# LogSentinel v2: Building a Multi-Agent SOC War-Room for RLVR Training

**Author:** Surya-sj  
**HF Space:** [Surya-sj/logsentinel](https://huggingface.co/spaces/Surya-sj/logsentinel)  
**Colab:** [Open Notebook](https://colab.research.google.com/drive/1N-We4n7g9vtH1A1Emtjndqpc8X0CH8zl)

---

## The Problem I Set Out to Solve

Modern incident response is not a single-agent, single-turn task. When a production system goes down at 3am, four people are on a call simultaneously:

- The **Incident Commander** coordinating the response
- The **App SRE** watching pod logs and nginx errors
- The **DB SRE** staring at replication lag and connection pool metrics
- The **Security Analyst** hunting for injection patterns and exfiltration signals

Each of them sees a *different slice* of reality. They have to talk to each other, negotiate severity, and act fast — without stepping on each other's toes.

Current LLM benchmarks test none of this. They ask a single agent to classify logs or answer MCQs. That's not incident response. That's a quiz.

I built **LogSentinel v2** to fix that.

---

## What I Built

LogSentinel v2 is an OpenEnv-compliant reinforcement learning environment that simulates a realistic SOC (Security Operations Center) war-room. It supports:

- **4 agent roles** with different log views
- **5-phase episode lifecycle** driven by actions, not step counts
- **Composable verifiable rewards** with 5 components and 4 anti-hacking penalties
- **Procedural scenario generation** with seeds for reproducibility
- **Adaptive curriculum** that adjusts difficulty based on rolling success rate
- **GRPO training pipeline** built on TRL + Unsloth

---

## Environment Design

### The 5-Phase Lifecycle

```
[DETECT] → [TRIAGE] → [MITIGATE] → [VERIFY] → [FINAL_REPORT]
```

Phase transitions are **action-driven**. The agent can't skip phases — it has to actually do the work of each phase before advancing. Detecting an incident transitions to triage. Executing a mitigation transitions to verify. This forces long-horizon reasoning.

### Partial Observability

Each role sees only the logs from sources relevant to their domain:

```
incident_commander → all sources (but gets fewer metrics)
app_sre            → nginx, app-server-1, app-server-2, k8s
db_sre             → postgres-primary, app-server-1
security_analyst   → nginx, app-server-1, waf, audit-log
```

This means a DB SRE running a query about replication lag sees things the Security Analyst cannot — and vice versa. Agents must use **handoff actions** to share findings via a shared board that all roles can read.

### The Shared Board

When a DB SRE discovers replication lag, they write to the shared board:

```json
{
  "action_type": "request_handoff",
  "agent_role": "db_sre",
  "handoff_to": "incident_commander",
  "handoff_note": "Replication lag 15.2s on pg-replica-1. WAL 256MB behind. Recommend read failover."
}
```

Every agent sees this in their next observation. This is the coordination mechanism — explicit, structured, and gradeable.

---

## Reward Engineering

The reward formula is:

```
R_total = 0.40 × R_outcome
        + 0.20 × R_detection_f1
        + 0.15 × R_severity_accuracy
        + 0.10 × R_efficiency
        + 0.15 × R_teamwork
        − penalties
```

Every component is logged in `state.reward_breakdown` so you can plot each one during training.

### Why this formula works for RLVR

Each component requires **real work** to earn:

- **R_outcome** only goes up when `service_health > 0.7` — which only happens after evidence-backed mitigations
- **R_detection_f1** uses precision + recall, so spamming all incident types tanks your precision
- **R_severity_accuracy** gives partial credit for near-misses (P2 when truth is P1 = 0.5, not 0)
- **R_efficiency** rewards finishing in fewer steps, so the agent learns to be decisive
- **R_teamwork** rewards useful handoffs with non-trivial notes — empty handoffs don't count

### Anti-Hacking Mechanisms

I spent a lot of time thinking about how an RL agent would try to game this environment. Here's what I found and how I stopped it:

| Attack | What the agent would do | How I stop it |
|--------|------------------------|---------------|
| Incident spam | Propose every incident type for max recall | F1 precision term collapses |
| Blind mitigation | Execute mitigations without looking at logs | Penalty + no world state change |
| Instant report | Submit report without detecting anything | `report_before_detection` penalty |
| Noop farming | Repeat `observe_logs` to burn steps safely | Noop counter → efficiency drop |

---

## Scenario Generation

Scenarios are procedurally generated from a `ScenarioConfig`:

```python
ScenarioConfig(
    num_incidents=3,
    difficulty=DifficultyLevel.HARD,
    attack_subtlety=0.8,       # attacker traffic looks like normal POSTs
    observability_quality=0.7, # 30% of normal logs are dropped
    confounding_noise_ratio=0.5,
    seed=42,                   # fully reproducible
)
```

On **hard** difficulty, there are 3 simultaneous incidents:
1. A DB replication lag cascade causing nginx 502s
2. A memory leak causing OOMKills
3. A SQL injection + data exfiltration attack that initially looks like normal search traffic

The security incident is the hardest to catch — the first log from the attacker is a normal-looking `POST /api/search 200 OK`. You need to correlate the UNION SELECT pattern, the 4.2MB response anomaly, and the rate limit breach across multiple sources to identify it.

---

## Adaptive Curriculum

The `adaptive_curriculum` task adjusts difficulty automatically:

```
rolling_avg > 0.70 → promote to harder difficulty
rolling_avg < 0.35 → demote to easier difficulty
window = last 20 episodes
```

This means a model that's mastered easy single-incident scenarios gets automatically challenged with multi-incident hard scenarios — without any manual tuning.

---

## Training Pipeline

The GRPO training script uses TRL's `GRPOTrainer` with Unsloth for 4-bit quantization:

```python
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Qwen2.5-7B-Instruct-bnb-4bit",
    max_seq_length=2048,
    load_in_4bit=True,
)
```

The reward function passed to GRPO is the environment's own `grade_action` + episode-level `compute_episode_reward` — the same signal used during evaluation. This is what makes it RLVR: the reward is **verifiable** against ground truth, not a learned value function.

The Colab notebook walks through:
1. Installing deps
2. Running baseline (heuristic) evaluation
3. GRPO training
4. Post-training evaluation
5. Plotting baseline vs trained reward curves

---

## Results

### Baseline (Heuristic Agent)

The baseline agent uses simple rules: classify logs by level, propose incident based on keyword matching, assign severity heuristically.

```
Avg total reward:      0.45
Success rate:          ~60%
Avg detection F1:      0.51
Avg severity accuracy: 0.72
Avg teamwork score:    0.00  ← handoffs never used
```

The heuristic never uses handoffs (teamwork = 0) and never verifies recovery (efficiency suffers). It gets lucky on severity because the mapping is simple.

### Real LLM Benchmark Results

We benchmarked two Qwen3 models against the heuristic baseline — no fine-tuning, zero-shot:

| Metric | Heuristic Baseline | Qwen3-0.6B (local) | Qwen3-32B (Groq) |
|--------|-------------------|---------------------|------------------|
| Avg Total Reward | 0.41 | **0.49** | **0.49** |
| Efficiency Score | 0.00 | **0.81** | **0.81** |
| Avg Steps to Resolve | 45 | **8** | **8** |
| Success Rate | 66.7% | 66.7% | 66.7% |

The standout result: **efficiency**. The heuristic agent grinds through 45 steps because it follows a fixed sequence. Both Qwen3 models resolve incidents in 8 steps — they read the phase, output the right JSON action, and advance. The reward gap comes almost entirely from not wasting steps.

Teamwork score (handoffs) remains 0.0 for all zero-shot models — this is exactly the gap that GRPO training is designed to close. A trained agent learns that `request_handoff` actions earn `R_teamwork` reward, creating emergent coordination behaviour that no prompt engineering alone achieves.

---

## What I Learned

**1. Phase design matters more than reward weights.**  
Early versions had step-count phase transitions. The agent learned to stall in easy phases. Switching to action-driven transitions immediately fixed this — the agent can only advance by doing the right thing.

**2. Anti-hacking is a first-class design concern.**  
Every reward component I added, I immediately asked: how would an RL agent game this? The answer was almost always "spam the easiest action." Designing against that upfront saved a lot of debugging later.

**3. Partial observability creates emergent coordination pressure.**  
When each role sees different logs, a single agent acting as all roles simultaneously has to explicitly decide what to share. This pressure is what makes handoffs meaningful rather than just a formality.

**4. Verifiable rewards are worth the engineering effort.**  
Ground truth is pre-computed at scenario generation time. Grading is deterministic and fast (<1ms per step). This makes the training loop extremely tight compared to learned reward models.

---

## Try It

```bash
# Reset a hard SOC scenario
curl -X POST https://Surya-sj-logsentinel.hf.space/reset \
  -H "Content-Type: application/json" \
  -d '{"task_name": "soc_warroom_hard", "seed": 42}'

# See all tasks
curl https://Surya-sj-logsentinel.hf.space/tasks

# Full API docs
https://Surya-sj-logsentinel.hf.space/docs
```

Or open the [Colab notebook](https://colab.research.google.com/drive/1N-We4n7g9vtH1A1Emtjndqpc8X0CH8zl) and run the full baseline evaluation — no GPU needed for that part.

---

## Links

- **HF Space:** https://huggingface.co/spaces/Surya-sj/logsentinel
- **Colab Notebook:** https://colab.research.google.com/drive/1N-We4n7g9vtH1A1Emtjndqpc8X0CH8zl
- **API Docs:** https://Surya-sj-logsentinel.hf.space/docs