_seed, task_id)
+
+        config = self._episode.task_config
+        self._state = AdFraudState(
+            episode_id=episode_id or str(uuid4()),
+            step_count=0,
+            task_id=task_id,
+            total_ads=config.queue_size,
+            reviewed_count=0,
+            remaining_budget=config.action_budget,
+            verdicts={},
+            grader_score=None,
+        )
+        self._verdicts = {}
+        self._links = []
+        self._investigations = {}
+        self._cumulative_reward = 0.0
+        self._done = False
+        self._last_feedback = "Episode started. Review the ad queue and begin your investigation."
+        self._focused_ad_id = self._episode.ads[0].ad_id if self._episode.ads else None
+
+        return self._build_observation(reward=0.0, done=False)
+
+    def step(
+        self,
+        action: AdReviewAction,
+        timeout_s: float | None = None,
+        **kwargs: Any,
+    ) -> AdReviewObservation:
+        if self._done:
+            return self._build_observation(
+                reward=0.0, done=True,
+                feedback_override="Episode is already complete. Call reset() to start a new episode.",
+            )
+
+        if self._episode is None:
+            return self._build_observation(
+                reward=0.0, done=False,
+                feedback_override="Environment not initialized. Call reset() first.",
+            )
+
+        self._state.step_count += 1
+
+        ad_ids = {a.ad_id for a in self._episode.ads}
+        if action.ad_id not in ad_ids:
+            self._last_feedback = f"Invalid ad_id '{action.ad_id}'. Valid IDs: {', '.join(sorted(ad_ids))}"
+            return self._build_observation(reward=-0.05, done=False)
+
+        if action.action_type == "investigate":
+            reward = self._handle_investigate(action)
+        elif action.action_type == "verdict":
+            reward = self._handle_verdict(action)
+        elif action.action_type == "link_accounts":
+            reward = self._handle_link(action)
+        else:
+            self._last_feedback = f"Unknown action_type '{action.action_type}'."
+            reward = -0.05
+
+        self._cumulative_reward += reward
+
+        done = self._check_done()
+        if done and not self._done:
+            end_reward = self._handle_episode_end()
+            reward += end_reward
+            self._cumulative_reward += end_reward
+            self._done = True
+
+        self._state.remaining_budget = max(0, self._state.remaining_budget)
+        self._state.reviewed_count = len(self._verdicts)
+        self._state.verdicts = {
+            ad_id: v.get("verdict", "") for ad_id, v in self._verdicts.items()
+        }
+
+        return self._build_observation(reward=reward, done=self._done)
+
+    @property
+    def state(self) -> AdFraudState:
+        return self._state
+
+    # ------------------------------------------------------------------
+    # Action handlers
+    # ------------------------------------------------------------------
+
+    def _handle_investigate(self, action: AdReviewAction) -> float:
+        if self._state.remaining_budget <= 0:
+            self._last_feedback = "No budget remaining. You must render verdicts on remaining ads or end the episode."
+            return -0.02
+
+        if action.investigation_target is None:
+            self._last_feedback = "investigation_target is required for action_type='investigate'."
+            return -0.05
+
+        ad_id = action.ad_id
+        target = action.investigation_target
+
+        prev = self._investigations.setdefault(ad_id, [])
+        if target in prev:
+            self._last_feedback = (
+                f"You already investigated '{target}' for {ad_id}. "
+                "Choose a different investigation target or render a verdict."
+            )
+            return -0.02
+
+        if ad_id in self._verdicts:
+            self._last_feedback = f"You already rendered a verdict on {ad_id}. Choose a different ad."
+            return -0.02
+
+        self._state.remaining_budget -= 1
+        prev.append(target)
+        self._focused_ad_id = ad_id
+
+        findings = self._episode.investigation_data.get(ad_id, {}).get(
+            target, "No data available for this investigation type."
+        )
+        self._last_feedback = (
+            f"Investigation complete: {target} for {ad_id}.\n"
+            f"--- Findings ---\n{findings}"
+        )
+        return -0.02
+
+    def _handle_verdict(self, action: AdReviewAction) -> float:
+        ad_id = action.ad_id
+
+        if ad_id in self._verdicts:
+            self._last_feedback = f"You already rendered a verdict on {ad_id}."
+            return -0.02
+
+        if action.verdict is None:
+            self._last_feedback = "verdict field is required for action_type='verdict'."
+            return -0.05
+
+        confidence = action.confidence if action.confidence is not None else 0.5
+        ad = self._get_ad(ad_id)
+        ground_truth = ad.ground_truth_label if ad else "legit"
+        severity = ad.severity if ad else 0.0
+
+        self._verdicts[ad_id] = {
+            "verdict": action.verdict,
+            "confidence": confidence,
+            "ground_truth": ground_truth,
+        }
+
+        reward = self._compute_verdict_reward(action.verdict, ground_truth, severity, confidence)
+
+        pending = [a.ad_id for a in self._episode.ads if a.ad_id not in self._verdicts]
+        self._last_feedback = (
+            f"Verdict recorded for {ad_id}: {action.verdict} "
+            f"(confidence: {confidence:.2f}). "
+            f"{len(pending)} ad(s) remaining in queue."
+        )
+
+        if pending:
+            self._focused_ad_id = pending[0]
+
+        return reward
+
+    def _handle_link(self, action: AdReviewAction) -> float:
+        if action.linked_ad_id is None:
+            self._last_feedback = "linked_ad_id is required for action_type='link_accounts'."
+            return -0.05
+
+        ad_ids = {a.ad_id for a in self._episode.ads}
+        if action.linked_ad_id not in ad_ids:
+            self._last_feedback = f"Invalid linked_ad_id '{action.linked_ad_id}'."
+            return -0.05
+
+        if action.ad_id == action.linked_ad_id:
+            self._last_feedback = "Cannot link an ad to itself."
+            return -0.05
+
+        link_key = tuple(sorted([action.ad_id, action.linked_ad_id]))
+        existing = {tuple(sorted([l["ad_id_1"], l["ad_id_2"]])) for l in self._links}
+        if link_key in existing:
+            self._last_feedback = f"Link between {action.ad_id} and {action.linked_ad_id} already recorded."
+            return -0.02
+
+        is_correct = self._check_link_correct(action.ad_id, action.linked_ad_id)
+
+        self._links.append({
+            "ad_id_1": action.ad_id,
+            "ad_id_2": action.linked_ad_id,
+            "reason": action.link_reason or "",
+            "correct": is_correct,
+        })
+
+        self._last_feedback = (
+            f"Network link recorded: {action.ad_id} <-> {action.linked_ad_id}. "
+            f"Reason: {action.link_reason or 'not specified'}."
+        )
+
+        return 0.4 if is_correct else -0.25
+
+    # ------------------------------------------------------------------
+    # Reward computation
+    # ------------------------------------------------------------------
+
+    def _compute_verdict_reward(
+        self, verdict: str, ground_truth: str, severity: float, confidence: float
+    ) -> float:
+        if verdict == "reject" and ground_truth == "fraud":
+            return 0.3 + 0.1 * severity
+        elif verdict == "approve" and ground_truth == "legit":
+            return 0.1
+        elif verdict == "escalate" and ground_truth == "escalate":
+            return 0.15
+        elif verdict == "reject" and ground_truth == "legit":
+            return -0.35
+        elif verdict == "approve" and ground_truth == "fraud":
+            return -0.5
+        elif verdict == "escalate":
+            return -0.05
+        elif verdict == "approve" and ground_truth == "escalate":
+            return -0.15
+        elif verdict == "reject" and ground_truth == "escalate":
+            return -0.1
+        else:
+            return -0.05
+
+    def _handle_episode_end(self) -> float:
+        """Apply end-of-episode adjustments for unreviewed ads, then delegate to graders."""
+        unreviewed_fraud = 0
+        for ad in self._episode.ads:
+            if ad.ad_id not in self._verdicts:
+                self._verdicts[ad.ad_id] = {
+                    "verdict": "approve",
+                    "confidence": 0.0,
+                    "ground_truth": ad.ground_truth_label,
+                    "auto_approved": True,
+                }
+                if ad.ground_truth_label == "fraud":
+                    unreviewed_fraud += 1
+
+        record = self._build_episode_record()
+        grader_score = grade_episode(record)
+        self._state.grader_score = grader_score
+
+        reviewed_count = len([v for v in self._verdicts.values() if not v.get("auto_approved")])
+        total_ads = len(self._episode.ads)
+        total_correct = sum(
+            1 for v in self._verdicts.values()
+            if not v.get("auto_approved")
+            and (
+                (v["verdict"] == "reject" and v["ground_truth"] == "fraud")
+                or (v["verdict"] == "approve" and v["ground_truth"] == "legit")
+                or (v["verdict"] == "escalate" and v["ground_truth"] == "escalate")
+            )
+        )
+        false_positives = sum(
+            1 for v in self._verdicts.values()
+            if not v.get("auto_approved")
+            and v["verdict"] == "reject" and v["ground_truth"] == "legit"
+        )
+        false_negatives = sum(
+            1 for v in self._verdicts.values()
+            if not v.get("auto_approved")
+            and v["verdict"] == "approve" and v["ground_truth"] == "fraud"
+        )
+        correct_links = sum(1 for l in self._links if l.get("correct"))
+        incorrect_links = sum(1 for l in self._links if not l.get("correct"))
+
+        global _last_grader_result
+        _last_grader_result = {
+            "task_id": self._state.task_id,
+            "grader_score": grader_score,
+            "episode_id": self._state.episode_id,
+            "total_steps": self._state.step_count,
+            "verdicts_rendered": reviewed_count,
+            "correct_decisions": total_correct,
+            "false_positives": false_positives,
+            "false_negatives": false_negatives,
+            "auto_approved": total_ads - reviewed_count,
+            "unreviewed_fraud": unreviewed_fraud,
+            "network_links_correct": correct_links,
+            "network_links_incorrect": incorrect_links,
+        }
+
+        feedback_lines = [
+            f"Episode complete. Grader score: {grader_score:.3f}/1.000",
+            f"Verdicts rendered: {reviewed_count}/{total_ads}",
+            f"Correct decisions: {total_correct}/{reviewed_count}",
+            f"False positives (legit rejected): {false_positives}",
+            f"False negatives (fraud approved): {false_negatives}",
+            f"Unreviewed ads auto-approved: {unreviewed_fraud}",
+        ]
+        if self._links:
+            feedback_lines.append(
+                f"Network links: {correct_links} correct, {incorrect_links} incorrect"
+            )
+        self._last_feedback = "\n".join(feedback_lines)
+
+        return 0.0
+
+    def _build_episode_record(self) -> EpisodeRecord:
+        """Convert internal state into an EpisodeRecord for the grader."""
+        verdict_results = []
+        for ad in self._episode.ads:
+            v = self._verdicts.get(ad.ad_id)
+            if v:
+                verdict_results.append(VerdictResult(
+                    ad_id=ad.ad_id,
+                    verdict=v["verdict"],
+                    confidence=v.get("confidence", 0.5),
+                    ground_truth=v["ground_truth"],
+                    auto_approved=v.get("auto_approved", False),
+                ))
+
+        link_results = [
+            LinkResult(ad_id_1=l["ad_id_1"], ad_id_2=l["ad_id_2"], correct=l["correct"])
+            for l in self._links
+        ]
+
+        ads_metadata = [
+            {"ad_id": ad.ad_id, "ground_truth": ad.ground_truth_label, "severity": ad.severity}
+            for ad in self._episode.ads
+        ]
+
+        return EpisodeRecord(
+            task_id=self._state.task_id,
+            total_steps=self._state.step_count,
+            action_budget=self._episode.task_config.action_budget,
+            verdicts=verdict_results,
+            links=link_results,
+            ads_metadata=ads_metadata,
+            n_fraud_rings=len(self._episode.fraud_rings),
+            ring_sizes=[len(r.member_ad_ids) for r in self._episode.fraud_rings],
+        )
+
+    # ------------------------------------------------------------------
+    # Helpers
+    # ------------------------------------------------------------------
+
+    def _check_done(self) -> bool:
+        if self._episode is None:
+            return True
+        all_reviewed = all(
+            ad.ad_id in self._verdicts for ad in self._episode.ads
+        )
+        steps_exhausted = self._state.step_count >= self._episode.task_config.action_budget
+        return all_reviewed or steps_exhausted
+
+    def _check_link_correct(self, ad_id_1: str, ad_id_2: str) -> bool:
+        """Check if two ads share a fraud ring."""
+        for ring in self._episode.fraud_rings:
+            if ad_id_1 in ring.member_ad_ids and ad_id_2 in ring.member_ad_ids:
+                return True
+        return False
+
+    def _get_ad(self, ad_id: str) -> Optional[Ad]:
+        if self._episode is None:
+            return None
+        for ad in self._episode.ads:
+            if ad.ad_id == ad_id:
+                return ad
+        return None
+
+    def _build_observation(
+        self,
+        reward: float,
+        done: bool,
+        feedback_override: str | None = None,
+    ) -> AdReviewObservation:
+        feedback = feedback_override or self._last_feedback
+
+        if self._episode is None:
+            return AdReviewObservation(
+                done=done,
+                reward=reward,
+                queue_summary="No episode loaded.",
+                current_ad_info="",
+                investigation_findings="",
+                verdict_history_summary="",
+                feedback=feedback,
+                available_ads=[],
+                queue_status={},
+            )
+
+        config = self._episode.task_config
+        pending = [a for a in self._episode.ads if a.ad_id not in self._verdicts]
+        reviewed = [a for a in self._episode.ads if a.ad_id in self._verdicts]
+
+        steps_remaining = max(0, config.action_budget - self._state.step_count)
+        queue_summary = (
+            f"Task: {config.name} ({config.difficulty})\n"
+            f"Total ads: {config.queue_size} | "
+            f"Reviewed: {len(reviewed)} | "
+            f"Pending: {len(pending)} | "
+            f"Steps remaining: {steps_remaining}/{config.action_budget} | "
+            f"Investigation budget: {self._state.remaining_budget} | "
+            f"Step: {self._state.step_count}"
+        )
+
+        current_ad_info = ""
+        if self._focused_ad_id and not done:
+            ad = self._get_ad(self._focused_ad_id)
+            if ad and ad.ad_id not in self._verdicts:
+                signals = ", ".join(ad.initial_risk_signals) if ad.initial_risk_signals else "None"
+                investigated = self._investigations.get(ad.ad_id, [])
+                inv_status = ", ".join(investigated) if investigated else "None yet"
+
+                # Contextual metadata visible before investigation
+                profile = self._episode.advertiser_profiles.get(ad.ad_id)
+                meta_lines = []
+                if profile:
+                    meta_lines.append(f"Advertiser country: {profile.country}")
+                    meta_lines.append(f"Account age: {profile.account_age_days} days")
+                    if profile.account_age_days < 30:
+                        meta_lines.append("Flag: New account (< 30 days)")
+                context_meta = "\n".join(meta_lines)
+
+                current_ad_info = (
+                    f"=== Ad in Focus: {ad.ad_id} ===\n"
+                    f"Category: {ad.category}\n"
+                    f"Ad copy: \"{ad.ad_copy}\"\n"
+                    f"Targeting: {ad.targeting_summary}\n"
+                    f"Initial risk signals: {signals}\n"
+                    f"{context_meta}\n"
+                    f"Investigations completed: {inv_status}\n"
+                    f"Available investigation targets: advertiser_history, landing_page, "
+                    f"payment_method, targeting_overlap, creative_similarity, campaign_structure"
+                )
+
+        investigation_findings = ""
+        for ad_id, targets in self._investigations.items():
+            for target in targets:
+                finding = self._episode.investigation_data.get(ad_id, {}).get(target, "")
+                if finding:
+                    investigation_findings += f"\n[{ad_id} / {target}]\n{finding}\n"
+
+        manual_verdicts = {
+            ad_id: v for ad_id, v in self._verdicts.items()
+            if not v.get("auto_approved")
+        }
+        if manual_verdicts:
+            counts = {"approve": 0, "reject": 0, "escalate": 0}
+            by_decision = {"approve": [], "reject": [], "escalate": []}
+            for ad_id, v in manual_verdicts.items():
+                counts[v["verdict"]] = counts.get(v["verdict"], 0) + 1
+                by_decision[v["verdict"]].append(ad_id)
+            summary_parts = [f"{c} {k}" for k, c in counts.items() if c > 0]
+            verdict_lines = [
+                f"Reviewed {len(manual_verdicts)} ad(s): {', '.join(summary_parts)}."
+            ]
+            for decision in ("reject", "approve", "escalate"):
+                if by_decision[decision]:
+                    verdict_lines.append(
+                        f"  {decision}: {', '.join(by_decision[decision])}"
+                    )
+            verdict_history_summary = "\n".join(verdict_lines)
+        else:
+            verdict_history_summary = "No verdicts yet."
+
+        available_ads = [a.ad_id for a in pending]
+
+        queue_status = {
+            "total_ads": config.queue_size,
+            "reviewed": len(reviewed),
+            "pending": len(pending),
+            "investigation_budget": self._state.remaining_budget,
+            "steps_remaining": steps_remaining,
+            "step": self._state.step_count,
+            "task_id": config.task_id,
+        }
+
+        return AdReviewObservation(
+            done=done,
+            reward=reward,
+            queue_summary=queue_summary,
+            current_ad_info=current_ad_info,
+            investigation_findings=investigation_findings.strip(),
+            verdict_history_summary=verdict_history_summary,
+            feedback=feedback,
+            available_ads=available_ads,
+            queue_status=queue_status,
+        )

server/requirements.txt

@@ -0,0 +1,9 @@
+openenv-core[core]>=0.2.2
+fastapi>=0.115.0
+pydantic>=2.0.0
+uvicorn[standard]>=0.24.0
+requests>=2.31.0
+faker==33.1.0
+networkx>=3.2
+openai>=1.0.0
+python-dotenv>=1.0.0

tests/test_data_generation.py

@@ -0,0 +1,141 @@
+"""Tests for deterministic data generation."""
+
+import json
+
+from ad_fraud_env.data.ad_generator import generate_episode
+
+
+class TestDeterminism:
+    def test_same_seed_produces_identical_output(self):
+        """Generate with seed=42 twice — output must be byte-identical."""
+        ep1 = generate_episode(seed=42, task_id="task_1")
+        ep2 = generate_episode(seed=42, task_id="task_1")
+
+        assert len(ep1.ads) == len(ep2.ads)
+        for a1, a2 in zip(ep1.ads, ep2.ads):
+            assert a1.ad_id == a2.ad_id
+            assert a1.ad_copy == a2.ad_copy
+            assert a1.ground_truth_label == a2.ground_truth_label
+
+        for ad_id in ep1.investigation_data:
+            for target in ep1.investigation_data[ad_id]:
+                assert (
+                    ep1.investigation_data[ad_id][target]
+                    == ep2.investigation_data[ad_id][target]
+                )
+
+    def test_different_seeds_produce_different_output(self):
+        ep1 = generate_episode(seed=42, task_id="task_1")
+        ep2 = generate_episode(seed=99, task_id="task_1")
+
+        copies_1 = {a.ad_copy for a in ep1.ads}
+        copies_2 = {a.ad_copy for a in ep2.ads}
+        assert copies_1 != copies_2
+
+    def test_task_configs_produce_correct_queue_sizes(self):
+        for task_id, expected_size in [("task_1", 5), ("task_2", 12), ("task_3", 20)]:
+            ep = generate_episode(seed=42, task_id=task_id)
+            assert len(ep.ads) == expected_size, f"{task_id}: expected {expected_size}, got {len(ep.ads)}"
+
+    def test_task3_has_fraud_rings(self):
+        ep = generate_episode(seed=42, task_id="task_3")
+        assert len(ep.fraud_rings) > 0, "Task 3 should have fraud rings"
+        for ring in ep.fraud_rings:
+            assert len(ring.member_ad_ids) >= 3
+            assert len(ring.shared_signals) >= 2
+            assert ring.topology in ("clique", "chain", "hub_spoke")
+
+    def test_investigation_data_exists_for_all_ads(self):
+        ep = generate_episode(seed=42, task_id="task_2")
+        expected_targets = [
+            "advertiser_history", "landing_page", "payment_method",
+            "targeting_overlap", "creative_similarity", "campaign_structure",
+        ]
+        for ad in ep.ads:
+            assert ad.ad_id in ep.investigation_data
+            for target in expected_targets:
+                assert target in ep.investigation_data[ad.ad_id], (
+                    f"Missing {target} for {ad.ad_id}"
+                )
+                assert len(ep.investigation_data[ad.ad_id][target]) > 0
+
+    def test_ground_truth_distribution(self):
+        ep = generate_episode(seed=42, task_id="task_2")
+        labels = [a.ground_truth_label for a in ep.ads]
+        assert "fraud" in labels
+        assert "legit" in labels
+
+
+class TestNoExplicitCrossAdReferences:
+    """Investigation text must not explicitly name other ad IDs."""
+
+    def test_payment_investigation_no_cross_refs(self):
+        ep = generate_episode(seed=42, task_id="task_3")
+        for ad_id, inv in ep.investigation_data.items():
+            text = inv["payment_method"]
+            for other_ad in ep.investigation_data:
+                if other_ad == ad_id:
+                    continue
+                assert other_ad not in text, (
+                    f"Payment investigation for {ad_id} references {other_ad}"
+                )
+
+    def test_targeting_investigation_no_cross_refs(self):
+        ep = generate_episode(seed=42, task_id="task_3")
+        for ad_id, inv in ep.investigation_data.items():
+            text = inv["targeting_overlap"]
+            assert "HIGH OVERLAP detected with:" not in text
+
+    def test_creative_investigation_no_cross_refs(self):
+        ep = generate_episode(seed=42, task_id="task_3")
+        for ad_id, inv in ep.investigation_data.items():
+            text = inv["creative_similarity"]
+            assert "STRONG SIMILARITY detected with:" not in text
+
+    def test_campaign_investigation_no_cross_refs(self):
+        ep = generate_episode(seed=42, task_id="task_3")
+        for ad_id, inv in ep.investigation_data.items():
+            text = inv["campaign_structure"]
+            assert "MATCH:" not in text
+
+
+class TestDecoysAndRealism:
+    def test_advertiser_profiles_have_temporal_signals(self):
+        ep = generate_episode(seed=42, task_id="task_2")
+        for ad_id, profile in ep.advertiser_profiles.items():
+            assert profile.account_created_date, f"Missing created date for {ad_id}"
+            assert profile.spend_velocity, f"Missing spend velocity for {ad_id}"
+            assert profile.ad_submission_pattern, f"Missing submission pattern for {ad_id}"
+
+    def test_temporal_signals_appear_in_investigation(self):
+        ep = generate_episode(seed=42, task_id="task_2")
+        for ad_id, inv in ep.investigation_data.items():
+            text = inv["advertiser_history"]
+            assert "Account created:" in text or "Account age:" in text
+            assert "Spend velocity:" in text or "spend" in text.lower()
+
+    def test_ring_members_share_creation_week(self):
+        """Ring members should have account creation dates within 7 days of each other."""
+        from datetime import date
+        ep = generate_episode(seed=42, task_id="task_3")
+        for ring in ep.fraud_rings:
+            dates = []
+            for ad_id in ring.member_ad_ids:
+                profile = ep.advertiser_profiles[ad_id]
+                d = date.fromisoformat(profile.account_created_date)
+                dates.append(d)
+            if len(dates) >= 2:
+                spread = (max(dates) - min(dates)).days
+                assert spread <= 7, (
+                    f"Ring {ring.ring_id} creation dates spread: {spread} days"
+                )
+
+    def test_investigation_has_whois_privacy_info(self):
+        ep = generate_episode(seed=42, task_id="task_2")
+        found_whois = False
+        for ad_id, inv in ep.investigation_data.items():
+            text = inv["landing_page"]
+            if "WHOIS privacy:" in text:
+                found_whois = True
+                break
+        assert found_whois, "At least one landing page should mention WHOIS privacy"

tests/test_environment.py

@@ -0,0 +1,178 @@
+"""Tests for the core AdFraudEnvironment."""
+
+from ad_fraud_env.models import AdReviewAction, AdReviewObservation, AdFraudState
+from ad_fraud_env.server.environment import AdFraudEnvironment
+
+
+class TestReset:
+    def test_reset_returns_observation(self):
+        env = AdFraudEnvironment()
+        obs = env.reset(seed=42, task_id="task_1")
+        assert isinstance(obs, AdReviewObservation)
+        assert obs.done is False
+        assert obs.reward == 0.0
+        assert len(obs.available_ads) == 5
+
+    def test_reset_clears_state(self):
+        env = AdFraudEnvironment()
+        env.reset(seed=42, task_id="task_1")
+        env.step(AdReviewAction(
+            action_type="verdict", ad_id="ad_001",
+            verdict="approve", confidence=0.9,
+        ))
+        obs = env.reset(seed=42, task_id="task_1")
+        state = env.state
+        assert state.step_count == 0
+        assert state.reviewed_count == 0
+        assert len(obs.available_ads) == 5
+
+    def test_reset_different_tasks(self):
+        env = AdFraudEnvironment()
+        for task_id, expected in [("task_1", 5), ("task_2", 12), ("task_3", 20)]:
+            obs = env.reset(seed=42, task_id=task_id)
+            assert len(obs.available_ads) == expected
+
+
+class TestStep:
+    def test_investigate_returns_findings(self):
+        env = AdFraudEnvironment()
+        env.reset(seed=42, task_id="task_1")
+        obs = env.step(AdReviewAction(
+            action_type="investigate",
+            ad_id="ad_001",
+            investigation_target="advertiser_history",
+        ))
+        assert obs.done is False
+        assert obs.reward == -0.02
+        assert "Advertiser" in obs.feedback or "Investigation complete" in obs.feedback
+
+    def test_verdict_correct_rejection(self):
+        env = AdFraudEnvironment()
+        env.reset(seed=42, task_id="task_1")
+        fraud_ads = [
+            a for a in env._episode.ads if a.ground_truth_label == "fraud"
+        ]
+        assert len(fraud_ads) > 0
+        ad = fraud_ads[0]
+        obs = env.step(AdReviewAction(
+            action_type="verdict", ad_id=ad.ad_id,
+            verdict="reject", confidence=0.9,
+        ))
+        assert obs.reward > 0
+
+    def test_verdict_false_negative_penalty(self):
+        env = AdFraudEnvironment()
+        env.reset(seed=42, task_id="task_1")
+        fraud_ads = [
+            a for a in env._episode.ads if a.ground_truth_label == "fraud"
+        ]
+        ad = fraud_ads[0]
+        obs = env.step(AdReviewAction(
+            action_type="verdict", ad_id=ad.ad_id,
+            verdict="approve", confidence=0.9,
+        ))
+        assert obs.reward < 0
+
+    def test_duplicate_verdict_rejected(self):
+        env = AdFraudEnvironment()
+        env.reset(seed=42, task_id="task_1")
+        env.step(AdReviewAction(
+            action_type="verdict", ad_id="ad_001",
+            verdict="approve", confidence=0.5,
+        ))
+        obs = env.step(AdReviewAction(
+            action_type="verdict", ad_id="ad_001",
+            verdict="reject", confidence=0.9,
+        ))
+        assert obs.reward == -0.02
+
+    def test_invalid_ad_id(self):
+        env = AdFraudEnvironment()
+        env.reset(seed=42, task_id="task_1")
+        obs = env.step(AdReviewAction(
+            action_type="investigate", ad_id="ad_999",
+            investigation_target="landing_page",
+        ))
+        assert obs.reward == -0.05
+        assert "Invalid" in obs.feedback
+
+    def test_episode_ends_when_all_reviewed(self):
+        env = AdFraudEnvironment()
+        obs = env.reset(seed=42, task_id="task_1")
+        for ad_id in list(obs.available_ads):
+            obs = env.step(AdReviewAction(
+                action_type="verdict", ad_id=ad_id,
+                verdict="reject", confidence=0.5,
+            ))
+        assert obs.done is True
+
+    def test_step_after_done_returns_done(self):
+        env = AdFraudEnvironment()
+        obs = env.reset(seed=42, task_id="task_1")
+        for ad_id in list(obs.available_ads):
+            obs = env.step(AdReviewAction(
+                action_type="verdict", ad_id=ad_id,
+                verdict="reject", confidence=0.5,
+            ))
+        obs = env.step(AdReviewAction(
+            action_type="investigate", ad_id="ad_001",
+            investigation_target="landing_page",
+        ))
+        assert obs.done is True
+        assert "already complete" in obs.feedback.lower()
+
+
+class TestState:
+    def test_state_tracks_progress(self):
+        env = AdFraudEnvironment()
+        env.reset(seed=42, task_id="task_1")
+        state = env.state
+        assert state.task_id == "task_1"
+        assert state.total_ads == 5
+        assert state.remaining_budget == 25
+        assert state.step_count == 0
+
+        env.step(AdReviewAction(
+            action_type="investigate", ad_id="ad_001",
+            investigation_target="landing_page",
+        ))
+        state = env.state
+        assert state.step_count == 1
+        assert state.remaining_budget == 24
+
+    def test_grader_score_set_on_completion(self):
+        env = AdFraudEnvironment()
+        obs = env.reset(seed=42, task_id="task_1")
+        for ad_id in list(obs.available_ads):
+            env.step(AdReviewAction(
+                action_type="verdict", ad_id=ad_id,
+                verdict="reject", confidence=0.5,
+            ))
+        state = env.state
+        assert state.grader_score is not None
+        assert 0.0 <= state.grader_score <= 1.0
+
+
+class TestAntiExploit:
+    def test_always_reject_scores_poorly(self):
+        """Always-reject on task_2 (5 legit / 5 fraud / 2 escalate) should be punished."""
+        env = AdFraudEnvironment()
+        obs = env.reset(seed=42, task_id="task_2")
+        for ad_id in list(obs.available_ads):
+            env.step(AdReviewAction(
+                action_type="verdict", ad_id=ad_id,
+                verdict="reject", confidence=0.9,
+            ))
+        score = env.state.grader_score
+        assert score < 0.7, f"Always-reject should score poorly, got {score}"
+
+    def test_always_escalate_scores_poorly(self):
+        env = AdFraudEnvironment()
+        obs = env.reset(seed=42, task_id="task_1")
+        for ad_id in list(obs.available_ads):
+            env.step(AdReviewAction(
+                action_type="verdict", ad_id=ad_id,
+                verdict="escalate", confidence=0.5,
+            ))
+        score = env.state.grader_score
+        assert score < 0.7, f"Always-escalate should score poorly, got {score}"

tests/test_graders.py

@@ -0,0 +1,203 @@
+"""Tests for the grading system."""
+
+from ad_fraud_env.graders.base_grader import (
+    BaseGrader,
+    EpisodeRecord,
+    LinkResult,
+    VerdictResult,
+    grade_episode,
+)
+
+
+def _make_record(
+    task_id: str = "task_1",
+    verdicts: list | None = None,
+    links: list | None = None,
+    total_steps: int = 5,
+    action_budget: int = 25,
+    ads_metadata: list | None = None,
+    n_fraud_rings: int = 0,
+    ring_sizes: list | None = None,
+) -> EpisodeRecord:
+    if verdicts is None:
+        verdicts = []
+    if links is None:
+        links = []
+    if ads_metadata is None:
+        ads_metadata = [
+            {"ad_id": v.ad_id, "severity": 0.8, "ground_truth": v.ground_truth}
+            for v in verdicts
+        ]
+    return EpisodeRecord(
+        task_id=task_id,
+        total_steps=total_steps,
+        action_budget=action_budget,
+        verdicts=verdicts,
+        links=links,
+        ads_metadata=ads_metadata,
+        n_fraud_rings=n_fraud_rings,
+        ring_sizes=ring_sizes,
+    )
+
+
+class TestGraderScoreRange:
+    def test_scores_in_valid_range(self):
+        verdicts = [
+            VerdictResult("ad_001", "reject", 0.9, "fraud"),
+            VerdictResult("ad_002", "approve", 0.9, "legit"),
+            VerdictResult("ad_003", "reject", 0.8, "fraud"),
+        ]
+        record = _make_record(verdicts=verdicts, total_steps=3)
+        score = grade_episode(record)
+        assert 0.0 <= score <= 1.0
+
+    def test_perfect_score_is_high(self):
+        verdicts = [
+            VerdictResult("ad_001", "reject", 0.95, "fraud"),
+            VerdictResult("ad_002", "approve", 0.95, "legit"),
+            VerdictResult("ad_003", "reject", 0.95, "fraud"),
+            VerdictResult("ad_004", "approve", 0.95, "legit"),
+            VerdictResult("ad_005", "reject", 0.95, "fraud"),
+        ]
+        record = _make_record(verdicts=verdicts, total_steps=5)
+        score = grade_episode(record)
+        assert score > 0.7, f"Perfect verdicts should score high, got {score}"
+
+    def test_all_wrong_scores_low(self):
+        verdicts = [
+            VerdictResult("ad_001", "approve", 0.9, "fraud"),
+            VerdictResult("ad_002", "reject", 0.9, "legit"),
+            VerdictResult("ad_003", "approve", 0.9, "fraud"),
+        ]
+        record = _make_record(verdicts=verdicts, total_steps=3)
+        score = grade_episode(record)
+        assert score < 0.3, f"All wrong verdicts should score low, got {score}"
+
+
+class TestTask2Grader:
+    def test_calibration_bonus(self):
+        well_calibrated = [
+            VerdictResult("ad_001", "reject", 0.9, "fraud"),
+            VerdictResult("ad_002", "approve", 0.9, "legit"),
+            VerdictResult("ad_003", "reject", 0.8, "fraud"),
+            VerdictResult("ad_004", "approve", 0.85, "legit"),
+        ]
+        poorly_calibrated = [
+            VerdictResult("ad_001", "reject", 0.2, "fraud"),
+            VerdictResult("ad_002", "approve", 0.2, "legit"),
+            VerdictResult("ad_003", "reject", 0.2, "fraud"),
+            VerdictResult("ad_004", "approve", 0.2, "legit"),
+        ]
+        r1 = _make_record(task_id="task_2", verdicts=well_calibrated, total_steps=4, action_budget=30)
+        r2 = _make_record(task_id="task_2", verdicts=poorly_calibrated, total_steps=4, action_budget=30)
+        s1 = grade_episode(r1)
+        s2 = grade_episode(r2)
+        assert s1 >= s2, f"Well calibrated ({s1}) should score >= poorly calibrated ({s2})"
+
+
+class TestTask3Grader:
+    def test_network_link_bonus(self):
+        verdicts = [
+            VerdictResult("ad_001", "reject", 0.9, "fraud"),
+            VerdictResult("ad_002", "reject", 0.9, "fraud"),
+            VerdictResult("ad_003", "reject", 0.9, "fraud"),
+        ]
+        links_correct = [
+            LinkResult("ad_001", "ad_002", True),
+            LinkResult("ad_002", "ad_003", True),
+        ]
+        links_wrong = [
+            LinkResult("ad_001", "ad_002", False),
+        ]
+
+        r1 = _make_record(
+            task_id="task_3", verdicts=verdicts, links=links_correct,
+            total_steps=5, action_budget=35, n_fraud_rings=1,
+            ring_sizes=[3],
+        )
+        r2 = _make_record(
+            task_id="task_3", verdicts=verdicts, links=links_wrong,
+            total_steps=4, action_budget=35, n_fraud_rings=1,
+            ring_sizes=[3],
+        )
+        s1 = grade_episode(r1)
+        s2 = grade_episode(r2)
+        assert s1 > s2, f"Correct links ({s1}) should score > wrong links ({s2})"
+
+    def test_graph_based_scoring(self):
+        """Task 3 grader should use edge coverage from ground truth graph."""
+        verdicts = [
+            VerdictResult("ad_001", "reject", 0.9, "fraud"),
+            VerdictResult("ad_002", "reject", 0.9, "fraud"),
+            VerdictResult("ad_003", "reject", 0.9, "fraud"),
+            VerdictResult("ad_004", "reject", 0.9, "fraud"),
+        ]
+        # 4 ads in a ring of 4 -> 6 ground truth edges
+        # Discover 3 of them
+        links = [
+            LinkResult("ad_001", "ad_002", True),
+            LinkResult("ad_002", "ad_003", True),
+            LinkResult("ad_003", "ad_004", True),
+        ]
+        r = _make_record(
+            task_id="task_3", verdicts=verdicts, links=links,
+            total_steps=7, action_budget=35, n_fraud_rings=1,
+            ring_sizes=[4],
+        )
+        score = grade_episode(r)
+        assert 0.0 <= score <= 1.0
+
+    def test_false_links_penalized(self):
+        """False link_accounts should reduce score."""
+        verdicts = [
+            VerdictResult("ad_001", "reject", 0.9, "fraud"),
+            VerdictResult("ad_002", "approve", 0.9, "legit"),
+        ]
+        no_links = _make_record(
+            task_id="task_3", verdicts=verdicts, links=[],
+            total_steps=2, action_budget=35, n_fraud_rings=1,
+            ring_sizes=[3],
+        )
+        false_links = _make_record(
+            task_id="task_3", verdicts=verdicts,
+            links=[LinkResult("ad_001", "ad_002", False)],
+            total_steps=3, action_budget=35, n_fraud_rings=1,
+            ring_sizes=[3],
+        )
+        s_none = grade_episode(no_links)
+        s_false = grade_episode(false_links)
+        assert s_none >= s_false, (
+            f"No links ({s_none}) should score >= false links ({s_false})"
+        )
+
+    def test_coverage_bonus(self):
+        """Agents that review more ads should get a coverage bonus."""
+        few_verdicts = [
+            VerdictResult("ad_001", "reject", 0.9, "fraud"),
+        ]
+        many_verdicts = [
+            VerdictResult("ad_001", "reject", 0.9, "fraud"),
+            VerdictResult("ad_002", "approve", 0.9, "legit"),
+            VerdictResult("ad_003", "reject", 0.9, "fraud"),
+            VerdictResult("ad_004", "approve", 0.9, "legit"),
+        ]
+        ads_meta = [
+            {"ad_id": "ad_001", "severity": 0.8, "ground_truth": "fraud"},
+            {"ad_id": "ad_002", "severity": 0.5, "ground_truth": "legit"},
+            {"ad_id": "ad_003", "severity": 0.8, "ground_truth": "fraud"},
+            {"ad_id": "ad_004", "severity": 0.5, "ground_truth": "legit"},
+            {"ad_id": "ad_005", "severity": 0.5, "ground_truth": "legit"},
+        ]
+        r_few = _make_record(
+            task_id="task_3", verdicts=few_verdicts, total_steps=1,
+            action_budget=35, ads_metadata=ads_meta, ring_sizes=[3],
+        )
+        r_many = _make_record(
+            task_id="task_3", verdicts=many_verdicts, total_steps=4,
+            action_budget=35, ads_metadata=ads_meta, ring_sizes=[3],
+        )
+        s_few = grade_episode(r_few)
+        s_many = grade_episode(r_many)
+        assert s_many > s_few, (
+            f"More coverage ({s_many}) should score > less coverage ({s_few})"
+        )